Changelog 2026

​

ClickHouse release 26.10, 2026-10-22

​

Backward Incompatible Change

• The query analysis that preceded the analyzer is removed. It has not been selectable since 26.9, when enable_analyzer became obsolete; now a server no longer uses it for a query another server sends either, so a cluster whose initiator is older than 26.10 gets the analyzer on this server regardless of what the initiator asks for. Such a mixed cluster can see NOT_FOUND_COLUMN_IN_BLOCK on the old initiator, because the two analyses name the result columns differently; set enable_analyzer = 1 on every server before upgrading any of them. The use_analyzer_for_mutations server setting is removed, and parallel_replicas_only_with_analyzer is obsolete. #120311 (Alexey Milovidov).
• The predicate push-down that the query analysis preceding the analyzer performed on the AST is removed, and enable_optimize_predicate_expression is now obsolete and has no effect. Push-down of a predicate into a subquery is performed by the analyzer as a query plan optimization and is not affected; push-down into a distributed subquery still happens on the AST level and is controlled by allow_push_predicate_ast_for_distributed_subqueries. The visible difference is that EXPLAIN SYNTAX of a query that is not a SELECT, for example EXPLAIN SYNTAX INSERT INTO ... SELECT ..., no longer prints a predicate copied into a subquery, a non-aggregate HAVING conjunct moved to WHERE, or the right-hand table of a JOIN wrapped into a subquery. #121465 (Alexey Milovidov).
• system.metric_log now uses the bucketed schema by default: all profile events and current metrics are stored in a single metrics column of type Map(Enum16(...), Int64) with the bucketed Map serialization and a constant number of 128 buckets, so the table has a few columns instead of thousands, zero values are not stored, and reading one metric reads only one of the 128 buckets. Every metric is also exposed as an ALIAS column, so queries such as SELECT ProfileEvent_Query FROM system.metric_log continue to work unchanged. On upgrade the existing table is renamed to metric_log_0, as with any other change of the structure of a system log table, and the built-in dashboards read merge('system', '^metric_log'), so they keep showing the data collected before the upgrade. Add <schema_type>wide</schema_type> to the <metric_log> section of the server configuration to keep the previous schema. #118318 (Alexey Milovidov). #124016 (Raúl Marín).
• clickhouse-server can now apply a seccomp system call filter at startup. The seccomp setting controls refused system calls: log only records them, while trap, kill, and errno enforce the policy; disabled installs no filter. The config.xml shipped with the server packages and the Docker image now sets seccomp to trap, so a server that runs in a sandbox which refuses the seccomp system call must switch it to log or disabled before the upgrade. Extended-attribute probes used by some hostname-resolution libraries are rejected with ENOSYS instead of the configured seccomp action, so server startup still succeeds under the default trap mode. You might need to disable seccomp when using executable user-defined functions. #119545 (Alexey Milovidov). #122577 (Alexey Milovidov). #124856 (Azat Khuzhin).
• Settings constraints and allow_feature_tier are now enforced for users, roles and settings profiles defined by SQL. A statement is refused if it changes a restricted setting of some user indirectly (such as GRANT, DROP ROLE, DROP SETTINGS PROFILE or CREATE USER OR REPLACE), sets a value outside its author’s constraints, or removes a setting its author is constrained on. A user defined by SQL whose stored settings, roles or assigned settings profiles violate the constraints or allow_feature_tier can no longer log in. The compatibility setting no longer applies values that settings constraints or allow_feature_tier refuse; such settings keep their current value. Login checks use the user’s resolved constraints, preserving bounds inherited from the default profile. Closes #124966. #119472 (Raúl Marín). #125582 (Raúl Marín).
• Overriding a key that is stored in a named collection now requires SHOW NAMED COLLECTIONS SECRETS on that collection. This also applies to overrides through an alias (including flat kafka_* settings over nested kafka.* keys), through a SETTINGS clause of a table engine or table function, and through keys in a dictionary source (checked when the dictionary is created). Resolving a relative url stored in a named collection against s3_base or url_base also counts as an override of url. Adding a key that the collection does not store remains allowed, and keys marked NOT OVERRIDABLE cannot be overridden. The allow_named_collection_override_by_default setting is obsolete. Existing tables whose SETTINGS or alias arguments override a NOT OVERRIDABLE key cannot be loaded after the upgrade, and the server does not start: before upgrading, recreate such tables without the override, or make the key overridable. #123133 (Nikita Mikhaylov).
• max_bytes_before_external_join and max_bytes_ratio_before_external_join are now the spill threshold for every hash-based join_algorithm, including grace_hash, which previously ignored them and spilled on max_rows_in_join / max_bytes_in_join instead. Those two settings are a hard cap again for every algorithm: they no longer cause spilling, and a join that spills no longer exceeds them. All join algorithms remain available; join_algorithm = 'grace_hash' now needs a non-zero spill threshold, and is skipped in favour of the next algorithm in the list when it has none. Set legacy_join_size_limits_trigger_spilling = 1 (or compatibility to 26.8 or older) to restore the previous behavior. #117356 (Yarik Briukhovetskyi). #119335 (Alexey Milovidov).
• arrayCount now returns UInt64 instead of UInt32, so it stays exact for arrays with more than 4294967295 matching elements. Set array_count_legacy_uint32_result = 1 (or compatibility = '26.9') to restore the previous result type. During a rolling upgrade, enable array_count_legacy_uint32_result on upgraded servers for the users under which shard-side queries execute, so distributed queries initiated by not-yet-upgraded servers keep the previous behavior on upgraded shards; remove it after the upgrade is complete. In addition, length(arrayFilter(func, arr)) is now rewritten to arrayCount(func, arr), which is up to 1.29x faster on long arrays, controlled by optimize_rewrite_array_filter_length_to_array_count. #113023 (Alexey Milovidov).
• accurateCast from String or FixedString to a Nullable target now throws the error its non-Nullable target already raised (CANNOT_PARSE_TEXT, TOO_LARGE_STRING_SIZE, …) instead of returning NULL, and no longer throws for a row the source marked NULL. This affects every use of accurateCast: DeltaLake writes stop committing an unrepresentable value as NULL, dictGetOrDefault rejects a default its Nullable attribute cannot hold and now converts only the defaults it uses, IN under transform_null_in = 1 no longer matches an unparsable key against a NULL in the set, and cast_ipv4_ipv6_default_on_conversion_error yields the type’s default for an unparsable address. CAST, accurateCastOrNull, accurateCastOrDefault, and ordinary INSERT are unchanged, as is what a non-text source does with a value it cannot represent; use accurateCastOrNull for the previous behavior. Closes #120524. #120674 (Groene AI).
• SELECT ... FINAL from SummingMergeTree no longer removes the rows whose summed columns are all zero, unlike background merges. Before, this removal looked only at the columns the query read, so results depended on them: for example, SELECT count() FROM table FINAL could return fewer rows than SELECT * FROM table FINAL. Background merges and OPTIMIZE TABLE ... FINAL still remove such rows, so filter out the all-zero rows explicitly if a query relied on their removal. #115032 (Alexey Milovidov).
• A read of a Distributed or Remote table with skip_unavailable_shards = 1 now fails with ALL_CONNECTION_TRIES_FAILED when every shard was skipped before returning any data, instead of returning an empty result, matching what already happens when the only shard is local. Previously an aggregate without GROUP BY silently returned zero rows instead of one. The error fails the whole query, so a query that combines such a read with other data, for example a UNION ALL with local rows, no longer returns those rows. When a shard is itself a Distributed table whose shards were all skipped, the outer query now fails even if another shard answered; skip_unavailable_shards_mode = 'unavailable_or_exception_before_processing' keeps the partial result in that case. No setting restores the previous empty result. Closes #115646. #115775 (Groene AI).
• A MATERIALIZED column computed from a column with a TTL is now recomputed when the TTL resets that column, both during a merge and in ALTER TABLE ... MATERIALIZE TTL. Previously it kept the pre-expiry value forever. If the dependent expression is not defined on the expired column’s default (for example intDiv(100, x) when x expires to 0), the TTL merge now throws instead of writing the stale value; make the expression total on the default (intDivOrZero, if(x = 0, ...)) if such a table must keep merging. Closes #117310. #117815 (Alexey Milovidov).
• With input_format_null_as_default = 0, a JSON null for a non-Nullable numeric column now raises an exception instead of being silently stored as 0 (or nan), consistent with other types and formats. Keep input_format_null_as_default at its default 1 to insert the column default instead. #122137 (Alexey Milovidov).
• ALTER TABLE ... ADD COLUMN and MODIFY COLUMN now apply the data_type_default_nullable setting the same way CREATE TABLE does: with the setting enabled, a column type written without an explicit NULL or NOT NULL modifier becomes Nullable, so the rows that existed before ADD COLUMN read NULL, ADD COLUMN x Int32 DEFAULT NULL is accepted, and MODIFY COLUMN x Int64 keeps a Nullable column Nullable instead of failing. As in CREATE TABLE, an Array, Map, LowCardinality, or another type that cannot be inside Nullable then requires NOT NULL, and so does a MODIFY COLUMN of a sorting key column. To keep the previous result of such an ALTER, write NOT NULL explicitly or run it with data_type_default_nullable = 0. Closes #124504. #124597 (Groene AI).
• arrayJoin is now lowered to an ARRAY JOIN step by default (setting query_plan_lower_array_join_function). The expansion streams through the pipeline instead of being materialized inside the projection, filters on the joined elements are applied before the rows are multiplied, and queries with several arrayJoin calls no longer build the whole cross product in memory. As a consequence, a non-deterministic function such as rand or generateUUIDv4 next to arrayJoin in the same SELECT is now evaluated once per output row, as with the ARRAY JOIN clause, instead of once per source row. Set arrayjoin_nondeterministic_functions_before_expansion = 1 (or compatibility to 26.9 or older) to keep the previous behavior. #117882 (Yarik Briukhovetskyi). #120390 (Yarik Briukhovetskyi). #122240 (Yarik Briukhovetskyi).
• A function that depends on blocks or row order, such as runningDifference, neighbor or rowNumberInBlock, next to the arrayJoin function now sees the blocks of the expansion, like with the ARRAY JOIN clause. Set legacy_array_join_function_nondeterministic_evaluation = 1 to get the previous behavior. #123143 (Yarik Briukhovetskyi).
• A row policy on a Distributed table (or another table that processes the query itself) read through a Merge table now makes queries to the Merge table fail with ILLEGAL_PREWHERE for the users the policy applies to, like queries to that table itself. Before, the Merge table applied such a policy after the table had already sorted, limited and computed extremes, so queries could return wrong rows and show values of hidden rows in extremes. Define the row policy on the underlying tables of the remote servers and on the local child tables instead. This also fixes NOT_FOUND_COLUMN_IN_BLOCK on every query over a Merge table whose child has a row policy when the Merge table also reads a Distributed table. #125765 (Groene AI).
• clickhouse-local --ignore-error now reports a statement that failed at execution instead of skipping it silently: previously it printed no message and no exception code, so a failing multi-statement script looked exactly like a passing one, while clickhouse-client --ignore-error reported every error all along. clickhouse-client also ignored the option entirely when it was combined with --interactive and --query/--queries-file on a terminal, and now applies it to those queries too. The exit code of clickhouse-client --ignore-error changes accordingly: it used to be the code of whichever error the last statement of the batch happened to hit, which made the result depend on the position of the failing statement rather than on the batch as a whole, and it is now a success code, matching what clickhouse-local --ignore-error has always returned. Scripts that relied on the old exit code to detect failures should drop --ignore-error so that the run stops and fails at the first error, or check the reported errors on stderr. #119572 (Yakov Olkhovskiy).
• ATTACH TABLE ... FROM '<path>' now requires READ ON FILE and WRITE ON FILE on the source directory in user_files, in addition to CREATE TABLE, because the query reads the directory and then moves it out of user_files on success. Previously a user could attach and read staged data without those FILE grants. Grant the corresponding FILE privileges to users that rely on this form. #120582 (Alexander Gololobov).
• A MySQL dictionary source created with a CREATE DICTIONARY query can no longer enable enable_local_infile, at the source or at one of its REPLICA entries; loading such a dictionary fails with BAD_ARGUMENTS, which with the default dictionaries_lazy_load is on first use rather than at CREATE. A REPLICA that does not set the key inherits the source’s value, as the connection pool does. To keep the option, define the dictionary in a server configuration file; enable_local_infile = 0, and a SOURCE(MYSQL(NAME ...)) reference to a named collection where the key was never read, are unaffected. #120972 (Groene AI).
• The definitions of disks in storage_configuration and in disk are validated: an element that nothing reads is reported as UNKNOWN_ELEMENT_IN_CONFIG instead of being silently ignored, which finds typos and options that do not exist for this disk type, such as path in an S3 disk. If a server does not start after the upgrade because of an element it does not know, remove that element from the configuration, or set <skip_check_for_incorrect_settings>1</skip_check_for_incorrect_settings>. #118440 (Alexey Milovidov).
• system.*_log tables can now be configured uniformly via system_tables.*. In system.opentelemetry_span_log, finish_date is renamed to event_date and kept as an alias. Aliases are now supported in configured PARTITION BY, ORDER BY, TTL, PRIMARY KEY, and SAMPLE BY expressions, so custom definitions can continue to use finish_date. Some tables now fall back to the default TTL when none is configured, so set ttl explicitly if that matters. #121563 (Azat Khuzhin). #125564 (Azat Khuzhin).
• The setting validate_mutation_query is deprecated and ignored: mutation queries (ALTER TABLE ... UPDATE, ALTER TABLE ... DELETE, DELETE FROM) are now always validated before being accepted. It was enabled by default since 24.11, and disabling it did not avoid the failure but only deferred it to the background, where the mutation got stuck and required manual intervention. Setting it to 0 no longer lets an invalid mutation be accepted: it is rejected when submitted. #121419 (Alexey Milovidov).
• Settings query_plan_read_in_order, query_plan_optimize_prewhere, and query_plan_aggregation_in_order are obsolete and no longer have any effect. Use optimize_read_in_order, optimize_aggregation_in_order, and optimize_move_to_prewhere instead. #121390 (Anton Popov).
• The setting use_compact_format_in_distributed_parts_names is obsolete. The non-compact directory format for the async INSERT queue of a Distributed table has been removed. The setting was 1 by default since version 21.1. Before upgrading, set it to 1 and let the queue drain. #120186 (Alexander Tokmakov).
• The MergeTree settings use_minimalistic_part_header_in_zookeeper and use_minimalistic_checksums_in_zookeeper are obsolete and have no effect: new data parts of ReplicatedMergeTree tables are always stored in ZooKeeper as a single node with the compact header, the default since 20.1. An explicit 0 in table metadata or in the merge_tree configuration is ignored, and such tables and configurations still load. Part nodes written in the old format by earlier versions are still read. Closes #122744. #122790 (Groene AI).
• Settings s3_max_connections and the S3 disk’s max_connections key are obsolete. They have had no effect since 21.4, when the per-endpoint connection pool of the S3 client they bounded was removed, but were still documented as limiting the number of connections per server. Both are still accepted and do nothing. To bound the connections to object storage, use the server settings disk_connections_soft_limit, disk_connections_store_limit, disk_connections_warn_limit, and disk_connections_hard_limit and their storage_connections_* and http_connections_* counterparts; to bound the request rate, use s3_max_get_rps and s3_max_put_rps. #119550 (Alexey Milovidov).
• Removed the zero-argument year function. Queries that used year() must switch to toYear(today()) instead. #120268 (Robert Schulze).
• Removed the system.codecs.is_experimental column. Queries that filtered or selected it should use tier = 'Experimental' instead. #118523 (Raufs Dunamalijevs).
• Removed the experimental function detectTonality. Queries that use it now fail with UNKNOWN_FUNCTION. #123327 (Groene AI).
• Update the zxc library to v0.14.0, which changes the compressed block format of the experimental ZXC codec. Data written with CODEC(ZXC) by earlier versions cannot be read by the new version and fails with CANNOT_DECOMPRESS. If you used the experimental ZXC codec, re-encode the affected columns to a different codec before upgrading, or re-insert the data after the upgrade. #119097 (Actuele AI).
• An aggregate function state that serializes to zero bytes cannot be read back, so the two parameterizations that produced one are now rejected: a -Resample combinator over an empty range (begin >= end, such as countResample(10, 5, 1)) throws ARGUMENT_OUT_OF_BOUND, and covarSampMatrix, covarPopMatrix, and corrMatrix with no arguments throw NUMBER_OF_ARGUMENTS_DOESNT_MATCH. Both returned an empty array before. No serialized format changes, but a table with such a column no longer loads. Before upgrading, drop the column where the engine supports DROP COLUMN, and the whole table where it does not (Log, TinyLog, StripeLog, Set, Join, File). #119728 (Groene AI).
• A data type naming the state of a function that only works as a window function, such as AggregateFunction(rank), is now rejected with BAD_ARGUMENTS: such a state has no serialization, so the column could never hold a value. This applies wherever a type is named, including inside Array, Map, Tuple, and another state’s argument list. A table that already stores such a column will no longer attach after the upgrade; drop or rewrite the column before upgrading. #122148 (Groene AI).
• rankState and every other -State of a function that only works as a window function, such as denseRankStateIf or initializeAggregation('lagInFrameState', x), are now rejected with BAD_ARGUMENTS when the query is analyzed, also where the state is never computed (toTypeName, EXPLAIN, a column the outer query drops), and so is such a type in a binary-encoded Native header. Computing such a state always failed, but a view or CREATE TABLE ... AS SELECT could still store its type. A table whose engine expressions build such a state, for example in a sorting key, partition key, skip index, or TTL, no longer attaches. #122463 (Groene AI).
• CREATE TABLE and full-definition ATTACH TABLE now refuse a CONSTRAINT name declared more than once, and ALTER TABLE ... ADD CONSTRAINT / MODIFY CONSTRAINT now refuse an expression containing arrayJoin even when the command would install nothing. A table already stored with a repeated constraint name still attaches and keeps enforcing each of its declarations; give each constraint a unique name in new definitions. #121235 (Groene AI).
• The Remote, RemoteSecure, and Cluster database engines now reject an empty remote database name with BAD_ARGUMENTS instead of creating a database that lists and reads no tables. A server that holds such a database, created by an earlier version, does not start after the upgrade, so drop these databases before upgrading. Closes #122616. #122675 (Groene AI).
• Connections to MySQL (the mysql table function, the MySQL table and database engines, and MySQL dictionaries) that set ssl_ca or ssl_ca_pem now verify the server certificate chain and host name, after updating mariadb-connector-c to v3.4.10. Previously the CA was not enforced, so a connection to a server whose certificate is signed by another CA, or whose certificate does not list the host name or IP address used to connect, succeeded and now fails with self-signed certificate in certificate chain or Hostname verification failed. Before upgrading, make sure the configured CA signs the MySQL server certificate and that the certificate lists the host name or IP address ClickHouse connects to. #119510 (Actuele AI).

​

New Feature

• Sets built for IN with a subquery or a table now spill to disk by default once query memory usage exceeds half the available memory under the applicable server or user memory limit, calculated at the start of execution (max_bytes_ratio_before_external_set = 0.5). The max_bytes_before_external_set setting can set an absolute threshold (max_memory_usage is not taken into account by the ratio, so use it to spill before that limit is reached), and setting both thresholds to 0 keeps sets in memory. This feature is now generally available. #123860 (Nihal Z. Miaji). #124907 (Nihal Z. Miaji).
• Support parts of the Trino SQL dialect in regular ClickHouse SQL: TRY_CAST(x AS T), the ARRAY[...] array constructor, the type names ARRAY(T) and MAP(K, V), and aliases for Trino functions such as starts_with, slice, regexp_like, day_of_week, url_extract_host, bitwise_and, arbitrary, and bool_and. #124763 (Alexey Milovidov).
• Implemented a generic block nested loop join algorithm. JOIN now supports an arbitrary ON condition with no equality between the two tables, executing it as a block nested loop join, as slow as a CROSS JOIN with a filter, instead of failing with INVALID_JOIN_ON_EXPRESSION. The new setting allow_block_nested_loop_join turns it on/off. #123642 (Alexey Milovidov). #123792 (Groene AI).
• Added the function-composition operator f | g and the argument placeholders _, _1, _2, …: for example, arrayMap(plus(_, 1) | multiply(_, 2), [1, 2, 3]) returns [4, 6, 8]. A composition can be used wherever a function is expected, including as an argument of a higher-order function. The operands can be lambdas, function names, nested compositions, or expressions with placeholders, which also work on their own, for example arrayMap(_1 * 2 + 1, [1, 2, 3]). The composition is resolved at query analysis time into an ordinary lambda. Note that | here is not bitwise OR; use bitOr for that. #116029 (Alexey Milovidov).
• The query cache can now store its entries on disk, backed by a preconfigured filesystem cache: this gives the entries more space, makes them survive server restarts, and, with a distributed filesystem cache, provides a way to share them between servers. Select the filesystem cache with the new setting query_cache_on_disk_cache_name; see also the new settings query_cache_on_disk_codec, enable_writes_to_query_cache_on_disk, and enable_reads_from_query_cache_on_disk. Closes #115361. #115382 (Alexey Milovidov).
• Added a notebook mode to the Web UI (/play). After a query is run, a [+] button under the result adds further cells, each either a query with its own result or a block of Markdown text, turning the tab into a notebook in the style of Jupyter or Observable. Cells can be run, reordered, inserted and deleted individually; the query editor collapses to a single line when it is not focused, results are bounded and scroll inside themselves, and the whole notebook is restored on reload. Notebook mode is per tab, so a tab that has only one cell looks and behaves exactly as before. #115378 (Alexey Milovidov).
• Added support for reading and writing data in the SQLite format. #104510 (Yue Ni).
• Added the Cluster database engine, which provides real-time access to the tables of a configured cluster database and forwards SELECT and INSERT queries to it. #114417 (Alexey Milovidov).
• Added the Overlay database engine: CREATE DATABASE db ENGINE = Overlay(db1, db2, ...) exposes the union of the tables of existing databases, each table working as an Alias table to the table of the first source database that has it. #121878 (Alexey Milovidov).
• CREATE DATABASE and ATTACH DATABASE now accept a SETTINGS clause without an explicit ENGINE clause, defaulting to Atomic, so a database can define its metadata disk inline the same way a single table does. A single SETTINGS clause can now also mix database-engine settings with ordinary query settings. #118423 (Alexey Milovidov).
• MergeTree tables in Ordinary databases can now be converted to ReplicatedMergeTree with ATTACH TABLE ... AS REPLICATED or the convert_to_replicated flag. #118757 (Kirill).
• Added query-aware scheduling in workloads: the new workload setting scheduler selects how queries of a workload are scheduled — fifo (default), fair (weighted fair queueing with per-query weights that can be lowered as a query runs), las (least-attained-service), or priority (by the query setting workload_priority). The leaf-node type in system.scheduler is now request_queue instead of fifo. #116800 (Sergei Trifonov).
• Queries can now spill to disk under memory pressure based on limits set per workload. The max_memory_before_spill and max_memory_to_spill_ratio options of WORKLOAD define a soft memory limit that triggers spilling for ORDER BY, GROUP BY, IN, DISTINCT, and JOIN with grace_hash. #119443 (Azat Khuzhin).
• Quotas can now limit the number of rows written by INSERT queries with the new written_rows resource, for example CREATE QUOTA q KEYED BY ip_address FOR INTERVAL 1 minute MAX written_rows = 100000, query_inserts = 250 TO user. Until now only the written bytes could be limited (written_bytes), which is awkward when the natural unit of the workload is rows. The current and maximum values are exposed as written_rows and max_written_rows in system.quota_usage, system.quotas_usage, and system.quota_limits, and the resource can also be set in users.xml as <written_rows>. #118272 (Alexey Milovidov).
• Added the settings max_temporary_tables, max_temporary_table_memory_usage, max_temporary_table_size_bytes_compressed, and max_temporary_table_size_bytes_uncompressed to limit the number of temporary tables in a session, the memory usage of Memory temporary tables, and the size of MergeTree temporary tables. #122165 (Alexey Milovidov).
• Added the SQL-standard SIMILAR TO pattern-matching predicate. #101514 (Zhao Zheguang).
• Allow MATERIALIZED common table expressions inside recursive WITH. A materialized helper CTE is evaluated once and every recursive step reads the snapshot, instead of the subquery being re-evaluated on each iteration. The recursive CTE itself still cannot be materialized. #123452 (Alexey Milovidov).
• Query parameters can now be used for primary USER and ROLE names in CREATE, ALTER, and DROP, and for role and grantee names in GRANT and REVOKE, for example GRANT {role:Identifier} TO {user:Identifier}. #109696 (Pervakov Grigorii).
• SET DEFAULT ROLE now accepts an ON CLUSTER clause, in either position: SET DEFAULT ROLE role1 TO user ON CLUSTER my_cluster and SET DEFAULT ROLE ON CLUSTER my_cluster role1 TO user. CURRENT_USER is resolved on the initiator before the query is distributed. SET ROLE only affects the current session and still takes no ON CLUSTER clause. #118429 (Alexey Milovidov).
• Added CREATE OR REPLACE NAMED COLLECTION ... to replace an existing named collection entirely. #121569 (Pervakov Grigorii).
• The postgresql table function and the PostgreSQL table engine can now connect to another ClickHouse server over the PostgreSQL protocol. The query(...) variant still needs an explicit column list. Added the PostgreSQL-compatibility functions format_type and current_setting. Also fixed cancellation over the PostgreSQL protocol: a cancel request never cancelled anything. #112940 (Alexey Milovidov).
• Disks with the plain_rewritable metadata (s3_plain_rewritable, and azure/local object storages with metadata_type = plain_rewritable) support hard links if configured with the new enable_hard_links setting, which enables ALTER operations and mutations on MergeTree tables stored on them. Mutations and partition operations share the unchanged blobs between parts instead of copying them. The setting is off by default, because a disk that has hard links cannot be read by older servers. #110939 (yanglongwei). #119304 (Alexey Milovidov).
• Added EXCEPT DATA FROM TABLE / EXCEPT DATA FROM TABLES clauses to BACKUP queries, so a table’s definition can be backed up without its rows, for any table engine that supports backups. Formatting now quotes keyword table names correctly in EXCEPT DATA FROM TABLES, so BACKUP ... ON CLUSTER round-trips through the formatter. Closes #111827. #111895 (Aditya Kumar). #122468 (David Meng).
• Added five columns to system.query_log describing what a query really executed: used_number_of_joins, used_join_algorithms, used_join_kinds, used_join_strictness, and spilled_to_disk. This makes it possible to find problematic JOIN patterns and queries that fall back to external memory across a fleet without reproducing each query. A spilling SELECT DISTINCT now records distinct in spilled_to_disk instead of [], and a spilling block nested loop join records join. #113833 (Matheus Nerone). #121830 (Groene AI). #124681 (David Meng).
• Added table function traceView that renders an OpenTelemetry trace from system.opentelemetry_span_log as a call tree with a per-span timeline bar, for debugging where the time of a traced, for example distributed, query was spent. It has a host column with the hostname of the server that recorded each span, so the spans of the initiator, remote shards, replicas, and stateless workers can be distinguished and filtered with WHERE host = .... #116325 (Diego Gomes Tomé). #122169 (Diego Gomes Tomé). #123528 (Diego Gomes Tomé).
• clickhouse-client --login can now authenticate against arbitrary OAuth/OIDC identity providers via endpoint discovery, with explicit endpoint and scope overrides and improved RFC 8628 device-code handling. On Linux desktops it now also passes the client’s environment to xdg-open when opening the identity-provider login page, so the system browser is detected correctly. Closes #120127. #110718 (yanglongwei). #121337 (Konstantin Bogdanov).
• Added interactive client commands /dialect, /lang, and /language to clickhouse-client and clickhouse-local, so you can switch the active SQL dialect before parsing, including when the current dialect cannot express a SET query, for example to leave kusto or promql. The current non-default dialect is now shown in the prompt, and SET dialect = ... keeps the prompt in sync. #114568 (Alexey Milovidov).
• chdig v26.10.1: ZooKeeper browser, per-host summary, and better filters. #124100 (Azat Khuzhin).
• Support reading the Parquet VARIANT type (unshredded, as written for example by Spark 4.0) into Dynamic, with object values read as JSON. A variant column can also be read directly as JSON if all its values are objects. Shredded variants are not supported yet. #117106 (Konstantin Vedernikov).
• Support the variant type of Iceberg. Based on the Parquet variant reader in #117106. #117198 (Konstantin Vedernikov). #125458 (Alexey Milovidov).
• Support the Parquet FILE logical type (a reference to bytes stored inline or in an external file): it is read as a named Tuple, and can be written from a Tuple with the corresponding element names with the new setting output_format_parquet_write_file_logical_type. #125132 (Alexey Milovidov).
• Lazy type hints for the JSON type are now generally available under the enable_json_lazy_type_hints setting. The setting remains disabled by default. #122711 (Pavel Kruglov).
• Add a new input format Freeform, which does schema inference for arbitrary tabular text files. Closes #34006. #43449 (Tri Dao).
• The generateRandom table function and the GenerateRandom table engine now support JSON, Dynamic, Variant, BFloat16, Time, and Time64. Generated JSON documents have a realistic schema derived from the seed. The new settings null_ratio, max_json_depth, and max_json_keys_per_object control the share of nulls and absent keys, the nesting depth, and the number of keys per object. Generation respects memory limits. CREATE TABLE ... AS generateRandom(...) also keeps the SETTINGS argument in the stored definition. #107966 (apst41). #124772 (Vadim Lopatin). #123247 (Raúl Marín).
• Add the obfuscate table function for generating anonymized query results. #115004 (Alexey Milovidov).
• Text indexes with the keyValuePairs tokenizer now answer map['key'] IN ('value1', 'value2') directly from the index, as they already did for map['key'] = 'value'. Added the function mapContainsKeyValue(map, key, value), which returns whether a map contains an entry with the given key and value; in contrast to map[key] = value, all entries of a repeated key are considered, and such an index evaluates the function directly. The mergeTreeTextIndex table function returns the decoded parts of each token of such an index in the columns token_key and token_value. #120244 (Jimmy Aguilar Mena). #120010 (Jimmy Aguilar Mena). #120183 (Anton Popov).
• The mergeTreeTextIndex table function now returns posting lists and token positions for a text index in the new virtual columns row_id and positions. If one of them is selected, the function returns a row per posting. The positions column is available for indexes created with support_phrase_search = 1. #125536 (Anton Popov).
• Added mapRemove, which removes all entries with a specified key from a map, and mapEntries, which returns a map’s entries as an array of key-value tuples. #123322 (Minh Vu). #123344 (Minh Vu).
• The -Map combinator can now be used with non-Map columns (the key is passed as the last argument, e.g. sumMap(value, key)) and supports aggregate functions with multiple arguments, e.g. argMaxMap(arg, val, key). #123449 (Alexey Milovidov).
• The aggregate functions avg, sum, sumWithOverflow, and sumKahan now support the Interval data types. The result is an interval of the same unit. #121458 (Alexey Milovidov). #121600 (Alexey Milovidov).
• The linear regression aggregate functions can now also be called under ClickHouse-style names: regrSlope, regrIntercept, regrR2, regrCount, regrAvgX, regrAvgY, regrSXX, regrSYY, and regrSXY. #119796 (mosya415).
• Added the arrayMinIndex and arrayMaxIndex array functions. #120097 (Minh Vu). #123686 (Minh Vu).
• Added the arrayRemoveAt array function. #120445 (Minh Vu).
• Added higher-order array functions to pack bits: arrayPackBitsToUInt64, arrayPackBitsToString, arrayPackBitsToFixedString, arrayPackBitGroupsToUInt64, arrayPackBitGroupsToString, and arrayPackBitGroupsToFixedString. #123478 (Alexey Milovidov).
• Added SQL functions normalizedQueryHashUnordered and normalizedQueryHashUnorderedOrNull. Unlike normalizedQueryHash, they ignore the order of expression lists such as the SELECT list, GROUP BY keys, and function arguments, so logically equivalent queries can collapse to one pattern in query_log; the OrNull variant returns NULL instead of throwing when the input is not valid ClickHouse SQL in the current session. #116345 (Yarik Briukhovetskyi).
• Add functions encodeQuotedPrintable and decodeQuotedPrintable for Quoted-Printable encoding (RFC 2045), commonly used in email. #125454 (Jay Pokale).
• Added the sql_compatible_division_for_integers setting, disabled by default, to make / and divide return integer quotients truncated toward zero when both operands are integers. It supports integer types up to 256 bits and checks division by zero and result overflow; division involving Decimal or floating-point operands is unchanged. #124490 (Ilya Yatsishin).
• Added SYSTEM RESET FILELOG [db.]table [FILE 'file_name' [OFFSET n | TO END]]: a FileLog table reads all its files again from the beginning, or one file from the beginning, from a given offset, or only from data appended after the command, without recreating the table. Requires the new SYSTEM RESET FILELOG privilege. #123545 (Groene AI).
• The FileLog table engine now accepts globs (*, ?, {a,b}, {N..M}) in the file name of its path, for example FileLog('/var/lib/clickhouse/user_files/my_app/*.log', 'JSONEachRow'), and reads only the matching files of that directory, so the compressed files logrotate leaves next to the log are not read. A file that is already read keeps being read after it is renamed to a non-matching name, such as app.log rotated to app.log.1. #123590 (Groene AI).
• Added the flat_namespaces setting to the DataLakeCatalog database engine. Set it for Iceberg REST catalogs that have only single-level namespaces and reject the parent query parameter of the list-namespaces endpoint, such as Apache Polaris federated to AWS Glue: only top-level namespaces are then listed and sub-namespaces are not requested. #117329 (Konstantin Vedernikov).
• Added the ClickHouse Keeper ListWithOptions operation for filtered direct or recursive listing with optional stats, data, limits, and shuffled results. #117521 (Antonio Andelic).
• Added the MergeTree setting min_unreserved_disk_space_for_merge: background merge selection subtracts the specified amount from the unreserved disk space, so large merges stop being selected before they can take the last free space and starve inserts. Unlike keep_free_space_bytes, the protected space remains usable by ClickHouse itself. #115723 (Valery Petrov).
• You can now disable reading through the executable table function and from Executable/ExecutablePool tables with the new allow_executable_tables setting, for a deployment that must not run arbitrary server-side scripts. It defaults to true, so existing behavior is unchanged. #120713 (Hanzi Jiang).
• Added server setting ai_function_allow_sql_named_collections. When disabled, AI functions accept only named collections defined in the server configuration file. #124999 (George Larionov).
• Added <openSSL><server><cipherSuites> and <openSSL><client><cipherSuites> settings to restrict which TLS 1.3 cipher suites a TLS port negotiates. Without them, cipherList controls only TLS 1.2 and below. An absent or empty value keeps the OpenSSL defaults. Closes #119929. #120480 (Groene AI).
• You can now set a custom syslog program name for ClickHouse logs using logger.syslog.programname, making it easier to distinguish multiple ClickHouse instances in syslog. The default syslog identifier is now command-specific: remote syslog previously sent - as the RFC 5424 APP-NAME and now sends the command name (for example clickhouse-server), and the multi-purpose binary started as clickhouse server or clickhouse keeper now uses clickhouse-server / clickhouse-keeper instead of clickhouse for local syslog. #98789 (François Martin).
• Added the Copy event and the source_bucket and source_remote_path columns to system.blob_storage_log for server-side object copies (S3 CopyObject and UploadPartCopy, Azure Blob Storage copies). Azure backups, restores, and AzureQueue moves now write to system.blob_storage_log too. Closes #123163. #123178 (Groene AI).
• Added an is_lazy column to system.dictionaries. It shows whether each dictionary loads on first use, taking into account both its dictionary_lazy_load setting and the server setting dictionaries_lazy_load. This makes it possible to tell a lazy dictionary that has not been used yet from an eager one that has not been loaded yet; both show NOT_LOADED. #124260 (Miсhael Stetsyuk).
• Added the replica_group column to system.clusters. For clusters that belong to a Replicated database, it shows the replica group of each node as published from the server configuration replica_group_name, or its alias role. #120537 (RinChanNOW).
• Added SYSTEM DISABLE ALL FAILPOINTS, which disables every fail point at once and resumes any thread blocked on a pauseable one. #118877 (Alexey Milovidov).

​

Experimental Feature

• Added an experimental columns cache that keeps deserialized columns of wide MergeTree parts in memory, one entry per stripe of granules of a column. Compared with the uncompressed cache, this saves time on deserialization and copying. It is enabled by the experimental setting use_columns_cache, with enable_reads_from_columns_cache and enable_writes_to_columns_cache as beta sub-settings, and it applies to tables in databases that assign UUIDs (Atomic and Replicated). The cache gives its memory back to queries when the server is short of it. #96844 (Alexey Milovidov).
• Added experimental LATERAL JOIN support, allowing correlated subqueries in the FROM clause. This enables the standard SQL pattern for “top N per group” and per-row aggregation queries. Controlled by the allow_experimental_lateral_join setting. #97833 (Dimitri König).
• Added an experimental logsql dialect: LogsQL, the log query language of VictoriaLogs, can now be used to query ClickHouse tables with logs. Enable it with SET enable_logsql_dialect = 1, logsql_table = '<table>', dialect = 'logsql' (plus logsql_database, logsql_time_column, and logsql_message_column as needed). LogsQL queries are translated at parse time into ordinary SELECT queries, so text filters expect String-backed log columns, while the math pipe applies raw arithmetic and needs numeric operand columns (String field values are not coerced to numbers). #113207 (Alexey Milovidov). #123891 (Alexey Milovidov).
• Added the beta PCO compression codec for fixed-width numeric columns, based on the pcodec library. Enable it with enable_pco_codec = 1. #110613 (Alexey Milovidov).
• Added a new implementation of the unity catalog type in the DataLakeCatalog database engine. It serves both DeltaLake and Iceberg tables from a single Unity Catalog, detecting the format of each table and routing Iceberg tables through the Unity Catalog Iceberg REST endpoint. Enable it with the database setting use_unity_catalog_v2, or set the session setting of the same name before CREATE DATABASE. Switch an existing database with ALTER DATABASE ... MODIFY SETTING use_unity_catalog_v2 = 1. #115879 (Asya Shneerson).
• Added an experimental embedded Iceberg REST catalog server with ClickHouse HTTP authentication. It supports GET /v1/config, listing, creating, and checking the existence of namespaces, stored either in memory or in Keeper, and listing, creating, loading, checking the existence of, and dropping tables, with their metadata files stored in S3. The catalog can also serve TLS via secure_port, using the certificate and key in openSSL.server, and correctly handles / in table and namespace names. Closes #125762. #118924 (Asya Shneerson). #123000 (Asya Shneerson). #123331 (Asya Shneerson). #124978 (Asya Shneerson). #125791 (Groene AI).
• Added experimental XGBoost dictionaries: the XGBOOST layout trains a gradient-boosted model from the source table at load time, and predictXGBoost runs inference. Enable it with enable_xgboost. Available only in builds that include the XGBoost contrib library. #109999 (Matheus Nerone).
• Added ipcryptEncrypt, ipcryptDecrypt, ipcryptPrefixEncrypt and ipcryptPrefixDecrypt functions for format-preserving encryption of IP addresses following the IPCrypt specification. The deterministic mode produces consistent output for deduplication, while the prefix-preserving mode maintains subnet relationships. Enable with SET enable_ipcrypt = 1. Contributed by Frank Denis. #123680 (Alexey Milovidov).
• Added read support for Iceberg v3 deletion vectors stored in Puffin files. #110781 (linjiayu1025-collab).
• Added experimental support for ALTER TABLE ... DROP PARTITION on standalone Iceberg tables and Iceberg REST catalogs. On standalone tables, the operation currently requires Iceberg format-version 2 with a single non-evolved partition spec. Old date-transform spellings in the partition expression are now evaluated with the Iceberg transform semantics too. #111219 (Den Kalantaevskii). #123273 (Groene AI). #122269 (Den Kalantaevskii).
• parallel_replicas_plan_based can now distribute queries with an IN (subquery) predicate instead of falling back to single-node execution. Closes #111876. #114086 (Igor Nikonov).
• With parallel_replicas_plan_based enabled, distributed INSERT ... SELECT again uses distributed execution when parallel_distributed_insert_select = 2, instead of running the SELECT on the initiator alone. Closes #118276. #121097 (Igor Nikonov).
• Distributed query plans built with make_distributed_plan now compress exchanged data with network_compression_method; for ZSTD, they also use network_zstd_compression_level instead of the server’s default codec settings. #120333 (Alexander Gololobov).
• Queries executed with the experimental make_distributed_plan setting now propagate the OpenTelemetry trace context to stateless workers via the traceparent and tracestate HTTP headers of the task dispatch request. Each dispatched task produces a StatelessWorkerClient::sendTask span on the initiator plus InterserverIOHTTPHandler and DistributedPlanTask::execute spans on the worker, with the task id, initial query id, plan hash, and exchange streams as attributes, so the whole distributed plan is visible as one trace in system.opentelemetry_span_log. #119010 (Diego Gomes Tomé).
• Join runtime filters can now travel across distributed query plan stages with the experimental setting distributed_plan_join_runtime_filters (default off). Build-side tasks send partial filters to probe-side scans before their shuffle exchange, reducing the number of rows crossing the network. A filter that is lost, late, or never built cannot change results. #114822 (Harikrishnan Prabakaran).
• Distributed index analysis is now beta. #123577 (Azat Khuzhin).
• Added the experimental setting enable_join_seal_gated_reading (default off). When enabled together with enable_join_runtime_filters and the probe-side join key is a prefix of the table’s primary key, the probe side of a hash JOIN starts reading only after the build side completes, and the completed runtime filter prunes whole mark ranges by the primary key before read tasks are created, instead of only filtering rows after they were read. #113512 (Nikolai Kochetov).
• Replaced the experimental settings query_plan_optimize_join_order_use_conflict_detector_a and query_plan_optimize_join_order_use_conflict_detector_c with query_plan_optimize_join_order_conflict_detector, which takes a, c, or '' (the default). #120630 (Fisnik Kastrati).
• Adaptive codec selection (enable_adaptive_codec_selection) also tries each specialized codec followed by the table’s default codec, and makes merges of Float32 and Float64 columns faster. #121630 (Raufs Dunamalijevs). #122248 (Raufs Dunamalijevs).
• Added the setting json_extract_named_tuples_as_objects (default off): when enabled, the JSONExtract functions fill named tuples from JSON objects only. By default a JSON array fills a named tuple positionally, so which element lands in which field depends on the order the fields were declared in. #122637 (Alexey Milovidov).
• Experimental UNIQUE KEY SST probes now batch lookups with MultiGet, so absent keys can be filtered by the SST bloom filter. More unsupported UNIQUE KEY operations are now rejected early. #119997 (johnjing). #123002 (Murphy).
• The experimental reader executor (use_reader_executor = 1) now plans filesystem cache reads ahead over a window set by reader_executor_plan_look_ahead instead of re-checking the cache on every read window. It also no longer fails spuriously with CANNOT_READ_ALL_DATA when a concurrent download advances the filesystem cache during a cold read. #115816 (Sema Checherinda). #122214 (Sema Checherinda).
• Streaming queries now prefer a commit_order projection when one is available for the current read round. #120420 (Mikhail Artemenko).
• PromQL again supports the range functions present_over_time, absent_over_time, quantile_over_time, and predict_linear, and now also supports stddev_over_time, stdvar_over_time, first_over_time, ts_of_first_over_time, ts_of_last_over_time, and mad_over_time. #112355 (Valery Petrov). #120352 (Vitaly Baranov). #121578 (Vitaly Baranov). #121800 (Vitaly Baranov).
• Added PromQL functions min_of and max_of to return the smaller or larger of two scalar values. #121250 (Minh Vu).
• PromQL now implements timestamp for both direct selectors and general expressions, matching Prometheus 3.5.0 semantics. #112356 (Valery Petrov).
• Support standalone PromQL step and range scalar functions. The PromQL parser now also accepts bare calls to start, end, sort_by_label, sort_by_label_desc, double_exponential_smoothing, and limit_ratio, so these not yet implemented functions are rejected after parsing instead of failing with a parse error. #121118 (Minh Vu). #123589 (James Cunningham).
• prometheusQuery now always returns Float64 values and DateTime64 timestamps with at least millisecond precision. The time zone of the returned timestamps comes from its DateTime or DateTime64 arguments when they agree, and otherwise from the table. #120285 (Vitaly Baranov).
• PromQL now treats Prometheus stale markers as absent in instant selectors and omits them from range selectors. #121098 (Minh Vu).
• PromQL now rejects queries that use the same label in both on(...) and group_left(...) or group_right(...). #114376 (Minh Vu).
• PromQL now rejects vector selectors whose matchers can all match the empty string. #121099 (Minh Vu).
• PromQL’s stddev and stdvar now return NaN instead of 0. #122743 (James Cunningham).
• PromQL now returns NaN from date functions when the sample is not finite. #124300 (James Cunningham).
• PromQL no longer fails on empty joins. #124362 (James Cunningham).
• Aligned PromQL string and comment parsing with Prometheus: quoted strings cannot contain a line feed, and a line comment can end at the end of the query without a space after #. Closes #117812. #124584 (Vitaly Baranov).
• PromQL range functions over scalars now handle windows with no samples without failing on NULL, including when used in nested range functions. #124586 (Vitaly Baranov).
• Added deduplication caches for TimeSeries metric families and tags, and fixed stale cache entries when the pending rows are full. #121915 (Vitaly Baranov). #122771 (Vitaly Baranov).
• The tags inner table of the TimeSeries table engine is split into a tags table without min_time and max_time and a separate table of time ranges. #121924 (Vitaly Baranov).
• The inner tables of the TimeSeries table engine now use the outer column name metric_family instead of metric_family_name. #119226 (Vitaly Baranov).
• Fixed ALIAS_REQUIRED for a join with a VIEW written without an alias under the experimental setting analyzer_inline_views = 1: the inlined subquery now keeps the view’s name as a table expression name, so it needs no alias and stays usable as a column qualifier, with or without the database name. #118401 (Alexey Milovidov).
• Reading subcolumns of the experimental Quantized codec (vec.quantized, vec.product_quantization_codebook) through a Merge table no longer breaks when another child table of the Merge lacks the base column. #123895 (Groene AI).

​

Performance Improvement

• Large hash tables used for GROUP BY, uniqExact, and JOIN now use up to two times less memory, and aggregation with many keys is faster on many threads. #125001 (Alexey Milovidov).
• The query condition cache now serves conditions that involve the current time, e.g. WHERE time >= now() - INTERVAL 10 DAY, when reading from MergeTree tables, by deriving a deterministic condition from them with the current-time constant rounded onto a time grid. Closes #115504. #115522 (Alexey Milovidov).
• Enabled the setting apply_string_filters_during_scan by default. Substring search conditions on String columns (LIKE, position, startsWith, endsWith) from PREWHERE are pushed down into the column scan, and non-matching values are read as empty strings without copying their data. This makes such queries faster and lowers their memory usage when reading from MergeTree tables and from Parquet pages whose string values use the PLAIN, DELTA_LENGTH_BYTE_ARRAY, DELTA_BYTE_ARRAY, or dictionary encodings, including reads using the uncompressed cache on object storage. #115848 (Alexey Milovidov). #124503 (Alexey Milovidov). #125079 (Groene AI). #124507 (Alexey Milovidov).
• Queries GROUP BY key LIMIT n and GROUP BY key ORDER BY key LIMIT n over MergeTree tables now read much less data: the aggregation publishes the boundary of its top-K heap to the reading step, which drops rows beyond it before reading the other columns and can skip whole granules through the primary key or a minmax skip index. Controlled by the new setting enable_group_by_top_k_dynamic_filtering. The same primary-key granule skipping now also applies to ORDER BY column LIMIT n when the column is in the primary key. #121145 (Alexey Milovidov).
• Queries with ORDER BY <column> LIMIT n over Parquet files now use TopN dynamic filtering with use_top_k_dynamic_filtering: rows that cannot enter the top-K are filtered inside the reader before the remaining columns are read, and whole row groups are skipped when their min/max statistics prove they cannot contribute. This works through the file and url table functions, the File and URL engines, object storage (s3, Azure, HDFS), and data lake tables such as Iceberg. Floating-point sort columns are excluded for now. #115798 (Alexey Milovidov). #122109 (Alexey Milovidov). #124861 (Alexey Milovidov).
• query_plan_join_shard_by_pk_ranges is enabled by default. A full_sorting_merge or parallel_full_sorting_merge JOIN of two MergeTree tables whose join keys contain a prefix of the primary key of both tables now runs as independent streaming merge joins over primary-key ranges on all threads instead of a single-threaded merge join, several times faster and with several times less memory in our tests. Hash joins are affected only when both join runtime filters and spilling to disk are disabled. Set query_plan_join_shard_by_pk_ranges = 0 to restore the previous behavior. #117856 (Alexey Milovidov). #119925 (Alexey Milovidov).
• Enabled the text-index postings cache by default with use_text_index_postings_cache. On a 1B-row table, a four-token phrase query that previously recomputed posting lists and phrase-search results within one query went from 19 s to 0.12 s. The shared cache is bounded by text_index_postings_cache_size (2 GiB by default). #120677 (Nikita Fomichev). #123228 (Anton Popov).
• With optimize_aggregation_in_order = 1, or with the default optimize_distinct_in_order = 1, a GROUP BY or DISTINCT whose keys extend past the table’s sorting key no longer re-packs the whole block’s grouping key once per run of equal sorting-key values. That cost was quadratic in max_block_size once the sorting prefix is near-unique, so GROUP BY a, b on a table sorted by a is up to 130x faster. Closes #45386. #120259 (Groene AI).
• Blocks sent back by remote replicas are now marshalled in parallel (enable_parallel_blocks_marshalling) when parallel_replicas_plan_based is enabled. #121642 (Igor Nikonov).
• GROUP BY with a Nullable key now uses the adaptive aggregator, except when that key is the only one and is a number of up to 8 bytes, a String, or a FixedString, or when any key is LowCardinality. Schema inference makes such keys common in CSV, JSON, and Parquet files; high-cardinality aggregations over them become up to 2x faster with up to 2.2x less memory. #124868 (Groene AI).
• Queries over Nullable columns now use the same optimizations as over non-Nullable ones: length, empty, notEmpty, and comparisons with an empty string on Nullable(String) read the size subcolumn, sum(x + c) is rewritten to sum(x) + c * count(x), and uniqExact over Nullable is merged in parallel. Some queries become several times faster. #125011 (Alexey Milovidov).
• The trivial GROUP BY ... LIMIT n optimization (setting optimize_trivial_group_by_limit_query) now also applies to queries with aggregate functions in the projection when the GROUP BY top-K optimization does not apply, for example above query_plan_max_limit_for_top_k_optimization: once any aggregation thread exceeds n + offset distinct keys, all threads are restricted to a single shared set of kept keys, so the hash tables stay small while the aggregate values of the returned keys remain exact, with up to 7-9x speedups. The optimization no longer fires for queries with DISTINCT, QUALIFY, LIMIT AFTER, LIMIT UNTIL, window functions, arrayJoin, or exact_rows_before_limit, where it could return wrong or inexact results. #114690 (Alexey Milovidov).
• When max_threads is not set explicitly, a query is planned with the fair share of CPU slots of the concurrency control instead of all CPU cores. This increases the throughput and lowers the memory usage under many concurrent queries, while a query on an idle server still uses all cores. It can be disabled with the setting use_concurrency_control_for_max_threads. #125085 (Alexey Milovidov).
• A single-level GROUP BY result smaller than max_block_size is now split into several blocks when it was aggregated by several threads, so the steps after it, such as the probe of a following join, run in parallel instead of on one thread. Previously such queries could be several times slower, most visibly under compatibility older than 26.8, which disables enable_parallel_single_level_merge and enable_adaptive_aggregator. The result of a GROUP BY on a single 8- or 16-bit integer key is split the same way. #121711 (Pratham Nayak).
• Restored the batched hash-and-prefetch insert path for uniq, uniqCombined, and uniqCombined64 in ordinary GROUP BY queries, where they had silently fallen back to a per-row loop and were up to 46% slower. Closes #120531. #120578 (Groene AI).
• Faster GROUP BY over keys that have to be serialized, such as several columns that include a String. #122857 (Samuel Krempaský).
• GROUP BY over multiple LowCardinality keys now serializes the keys from the per-block dictionaries instead of copying them into full columns first, reducing CPU time and memory usage of aggregation queries over LowCardinality dimensions. #121795 (artemkulyk).
• Speed up GROUP BY over several LowCardinality keys when the same key tuples repeat, by caching aggregate states per tuple of dictionary positions in the serialized aggregation method. #121998 (artemkulyk).
• GROUP BY over a single String column with only count is faster when the grouping state does not fit in the CPU cache, especially with long strings. #123655 (Groene AI).
• GROUP BY queries with many groups that end in ORDER BY count() ... LIMIT n and use no other aggregate function are slightly faster when they group by a String column or by several columns. #124277 (Groene AI).
• Share the GROUP BY top-K skip boundary between aggregation threads. #116150 (Konstantin Bogdanov).
• Allow using fused GROUP BY top-K with a serialized query plan. #116163 (Konstantin Bogdanov).
• Added a query-plan optimization, controlled by the new setting query_plan_aggregation_having_prefilter, for HAVING count() <comparison> <constant> over a GROUP BY. While a two-level bucket of the aggregation result is converted to chunks, a group whose count() cannot satisfy the bound is now skipped before its key columns are materialized, instead of being materialized and then discarded by the filter above. This speeds up “groups above a threshold” queries over a high-cardinality GROUP BY with wide keys, where the discarded keys are most of the conversion. A skipped group is neither filtered nor finalized, so a HAVING conjunct written before the bound, and a sibling aggregate’s finalization, stop being evaluated on the groups the bound rejects: a query that raised an exception from one of those can now return rows instead. Setting query_plan_aggregation_having_prefilter = 0, or compatibility to a version below 26.10, keeps the previous behavior. #120510 (Groene AI).
• optimize_aggregators_of_group_by_keys now works for LowCardinality GROUP BY keys, so a HAVING predicate such as HAVING min(s) = 'x' is pushed down to storage and can use skip indexes, as for a plain key. It also eliminates anyHeavy, quantileExact, quantileExactLow, quantileExactHigh, groupBitAnd, and groupBitOr of a GROUP BY key. #110059 (Groene AI).
• -If aggregate functions with one Nullable argument no longer allocate a buffer the size of the whole block for every group with optimize_aggregation_in_order or for every window frame. With small groups or frames this makes them 1.7 to 3.1 times faster on x86_64. #122505 (Groene AI).
• Speed up -ForEach aggregate functions such as sumForEach and avgForEach: rows are processed in batches instead of calling add for each row, and an array of two or more elements is passed to the nested function in one call instead of one call per element. Grouped multi-argument -If aggregate functions with a Nullable argument or condition are also processed in batches. #121137 (Valery Petrov). #124207 (Groene AI).
• Speed up serialization of AggregateFunction state columns with a fixed-size state, such as sum, count, avg, min, max, and any, and their -If, -Array, -Merge, and Nullable variants. This affects writes to AggregatingMergeTree, the Native format, distributed queries, and spilling aggregate states to disk. #113962 (Manuel).
• Speed up reading and merging of uniq aggregate function states, such as uniqMerge over an AggregatingMergeTree, by 10-35% depending on state sizes and merge shape. #111493 (Manuel).
• Faster merging of uniqCombined and uniqCombined64 states with many distinct values into empty states, e.g. in -Merge combinators and AggregatingMergeTree: up to 6x faster. #124389 (Manuel).
• Reduce memory usage and improve performance of external DISTINCT by avoiding spills for duplicate-heavy input, writing smaller spill blocks, budgeting tail spilling before the final merge, preserving constant payloads, combining chunks before sorting, and retaining fewer rows while restoring order with LIMIT. #121772 (Nihal Z. Miaji).
• Faster sorted-stream processing on a multi-column key whose leading column has long runs, in LIMIT BY, negative LIMIT BY, DISTINCT in order, full_sorting_merge and partial_merge JOIN, and ReplacingMergeTree FINAL without vertical FINAL: the run of a key prefix is now found once instead of again for every group in it. The query from #123700 is about 3.5x faster. #123808 (Groene AI).
• Short queries with many keyed aggregations or scalar subqueries over large constant-folded literals are faster because hash-table-stats cache keys no longer hash whole variable-size constant values. #122738 (Nikita Taranov).
• Reduce memory usage of some window functions by precomputing peer-group boundaries instead of keeping earlier blocks in memory. #123737 (Mikhail Artemenko).
• A filter on the PARTITION BY columns of a window function is no longer pushed below the window (this reverts #110114). The push-down made queries up to 8 times slower when the window reads a JOIN and the filter removes every row, for example IN over an empty subquery, because the join built its right side before the filter could stop the query. Closes #123875. #125463 (Groene AI).
• hash and parallel_hash joins now copy right-side output columns directly from the stored blocks instead of calling the stored column object once per row. This speeds up every column type, including String, Array, Tuple, Nullable, Variant, Map, LowCardinality, JSON, Dynamic, AggregateFunction, and QBit. #117460 (Harikrishnan Prabakaran).
• Speed up hash join output construction in the row-major layout: output reconstruction is up to 2x faster for joins with many rows per key and about 1.5x for wide or Nullable payloads. The row-major layout is now also skipped for payloads too narrow to benefit from it. #121981 (Hechem Selmi).
• Hash joins with many matches per key are faster when they output a LowCardinality column whose blocks have different dictionaries (several parts, FINAL, or a LowCardinality value computed per block): short ranges from another dictionary are now inserted key by key instead of through temporary columns. Such joins run 4-9x faster, at the 26.6 level or better. #123900 (Groene AI).
• RIGHT and FULL joins using hash, parallel_hash, or grace_hash skip the search for unmatched right rows when every right key has matched. With 50 million right keys, that search was about 10% of query time for hash (356 ms of 3.6 s) and about 15% for parallel_hash (106 ms of 0.7 s). #124775 (Groene AI).
• A full_sorting_merge join now advertises that its result is ordered by the join key, so a sort above it, the pre-sort of the next merge join in a chain or an ORDER BY, merges the already sorted streams instead of sorting from scratch. Joining several tables on the same key with full_sorting_merge no longer performs one full sort per join. Also fixed the interplay with query_plan_join_shard_by_pk_ranges, which could otherwise fail with a logical error or return wrong results for such a chain. #117786 (Alexey Milovidov).
• A join step that has a condition of its own now also joins on equalities implied through other tables, so for a.x = b.x AND b.x = c.x, joining a with c also uses a.x = c.x. With query_plan_optimize_join_order_algorithm = 'dpsize', this makes TPC-H Q02 join its correlated subquery to partsupp on the part key instead of on the supply cost alone. Closes #122396. #122445 (Groene AI).
• A WHERE condition that no row satisfies is pushed to every JOIN input again, including the non-preserved side of an OUTER JOIN and a side with no column used above the join, so that input is no longer read and hash-built for nothing. For example, a UNION ALL branch whose constant column fails the outer WHERE is skipped as in 26.6. Closes #123702. Closes #123878. #123829 (Groene AI).
• Constant comparison, logical, and null (IS NULL / IS NOT NULL) filters over materialized constants are now folded and pushed down, reducing unnecessary reads in WHERE, PREWHERE, and UNION branches. Always-true WHERE filters are replaced with expressions so join reordering still sees all joined tables as one graph. #105987 (Yarik Briukhovetskyi). #124056 (Yarik Briukhovetskyi).
• OUTER JOIN queries can now be faster: when a WHERE clause or another JOIN later in the query would discard the unmatched rows anyway, ClickHouse converts the join to an INNER JOIN. The new setting query_plan_convert_outer_join_to_inner_join_transitively controls this, and only takes effect when query_plan_convert_outer_join_to_inner_join is enabled. #120253 (Hechem Selmi).
• Restored filter push-down over ANY INNER JOIN, so primary-key pruning works again and such queries no longer read much more data than necessary. Because ANY keeps an arbitrary matching row, SELECT ... ANY INNER JOIN ... WHERE ... can now return a different matching row than before. #120765 (Kirill Shokhin).
• optimize_and_compare_chain again derives transitive conditions through a DateTime or lower-scale DateTime64 constant compared with a DateTime64 expression, as it did before 26.8. For example, t1.ts < toDateTime('2026-09-16 00:00:00') AND t1.ts >= t2.ts adds t2.ts < toDateTime('2026-09-16 00:00:00'), so the other side of a join can be pruned by its partition and primary key instead of being read in full. Closes #123706. #123796 (Groene AI).
• Improve join order selection when column statistics are missing: an equi join without NDV statistics is now treated more like a foreign-key-to-primary-key join instead of a cross product, and the greedy join-order algorithm prefers equi-connected pairs over pairs linked only by non-equi predicates. #120176 (Vladimir Cherkasov). #121075 (Alexey Milovidov).
• Improves JOIN cardinality estimation by propagating join-key distinct-value estimates according to join semantics. Outer joins retain the preserved-side estimate, anti joins keep their input estimates, and semi joins narrow only the output side, preventing underestimation that could lead to suboptimal query plans. #122305 (Christoph Viebig).
• IEJoin now treats opposite-direction conditions on unrelated columns as independent conditions instead of as a band, which avoids choosing a join-key pair that passes thousands of times more rows to the residual condition. #120213 (Vladimir Cherkasov).
• Function firstNonDefault, which computes the USING columns of FULL JOIN, now copies whole ranges of rows instead of making two virtual calls per row. Queries that use a USING column of a FULL JOIN after the join, for example in WHERE, are faster. #123302 (Groene AI).
• Enabled enable_join_runtime_filters_index_analysis by default, so join runtime filters can prune granules during the probe phase. #121859 (Robert Schulze).
• Join runtime filters now run after the query’s own PREWHERE conditions also for tables without column statistics, as they already did with statistics. Before, a runtime filter on a well-compressed join key could run first and read that key for every row, so enabling enable_join_runtime_filters could make a query read several times more data than without it. #122452 (Groene AI).
• The granule pruning of enable_join_runtime_filters_index_analysis now also applies when the left side of a JOIN is read from a normal projection. The setting description states that the pruning does not happen with parallel replicas, a distributed query plan, or FINAL. #119041 (Alexey Milovidov).
• The query condition cache now records the static WHERE conditions of a join whose PREWHERE also holds a join runtime filter, so a repeated join skips the granules those conditions exclude instead of reading them again. This helps whenever the runtime filter cannot skip those granules through an index itself, for example when no primary key or skip index covers the join key, or when enable_join_runtime_filters_index_analysis is disabled. #122380 (Groene AI).
• Repeated insertion of empty and single-element Array values into in-memory columns is now faster. This happens, for example, during joins. #117321 (Minh Vu).
• With use_top_k_dynamic_filtering = 1 (the default), a SELECT ... ORDER BY ... LIMIT n could read far more data than with the setting off: the internal __topKFilter took over the read’s PREWHERE slot and displaced the query’s own condition, and an explicit PREWHERE disabled the feature outright. The threshold filter now shares PREWHERE with the other conditions and takes the first slot. It is also no longer installed on a read that a sorting projection already delivers in ORDER BY order, where it forced a full scan. #117968 (Groene AI).
• Granules fully filtered by the dynamic threshold of an ORDER BY ... LIMIT n read are now recorded in the query condition cache, so repeat runs of such queries skip them at the mark-selection stage instead of re-reading and re-filtering most of the table. On the ClickBench Q24/Q26 shapes (SELECT SearchPhrase FROM hits WHERE SearchPhrase <> '' ORDER BY EventTime LIMIT 10), a warm run drops from about 95 million rows / 12000 granules read to about 1.3 million rows / 160 granules, and from about 15-30 ms to about 8 ms on a 96-core machine. This now also works for ORDER BY ... LIMIT n queries over Parquet files of File tables, with or without a WHERE clause: a repeated query reads only the row groups which can hold rows of its result. #114659 (Alexey Milovidov). #122064 (Alexey Milovidov).
• Improve performance of ORDER BY ... LIMIT queries that use minmax-index top-K pruning (use_skip_indexes_for_top_k) on large tables with many granules. #116512 (Anton Popov).
• Queries with FINAL that sort in reverse order of the sorting key, such as SELECT ... FROM t FINAL ORDER BY key DESC LIMIT n, now read in order through Merge tables over ReplacingMergeTree tables too, and the topKThroughJoin optimization defers to reverse-order reads below a LEFT JOIN when the table supports them. This extends #115345. #118008 (Alexey Milovidov).
• Fixed a performance regression in ORDER BY merge sort: queries with a sort followed by LIMIT no longer do redundant per-output-block work in result-block reservation. #108179 (Groene AI).
• Setting optimize_move_to_prewhere_if_final is now enabled by default, so queries with FINAL benefit from the automatic PREWHERE optimization. Only deterministic conditions over the sorting key are moved, and query results are not affected. #121350 (Yarik Briukhovetskyi).
• With short-circuit evaluation, non-throwing arguments of and and or are now ordered by measured cost and selectivity, so cheap filters can avoid evaluating expensive expressions on excluded rows. Controlled by short_circuit_function_evaluation_reorder_arguments, enabled by default. #122508 (Alexey Milovidov).
• Fix a performance regression in short-circuit evaluation of and and or by avoiding expansion of intermediate results that are only used to update the logical mask. #120520 (LB7666).
• On a table with no applicable data skipping index, a query no longer stores the same predicate under two query condition cache keys. This halves the number of O(marks per part) cache entries such a query allocates, and removes one cache lookup per part. Queries with distributed_index_analysis enabled keep both keys. #116259 (Groene AI).
• SELECT ... FINAL with a selective skip-index plan is much faster on tables with many granules: the use_skip_indexes_if_final_exact_mode recovery pass now iterates ranges instead of individual granules. #121799 (Nikolai Kochetov).
• Improve primary index analysis in queries that filter both by the primary key and by other columns, for example WHERE timestamp BETWEEN ... AND level = 'error' where level is not in the primary key. #123535 (Anton Popov).
• IS NULL and IS NOT NULL predicates can now prune MergeTree parts using the NULL row count tracked by basic statistics on Nullable columns. #116697 (zoomxi).
• A set skip index on a Nullable column again skips granules whose stored set contains NULL but not the searched value for IN and = conditions, restoring the pruning lost in 26.8. Closes #123705. #123778 (Groene AI).
• Skip indexes (text, bloom_filter, tokenbf_v1, ngrambf_v1, minmax) and the primary key are now used when the indexed column or the searched constant is wrapped in a conversion that does not change its value, such as toLowCardinality, toNullable, or a CAST to Nullable or LowCardinality. For example, m['key'] = 'value' on Map(String, LowCardinality(String)) and hasToken(s, toLowCardinality('x')) no longer scan all granules. bloom_filter and text indexes also work when a Nullable indexed column is cast to the non-Nullable type, such as CAST(x, 'String') or json.path::String. #120220 (Anton Popov). #123239 (Anton Popov). #124220 (Anton Popov).
• Speed up planning of queries with skip indexes on tables with many parts and pending patch parts of lightweight updates. #124931 (Anton Popov).
• Speed up queries with multiple PREWHERE conditions on AMD Zen 1 and Zen 2 CPUs. #120283 (Manuel).
• Analysis of a query over a Merge table no longer traverses every child table for each query-tree node when optimize_functions_to_subcolumns is enabled. The answer is now reused across the query tree, so analysis time grows with the query size plus the number of child tables instead of their product. Closes #113202. #113367 (Groene AI).
• Queries to the merge table function and creation of Merge tables are faster when the source tables have many columns: combining the source columns no longer takes quadratic time in the number of columns. #124115 (Groene AI).
• Faster analysis of queries to tables with many ALIAS columns, such as system.metric_log: the expression of an ALIAS column is analyzed only if the query uses that column. #123926 (Alexey Milovidov).
• Reduce aggregate projection analysis overhead for tables with many subcolumns. #110979 (Rory Shanks).
• A condition such as name IN ('a', 'b') on the system tables that enumerate tables (system.tables, system.columns, system.parts, system.replicas, and others) now narrows the enumeration instead of walking every table of every database, the same way an equality condition already did. This is most visible for databases whose table list comes from a remote system, such as PostgreSQL, SQLite, and the DataLake catalogs. #117589 (Alexey Milovidov).
• Added the external_storage_push_down_limit setting to reduce data read from external databases for safe LIMIT queries. Also fixed wrong results when LIMIT plus OFFSET does not fit in UInt64: numbers, system.numbers, generate_series, primes, system.primes, and the PostgreSQL table engine returned no rows for the common “no limit” idiom LIMIT 18446744073709551615 OFFSET n, and a LIMIT over a JOIN with numbers-like sources returned too few rows. #115238 (Alexey Milovidov). #118131 (Groene AI).
• Push filters into the query fragment executed by parallel replicas so replicas can filter rows and prune data before sending results to the initiator. This also allows join runtime filters to reach the relevant reads. Controlled by parallel_replicas_filter_pushdown, now enabled by default; SET parallel_replicas_filter_pushdown = 0 restores the behavior before 26.10. #117373 (Nikita Taranov).
• With parallel replicas, optimize_aggregation_in_order no longer disables projections: aggregate and normal projections, and _minmax_count_projection for count() with a partition predicate, are used as with optimize_aggregation_in_order = 0. Closes #123061. #123098 (Groene AI).
• Parallel replicas now apply enable_join_runtime_filters_index_analysis to prune probe-side granules locally on each replica, reducing JOIN read work. #123116 (Nikita Taranov).
• Automatic parallel replicas no longer distributes a query when most of its reading would be repeated in full by every replica rather than split between them. The limit is set by the new setting automatic_parallel_replicas_max_duplicated_read_ratio (default 0.5; 1 disables it). #123332 (Nikita Taranov).
• Automatic parallel replicas for multi-read queries now reuse local plan index analysis in the parallel-replicas plan, reducing extra reads when parallel replicas are enabled. #115153 (Nikita Taranov).
• Improve reuse of collected query statistics by automatic parallel replicas for equivalent local and distributed plans. #116781 (Nikita Taranov).
• Automatic parallel replicas now account correctly for the size of aggregated results on more aggregation paths, so the cost model no longer treats shipping aggregate states as free and is less likely to choose parallel replicas when a query is cheaper locally. #119499 (Nikita Taranov).
• Automatic parallel replicas now collect planning statistics for the table read that is actually split between replicas, including after join reordering and with lazy materialization, improving the decision to distribute a query. #122906 (Nikita Taranov).
• Avoid building the parallel replicas candidate plan when cached statistics already show the query should stay on a single replica. #124983 (Nikita Taranov).
• Added quick checks to skip automatic parallel replicas optimization for queries unsupported by parallel replicas, including keyless aggregations whose aggregate state owns memory and therefore never produces the probe statistics the optimization needs. #120861 (Nikita Taranov). #125470 (Nikita Taranov).
• Automatic parallel replicas (automatic_parallel_replicas_mode) can now be applied to queries with ORDER BY ... WITH FILL. #123929 (Alexey Milovidov).
• Queries that search for several tokens at once with a text index, such as hasAllTokens, are now faster. The automatic choice of algorithm for intersecting posting lists is improved too, and is controlled by the new setting text_index_postings_intersection_algorithm. #121038 (Anton Popov).
• Queries that use a text index read large posting lists faster and with fewer disk operations. The improvement is most noticeable for searches for frequent tokens in big tables. #121040 (Anton Popov).
• A query that searches a text index and also selects a column no index covers no longer reads that column for every row of every granule the index selected. On a 1M-row table with a 1055-byte payload and 64 matching rows, SELECT ts, payload FROM t WHERE hasToken(message, 'needle') now reads 578 KiB instead of 532 MiB. #120517 (Groene AI).
• A query that reads a text index no longer asks the filesystem for each index substream’s on-disk name and size once per part; both are now taken from the part’s checksums, which already hold them. On an 8M-row table split into 400 parts, SELECT count() FROM logs WHERE hasToken(message, 'needle') drops from a median 12 ms to 9 ms, with 42% less system time. #120875 (Groene AI).
• Speed up queries that filter by a text index together with other PREWHERE conditions, such as the dynamic filter of ORDER BY ... LIMIT queries. #124880 (Anton Popov).
• Speed up LIKE, ILIKE, startsWith, and endsWith evaluated by a text-index dictionary scan (use_text_index_like_evaluation_by_dictionary_scan): the scan now seeks the sorted dictionary to the block range an anchored pattern can match, and tests a block’s tokens for the pattern’s mandatory literal in one pass instead of running the matcher on every token. #120495 (Groene AI).
• Repeated broad LIKE and ILIKE queries over a text index no longer rescan the index dictionary of every part only to abandon the scan again: the decision is cached (setting use_text_index_pattern_bypass_cache). #113572 (Rory Shanks).
• A text index with the splitByNonAlpha or array tokenizer now answers single-needle multiSearchAny, multiSearchAnyUTF8, multiSearchAnyCaseInsensitive, and multiSearchAnyCaseInsensitiveUTF8 with the same dictionary scan that LIKE '%needle%' uses, so a needle inside a word, such as PointerExc in NullPointerException, no longer reads the whole column. #124520 (Groene AI).
• multiSearchAny over a map element can now use a mapValues(...) ngram or token text index instead of being rejected by index analysis. #121787 (Groene AI).
• Optimize SELECT count() queries answered from a text index with a compressed posting list codec (bitpacking, pfor). #123294 (Anton Popov).
• Reduced the CPU overhead of reading and intersecting text index posting lists in lazy apply mode. #123786 (Anton Popov). #123873 (Anton Popov).
• Merges of tables with a text index are now faster. Pages of a source part’s row ids that hold consecutive values are stored as a first value plus a count, so they are neither bit-packed nor decoded per posting row id and need no storage for packed values. #120982 (Groene AI).
• Merging positional text indexes is faster: position lists are now merged segment by segment instead of being materialized and sorted in a temporary roaringish structure. #121976 (Elmi Ahmadov).
• Added the pfor codec for text-index posting lists. It improves compression ratio and decoding speed. Enable it per index with posting_list_codec = 'pfor', or for all text indexes of a table with the text_index_posting_list_codec setting. #118123 (Elmi Ahmadov).
• Building text indexes during INSERT is now faster, especially for indexes with the bitpacking codec. #113856 (Anton Popov).
• Building a text index over a LowCardinality(String) or Array(LowCardinality(String)) column is faster: each distinct value is tokenized once per batch of rows instead of at every occurrence. MATERIALIZE INDEX over a column with 50 distinct values takes about 33% less time. #122409 (Groene AI).
• Improve performance of the ngrams tokenizer. This speeds up tokens(s, 'ngrams', N), ngrams(s, N), hasAnyTokens, hasAllTokens, and hasPhrase with the ngrams tokenizer, as well as inserts into tables with an ngrambf_v1 or text(tokenizer = ngrams(N)) index. tokens(URL, 'ngrams', 3) over ClickBench URLs takes 26% (aarch64) to 36% (x86_64) less time. #122328 (Groene AI).
• Reading columns and subcolumns of MergeTree tables no longer rebuilds the serialization of a column for every reader: the serialization the data part already holds is reused. The gain is largest for JSON columns with many typed paths, where the whole set of path serializations used to be rebuilt per read task and per filter step. #120632 (Pavel Kruglov).
• Reduce per-block overhead when reading MergeTree parts with small blocks: column-stream names and substream-cache keys are no longer rebuilt for every block, and reads skip patch-part handling when there are none, virtual-column name lookups, and some small-column reallocations. This makes queries with a low max_block_size faster. #124816 (Alexey Milovidov). #124817 (Alexey Milovidov).
• Reduce CPU overhead of reads from MergeTree tables by not formatting a tracing span name for every block. #124820 (Alexey Milovidov).
• Reading many columns from Compact parts is up to 2% faster by avoiding per-call copies of column metadata. #122840 (Samuel Krempaský).
• Restore parallel prefix deserialization when reading an Object or JSON column from a Wide part: marks loading no longer queues behind one mutex. Closes #114136. #114252 (Groene AI).
• Queries that read subcolumns of JSON columns with many typed paths are faster: resolving a subcolumn by name visits only the typed paths that can match it. For a JSON column with 500 typed paths, reading subcolumns from 96 parts takes about half the time. #122029 (Rory Shanks).
• Reduce serialization construction when reading JSON subcolumns from MergeTree parts by reusing safe cached subcolumn serializations. #110957 (Rory Shanks).
• Improve nested JSON path read performance with map_with_buckets by reusing thread-local serializations for Dynamic values. #113344 (Rory Shanks).
• Writing JSON columns with many paths in shared data is faster: the writer no longer rebuilds the statistics of the column for every block. For merges of parts with small granules, this saves up to 13% of the merge time. #121987 (Rory Shanks).
• Make constructing a JSON data type that uses SKIP REGEXP cheaper. The patterns are now compiled once per type instead of once for every declared path, so a column declaring 183 typed paths and 6 SKIP REGEXP entries performs 6 regular expression compilations where it previously performed 1098. This matters most for wide JSON schemas, where the patterns are often long alternations that are expensive to compile. #122941 (arc1999).
• Read only the matching Map subcolumn for safe LIKE predicates over Map values. #119809 (Minh Vu).
• Improve SELECT performance by avoiding lock contention when getting mutation snapshots with many concurrent background fetches. #102382 (c-end).
• Support PREWHERE (including the automatic move of WHERE conditions by optimize_move_to_prewhere) for Memory tables: only the columns of the conditions are read at first, and the remaining columns are read only for the blocks where some rows pass, and only for the passing rows. This is especially beneficial for tables with SETTINGS compress = true, because for a selective condition most columns are never decompressed. Additionally, SELECT count() FROM table on a Memory table is now served from metadata, and system.columns shows real per-column sizes for Memory tables. #116248 (Alexey Milovidov).
• PREWHERE over a Memory table is evaluated in multiple steps, like for MergeTree (enable_multiple_prewhere_read_steps): the columns of each next condition are read only for the rows that passed the previous ones. For a table with compress = true, a condition on a wide column next to a selective one no longer decompresses that column for every block, which makes such queries up to several times faster. #122604 (Alexey Milovidov).
• Queries with ORDER BY ... LIMIT n over Memory tables use TopN dynamic filtering: rows that cannot get into the result are dropped while reading, before the other columns are read or decompressed with compress = true. #123787 (Alexey Milovidov).
• Queries whose only aggregate functions are min and max of columns of Parquet files read with the file table function or the File engine, without GROUP BY or filters (for example ClickBench Q7), now take the values from the column chunk statistics instead of reading the data. Controlled by the new setting optimize_min_max_from_files. #124201 (Groene AI).
• ORDER BY ... LIMIT over a single Parquet file reads first the row groups whose min/max statistics are best for the result, so the top-K threshold skips the other row groups: SELECT URL FROM file('hits.parquet') ORDER BY UserID DESC LIMIT 10 skips 207 of its 226 row groups at 4 threads. New profile event ParquetTopKSkippedRowGroups. #123639 (Groene AI).
• Row group pruning by dictionaries in the Parquet reader is faster for an equality or a short IN list on a column with large dictionaries, such as WHERE UserID = ... over ClickBench hits.parquet: the hashes of the dictionary values are no longer sorted unless a column chunk is checked against many values. #123579 (Groene AI).
• Parquet readers now check a column chunk’s bloom filter before reading its dictionary page. A row group that the bloom filter rules out no longer reads the much larger dictionary page to reach the same conclusion, so input_format_parquet_dictionary_filter_push_down costs nothing on files that carry both, while pruning as much as before. #120907 (Alexey Milovidov).
• Parquet footer-size reads are now tuned by storage type: remote and object-storage reads use more optimistic footer-size estimates to fetch the footer in one read instead of two, reducing footer-read latency, while local-file reads again use a 64 KiB tail so short queries over many local Parquet files read less data. #115686 (UnamedRus). #124883 (Groene AI).
• Avoid a redundant object-storage GET request when reading random-access formats from object storage. The generic from-start read-ahead is no longer issued for files larger than a single read buffer, where it could not reach the footer. Also, input_format_allow_seeks = 0 is now honored by the Parquet reader, which previously sought regardless of the setting: with seeks disabled the whole object is read sequentially instead, which can raise memory usage for queries that set it. #115681 (UnamedRus).
• Speed up writing Parquet bloom filters: instead of deduplicating the value hashes in a hash set to determine the filter size, the writer now builds a filter sized for the number of values and folds it down to the size implied by output_format_parquet_bloom_filter_bits_per_value. This removes the worst-case slowdown on data that is adversarial for the hash set, and lowers the worst-case memory usage. #119230 (Armin Braun).
• Parse some GeoParquet metadata lazily. #122693 (Konstantin Vedernikov).
• Snappy decompression on AArch64 now uses NEON for short-offset copies, as upstream snappy does by default, which makes reading Snappy-compressed Parquet with repetitive string columns up to ~9% faster on ARM servers. The ARMv8.0 compatibility build (NO_ARMV81_OR_HIGHER) is unchanged. #124608 (Groene AI).
• Reading CSV with input_format_csv_trim_whitespaces = 0 no longer prevents the parser from scaling across CPU cores. Each field used to copy a shared_ptr shared by all parsing threads, turning its reference counter into a contention point. On a 96-core machine, reading 200 million rows became 4.6 times faster and used 23 times less CPU. #117251 (Alexey Milovidov).
• Iceberg partition pruning now uses the partition value of a day, month, year, or hour transform for a filter that wraps the partition source column in a monotonic function, such as WHERE toDate(ts) = '2024-01-20', including for partition values from before 1970. Closes #103433. #118606 (Konstantin Vedernikov). #119302 (Alexey Milovidov).
• Faster planning of queries to Iceberg tables with many manifest files: when a manifest file header has the schema-id key and that schema is already known from metadata.json, the schema copy stored in the manifest header is no longer parsed and compared with it. ClickHouse now also writes the schema-id key into the headers of manifest files it creates. #122574 (Konstantin Vedernikov).
• An INSERT into an Iceberg table that loses the commit race now reuses its manifests on retry and rewrites only the manifest list and the metadata file, instead of regenerating all manifests. #122096 (Asya Shneerson).
• Restore the summary-based trivial count() optimization for Iceberg tables when the snapshot summary contains a safe row count. The optimization is skipped when equality deletes are present, because their recorded count is not the number of deleted rows. #123267 (Konstantin Vedernikov).
• Iceberg min/max hyperrectangles are now built only for columns the filter uses. #122681 (Konstantin Vedernikov).
• Restore min/max pruning of Iceberg data files for a Decimal(38, 38) column. A manifest bound of magnitude above roughly 0.7 could not be widened inside Int128 and was dropped, leaving the column without a min/max condition; it is now saturated at the largest magnitude the precision allows, which is still outside every value the file can hold. #119712 (Alexey Milovidov).
• Queries to system.iceberg_history with filters on table now avoid listing unrelated data lake catalog namespaces. #113867 (Den Kalantaevskii).
• Improve HTTP upload and download efficiency for S3, Azure, the url table function and engine, dictionaries, and data lake catalogs. Uploads use 1 MiB chunks instead of 8 KiB chunks, reducing the number of system calls for a 15 MB body from 1817 to 15 and avoiding redundant copies. Reads from Azure also avoid an intermediate 8 KiB buffer. #116222 (Alexey Milovidov).
• Reading from object storage with remote_filesystem_read_method = 'read' now keeps full-size network reads. Before, a short read at the end of one range capped every later read of the same file, which could make a scan hundreds of times slower, for example a compact part on an encrypted S3 disk without the filesystem cache. #123777 (Groene AI).
• Reads that are served from the OS page cache no longer consume the budget of the local read bandwidth throttler (max_local_read_bandwidth_for_server and max_local_read_bandwidth), so a query over cached data is not slowed down to the configured disk bandwidth. This works with the default local_filesystem_read_method = 'pread_threadpool', with pread, and for the files of the filesystem cache; the read method cannot detect cached reads and is accounted as before. #118495 (Alexey Milovidov).
• Populating the filesystem cache on a cache miss is faster: the writer now reserves space ahead instead of taking the cache lock on every write. #123255 (Kseniia Sumarokova).
• Loading the metadata of a large s3_plain_rewritable disk is several times faster. Previously it made one ListObjects request per directory and enumerated the __meta directory by a single paginated listing, so a disk with tens of thousands of directories needed tens of thousands of sequential-ish requests. The listings are now split by key prefix and run in parallel, and the files of all directories are collected by a few listings of the whole disk. For a disk with 36 thousand directories holding half a million objects, the load time dropped from 116 to 37 seconds. #121412 (Alexey Milovidov).
• Metadata transactions of unrelated tables on a plain_rewritable disk no longer wait for each other: the disk-wide lock used to be held across all object storage requests of a commit, so one long MOVE PARTITION, DROP TABLE, or part removal blocked every merge, insert, and CREATE TABLE on the same disk. The lock is now taken per affected path, and the in-memory metadata is updated by replaying the recorded changes. Also fixes a permanent leak of objects when the server was killed during DROP TABLE (or another removal) on a plain_rewritable disk: the leftovers of such a removal are now marked and deleted when the disk metadata is loaded on the next start. #118490 (Alexey Milovidov).
• Removing data parts from a disk with metadata_type = plain_rewritable (cleanup of outdated parts, TRUNCATE, DROP PARTITION, DROP TABLE) no longer copies every file of the part to a temporary object before deleting it, and no longer holds the disk metadata lock for those copies. The part directory is removed with one batched delete request. Closes #123275. #123303 (Groene AI).
• Reduce metadata update cost and lock contention for s3_plain_rewritable disks with many parts by sharing unchanged directory entries between snapshots. #119625 (Rory Shanks).
• Avoid object reads for unchanged directories when a plain-rewritable disk refreshes its metadata. #119760 (Rory Shanks).
• OPTIMIZE TABLE ... FINAL on a non-replicated MergeTree table now assigns and runs the merges for all partitions at once instead of processing them one partition at a time. #109004 (Alexey Milovidov).
• TTL DELETE merges can now use the vertical merge algorithm on ReplacingMergeTree, CollapsingMergeTree, and VersionedCollapsingMergeTree, where vertical_merge_optimize_ttl_delete previously applied only to plain MergeTree. #119886 (Murphy).
• Allow vertical merges to produce Wide parts with Packed storage instead of forcing the horizontal merge algorithm. #125760 (alesapin).
• Improved merge performance for SummingMergeTree tables by skipping unnecessary aggregate-column materialization. #120852 (Eduard Karacharov).
• Many queued lightweight deletes with an IN condition that reach the same part in one mutation are now applied together, instead of evaluating each IN predicate over the whole part separately. Controlled by the new MergeTree setting enable_row_mask_update_coalescing. #122718 (Sergey Kuznetsov).
• Optimize the insert backpressure path in MergeTree by short-circuiting the parts-count check. #114754 (johnjing).
• Reduced merge-selection work when all storage volumes allow merges. #119782 (Rory Shanks).
• Reduce merge-selection CPU overhead for tables with column TTLs by avoiding copies of TTL descriptions. #119644 (Rory Shanks).
• Speed up decompression of codec chains. The first codec now reads the compressed block in place and the last codec writes the result in place, unless that codec requires a padded buffer (LZ4, ZXC). #120043 (Raufs Dunamalijevs).
• Faster decompression of T64-encoded columns. #120414 (Raufs Dunamalijevs). #122442 (Raufs Dunamalijevs).
• A query with arrayJoin and a small LIMIT again shrinks the read block size instead of reading whole blocks on every thread for a few output rows, fixing a regression from 26.6. Closes #119193. #120214 (Yarik Briukhovetskyi).
• Chained ARRAY JOINs no longer materialize the whole block of a lazily replicated array column before expanding it, which cuts their peak memory by several times. #121397 (Yarik Briukhovetskyi).
• Avoid copying whole columns when slicing them, including in ARRAY JOIN. #118788 (Nikita Mikhaylov).
• Improve ARRAY JOIN performance for queries with wide columns that are not expanded, by avoiding unnecessary copies and preserving lazy replication through filtering and sorting. #124727 (Yarik Briukhovetskyi).
• Reduced the peak memory usage of RESTORE and of incremental BACKUP: the metadata file of a backup is now parsed while it is read, instead of being read into memory whole first. For backups of millions of files this saves gigabytes. #125712 (Julia Kartseva).
• Faster startup of clickhouse-local and other tools. Time zone lookup tables are built 2.4 to 5.6 times faster and use half the memory (2.3 MiB instead of 4.6 MiB per loaded time zone). The clickhouse binary also starts faster when it is not in the page cache, with a 1.40x speedup measured for clickhouse local. #125665 (Raúl Marín). #125672 (Raúl Marín).
• A full scan of a complex-key hashed, sparse_hashed, hashed_array, or range_hashed dictionary (SELECT ... FROM <dictionary> or the dictionary table function) is faster: each reading stream now deserializes only the keys of its own block instead of one thread deserializing all keys before reading starts. A scan with LIMIT deserializes only the blocks it reads. A full scan also no longer keeps all deserialized keys in memory at once. #123883 (Groene AI).
• Added the SIEVE cache eviction policy for in-memory caches such as the mark cache and the uncompressed cache. Enable it with the corresponding *_cache_policy server settings, for example mark_cache_policy or uncompressed_cache_policy. It is not available for the filesystem cache. #123435 (Alexey Milovidov).
• The symbolized stack trace of a failed query is now built only when the error message will be logged or the trace will be stored in system.query_log. Previously each failing clickhouse-local invocation built the symbol index of the whole binary to produce a string that was then discarded. #120660 (Groene AI).
• Reduces the CPU overhead of the workload scheduler thread by updating the system.scheduler introspection metrics more cheaply. A saturated scheduler can now serve about 15% more resource requests per second. #121958 (Sergei Trifonov).
• PromQL now copies label matchers across binary operators for the labels the operator matches series by, so b / on(job) a{job="x"} reads only the matching series of b. As in Prometheus, duplicate left series of a one-to-one match in a group the right side does not have no longer fail the query. The new setting promql_push_down_label_matchers turns this off. #122972 (Valery Petrov). #125267 (Groene AI).
• Accelerate exact PromQL label matching on TimeSeries tables with a text index on the tags map. #120083 (Nikita Mikhaylov).
• Use more compact codecs for generated timestamp and value columns in the samples and recent samples tables of the TimeSeries engine: Delta, T64, and ZSTD(3) for timestamp, and ALP plus ZSTD(3) for value. #120454 (Nikita Mikhaylov).
• Reduce memory usage of the timeSeries*ToGrid aggregate functions. #124773 (Valery Petrov).
• Filtering and sorting are now faster: the helpers that build filter masks, count filter bytes, and check for an identity permutation use AVX2 or AVX-512 on x86 instead of SSE2, and are vectorized on ARM where they were not. #121403 (Raúl Marín).
• Faster filtering of numeric columns on ARM (up to 1.28x) and on AMD Zen 4 (up to 2.8x), speeding up queries with WHERE and multi-step PREWHERE. #122893 (Raúl Marín).
• Filters on numeric columns that keep few rows are faster. #124024 (Raúl Marín).
• Filtering numeric columns no longer over-allocates the result buffer (by up to 2x) when the size of the result is known in advance. Closes #123546. #123576 (Raúl Marín).
• Vectorized comparisons of UUID, IPv4, and IPv6 columns with columns or constants of the same type, which makes such comparisons several times faster. Comparisons of FixedString(16) values are now branchless too, making them faster when it is unpredictable whether the values are equal. #120193 (Manuel).
• Comparisons <, >, <=, and >= of IPv6 and FixedString(16) values are 1.4 to 1.8 times faster on CPUs with AVX-512. #125151 (Groene AI).
• LIKE, ILIKE, match, and their variants are faster when the pattern is not a constant, up to 5 times with 8 threads. #122764 (Groene AI). #122843 (Samuel Krempaský).
• Faster multiMatchAny and related functions with non-constant needles when reject_expensive_hyperscan_regexps is disabled. #122842 (Samuel Krempaský).
• CAST of strings to Nullable(Bool) (also accurateCastOrNull(s, 'Bool') and toBool of a Nullable(String)) is up to 7 times faster with 8 threads, and arrayIntersect, arrayUnion, arraySymmetricDifference, and arrayJaccardIndex over non-numeric elements such as DateTime64, UUID, or Tuple are up to 18 times faster with 8 threads: they no longer allocate memory and update profile events for every row. multiMatchAny and the other multi-regexp functions with non-constant needles now add their cache hit and miss counts once per block. #122920 (Groene AI).
• Speed up extracting a value from a Map with LowCardinality keys. The key is now resolved to its dictionary position once instead of being compared against every entry of a row, so the cost of m[key] no longer grows with the position of the key in the row. Also fix extracting a value from a Map whose key is NaN, or holds a NULL inside a composite key: such a key used to compare equal to every other key. #120453 (Pavel Kruglov).
• L2DistanceTransposed, cosineDistanceTransposed, dotProductTransposed, and their ...Quantized variants with precision 1 are now computed directly from the packed sign bit plane with per-byte lookup tables, 2 to 3 times faster than before. The new setting qbit_one_bit_symmetric_distance additionally reduces the reference vector to its signs and derives the result from the Hamming distance between the two sign vectors, computed with XOR and popcount, which is up to 10 times faster. #120594 (Alexey Milovidov).
• Speed up randomHadamardTransform and vector quantization for vector dimensions of the form 2^k × 12 or 2^k × 20, such as 768 or 1536, on x86 with AVX2: 1.9x to 2.5x higher throughput and 50% to 60% lower median latency. #120874 (Jingxin (Philip) Li).
• hasAll and hasAny on integer arrays use SIMD on both x86 and ARM, including Nullable arrays and hasAny, which previously used a scalar loop. hasAny is up to 15x faster, and hasAll becomes several times faster on ARM. #123420 (Groene AI).
• hasAny over Array(String) looks each element up in a hash set, instead of comparing it with every string, when its second argument is a constant array of four or more strings of up to 256 bytes and a block has enough rows to pay for building the set. With 32 strings this uses 87% less CPU. Closes #122333. #122507 (Groene AI).
• Vectorize least and greatest for Nullable numeric arguments and for three or more numeric arguments, and clamp for numeric arguments. These cases were previously evaluated row-by-row and are now 3-7 times faster. #109783 (Manuel).
• Speed up round, roundBankers, floor, ceil, and trunc with a scale argument on Float32 and Float64 columns by using AVX2 on x86 instead of SSE4.1, and speed up rounding of integers narrower than 64 bits on ARM by vectorizing it. #122032 (Raúl Marín).
• modulo and moduloOrNull of two unsigned integers of different widths are faster again, as before 26.9, when the wider one is UInt128 (about 3 to 4 times, for example UInt128 % UInt16) or the dividend is UInt256 (1.6 to 2 times). #123170 (Groene AI).
• exp10 is up to 2x faster by splitting its argument with trunc instead of the modf library call, with bitwise-identical results. #122263 (Konstantin Bogdanov).
• Speed up entropy and multi-argument uniqExact on short strings by optimizing SipHash for short inputs. Hash values are unchanged. #123658 (Groene AI).
• sumMap, minMap, maxMap, sumMapWithOverflow, and sumMapFiltered over arrays are up to 2.3 times faster for integer, Enum, Date, Date32, and DateTime keys. #124340 (Groene AI).
• h3PolygonToCells and h3PolygonToCellsWithContainment are several times faster on polygons much smaller than a cell (9x on the query from the issue). This brings h3PolygonToCells back to the speed it had on such polygons before 26.8.12 and 26.9. Closes #123877. #123911 (Groene AI).
• isValidUTF8 validates a String or FixedString column with one simdutf call over the whole column instead of one call per row, which makes it up to 2.8 times faster on short strings. #123835 (Groene AI).
• Speed up lowerUTF8 and upperUTF8 on text whose non-ASCII characters are mostly two-byte UTF-8 (U+0080 to U+07FF, e.g. Latin-1, Greek, Cyrillic): the column is processed in one pass, these characters are mapped from a table built from ICU, and ICU is called only for the rest of a row that contains other characters. #123867 (Groene AI).
• lower, upper, and the ASCII path of lowerUTF8 and upperUTF8 are now faster by using AVX2 or AVX-512 depending on the CPU. #121374 (Manuel).
• Speed up lengthUTF8 and other UTF-8 functions on long strings by counting code points 64 bytes at a time, on both x86 and ARM. #122049 (Raúl Marín).
• Fixed ARM performance regressions in indexOf with a constant needle and in text output formats over non-ASCII strings. Closes #123338. Closes #123346. #123377 (Raúl Marín).
• UTF-8 string functions such as substringUTF8 are faster on long ASCII strings on CPUs with AVX-512. #120088 (Minh Vu).
• Improved performance of functions extractKeyValuePairs and extractKeyValuePairsWithEscaping. #119361 (Anton Popov).
• Function mortonEncode now resolves the type of each argument and the ratios of a constant range mask once per block instead of once per row, which makes it up to 2x faster. #123354 (Groene AI).
• Speed up mortonDecode on AArch64 by decoding with compile-time shift-and-mask sequences instead of a lookup table. The output is unchanged on every architecture. #116205 (Groene AI).
• encrypt and aes_encrypt_mysql are faster on values shorter than 16 bytes in the aes-*-ecb and aes-*-cbc modes (up to 2x for encrypt with aes-128-ecb). #124103 (Groene AI).
• Speed up hex and unhex with optimized AVX and AVX2 implementations when they are available. #123511 (Alexey Milovidov).
• Faster toMillisecond, toMicrosecond, and toNanosecond on DateTime64. #122852 (Samuel Krempaský).
• Speed up reverse for String and FixedString values. #119462 (Minh Vu).
• arrayDistinct is faster: it now does a single hash-table lookup per element instead of two, and Decimal arrays use a specialized fast path. #113966 (Manuel). #121646 (Minh Vu).
• Specialize has and indexOf searches in non-nullable numeric arrays when the needle is constant. #119090 (Minh Vu).
• arrayFilter is faster when the lambda keeps all or none of the elements of a block. #119816 (Minh Vu).
• Improve performance of bitmapAndCardinality, bitmapOrCardinality, bitmapXorCardinality, bitmapAndnotCardinality, and bitmapHasAny for UInt64 and Int64 bitmaps, which previously had to materialize the intersection to compute the result. On AArch64, bitmapAndCardinality, bitmapOrCardinality, bitmapXorCardinality, and bitmapAndnotCardinality no longer hit a rare code-layout-dependent slow path. #120615 (Ashish kumar choubey). #123675 (Groene AI).
• Restored the existing Decimal fast paths in arrayCompact. #121245 (Minh Vu).
• uniqUpTo now stops hashing after it exceeds its threshold. #121247 (Minh Vu).
• Restored the FixedString fast path for arrayEnumerateUniq and arrayEnumerateDense. #121338 (Minh Vu).
• arrayReverse is now faster for Decimal and wide integer arrays. #121342 (Minh Vu). #121644 (Minh Vu).
• Avoid copying String columns for substring(s, 1). #119820 (Minh Vu).
• Skip KMP preprocessing for impossible hasSubstr searches. #119824 (Minh Vu).
• Speed up arrayMin and arrayMax for Decimal and DateTime64 arrays. #123507 (Minh Vu).
• Speed up arrayUniq for Decimal32, Decimal64, and Decimal128 arrays. #122121 (Minh Vu).
• Allow a single condition to constrain multiple key columns during index analysis. For example, if the primary key is (toYYYYMMDD(dt), toDate(dt), dt) and the query has WHERE dt >= toDateTime('2026-01-10 00:00:00'), all three key columns are now used for pruning, instead of only one of them as before. This behavior is controlled by the setting analyze_index_with_multiple_key_columns_per_condition (enabled by default). #96225 (Nihal Z. Miaji).
• Left-side columns absent from a hash join’s output (for example the join key in SELECT count() ... LEFT JOIN ...) are now pruned when only a spill threshold is set, as with the default max_bytes_ratio_before_external_join, instead of being carried through the pipeline. #108191 (Groene AI).
• Joins on null-safe comparison (<=>) of nullable keys use fixed-width hash keys instead of a serialized tuple. #120596 (Manuel).
• Fixed a quadratic slowdown of JOIN with join_algorithm = 'hash' when ON has a non-equality condition or several OR conditions. #121209 (nt).
• Faster parsing of floating-point numbers from text formats. Update fast_float to v8.3.2. #121383 (Actuele AI).
• Query plan optimizations can now use the sort order reported by any table engine, not only MergeTree, avoiding unnecessary sorting. #121858 (Ivan Zhukov).
• Slightly reduce the per-block overhead of aggregation by not copying the columns of the block when preparing the aggregate function instructions. #124819 (Alexey Milovidov).
• GROUP BY by several keys, which requires the serialization of keys, serializes the keys of a block in chunks that fit in the CPU cache instead of all at once. This makes such queries faster when many threads are used, for example, ClickBench Q18 is 37% faster on a 96-core machine. #124955 (Alexey Milovidov).
• Added the setting cluster_function_wait_for_replica_ms (default 0, disabled), which lets cluster table functions wait for the replica assigned a file to connect before giving that file to another replica. This improves file distribution when replicas start at different times. Work stealing now takes files from the active replica with the most remaining bytes without scanning all unprocessed files. #125611 (Konstantin Vedernikov).
• The SLRU filesystem cache policy now moves a partially downloaded file segment to the protected queue on a cache hit, as it does for a fully downloaded one. Before, a scan could evict the data of narrow reads from the probationary queue. #125657 (Sema Checherinda).

​

Improvement

• Lightweight updates (the UPDATE statement and patch parts) are now generally available instead of Beta. The enable_lightweight_update setting is now in the production feature tier. #125719 (Alexey Milovidov).
• Cluster discovery is no longer experimental: clusters with a discovery section in remote_servers work without enabling allow_experimental_cluster_discovery. The setting is now obsolete. #123917 (Alexey Milovidov).
• The Kusto (KQL) and Trino query dialects were promoted from experimental to beta. They are still enabled with the allow_experimental_kusto_dialect and enable_trino_dialect settings. #124750 (Alexey Milovidov).
• Tables and databases can now be addressed through URL paths in the HTTP interface by default: curl http://localhost:8123/my_db/my_table.csv now works without extra configuration. The settings http_allow_path_requests, http_allow_database_as_path, http_allow_table_as_file, and http_allow_filters_as_path are enabled by default; http_allow_filters_as_unrecognized_url_parameters stays disabled. #121143 (Alexey Milovidov).
• The all_... union tables over the system log tables (system.all_query_log, system.all_text_log, and so on) are now created by default: the default server configuration enables the create_union_system_log_tables section with merge_rotated_tables, so queries over the rotated versions of a log no longer have to remember constructs like merge(system, '^query_log'). The cluster setting is shown as a commented-out example, because the table definitions would reference the cluster and require it to exist. #117943 (Alexey Milovidov).
• In a DataLakeCatalog database with allow_experimental_iceberg_compaction = 1, only Iceberg tables are now cached. DeltaLake and Paimon tables no longer keep the first schema they saw, so SELECT * and DESCRIBE pick up added columns without recreating the database or restarting the server. Closes #122548. #122568 (Groene AI).
• The background memory worker now corrects the global memory tracker from real memory usage on every tick (server setting memory_worker_correct_memory_tracker, enabled by default). Previously the tracker was a pure counter that was re-baselined only at startup and when it went negative, so upward accounting drift was permanent: servers were observed refusing every query with MEMORY_LIMIT_EXCEEDED, reporting more than 100 GiB in use while actually using 7 GiB, until restart. The drift itself remains visible as the difference between the MemoryTrackingUncorrected and MemoryTracking metrics. #118094 (Alexey Milovidov).
• The AI function per-query quotas are now disabled by default: ai_function_max_input_tokens_per_query, ai_function_max_output_tokens_per_query, and ai_function_max_api_calls_per_query default to 0, meaning no limit. To keep the previous caps, set the three settings explicitly or use compatibility = '26.9'. #121424 (George Larionov).
• Named Tuple columns in Pretty formats are now displayed as subcolumns, with the element names shown in additional header and footer lines. Controlled by the new setting output_format_pretty_named_tuples_as_subcolumns (enabled by default). A named tuple rendered inside one cell, for example nested in an Array or when the setting is disabled, is still displayed according to output_format_pretty_named_tuples_as_json. #78927 (Engel Danila).
• Show the query result in Pretty formats as soon as it arrives. Previously a query that returns a few rows early and then keeps reading, such as SELECT DISTINCT over a large table, displayed nothing until it finished or was cancelled. A later fix prevents a background write error after the query has finished from surfacing as a late LOGICAL_ERROR. #117005 (Alexey Milovidov). #125028 (Groene AI).
• Display non-printable control characters as Unicode “Control Pictures” in the Vertical and Pretty* output formats by default, so they are no longer silently swallowed and no longer deform the table. TAB, the line feed, and ESC are excluded, because a terminal interprets them rather than swallowing them. The new setting output_format_pretty_display_control_characters can be disabled to restore raw-byte output. #101774 (Yash ).
• Expired authentication methods are now removed from a user automatically. Every ALTER USER drops the authentication methods whose VALID UNTIL deadline has already passed and keeps all the others, so short-lived credentials can be rotated indefinitely with ALTER USER ... ADD IDENTIFIED WITH ... VALID UNTIL ... without the expired ones accumulating against max_authentication_methods_per_user. A method added by the same statement is not affected, and a user whose every method is expired keeps them rather than being left with no way to authenticate. #118880 (Alexey Milovidov).
• On ReplicatedMergeTree, the sync value 3 (mutations_sync, lightweight_deletes_sync, alter_sync) now waits for the active replicas, as documented, instead of all of them, so an inactive replica no longer causes UNFINISHED. This also affects OPTIMIZE, TRUNCATE, DROP/DETACH PARTITION, REPLACE PARTITION, and MOVE PARTITION TO TABLE. #120592 (Nikolay Degterinsky).
• A number written directly as the first argument of CAST is now read with the target type when that type reads it more precisely than the number itself carries it, such as a Decimal or an integer wider than 64 bits. Nullable, Array, and LowCardinality of those types take the same exact path. #116043 (Alexey Milovidov).
• TTL ... GROUP BY (a, b, c) is now accepted and means the same as TTL ... GROUP BY a, b, c, matching how ORDER BY (a, b, c) works. Previously the parenthesized form was rejected with TTL Expression GROUP BY key should be a prefix of primary key. #118426 (Alexey Milovidov).
• sequenceMatch, sequenceCount, and sequenceMatchEvents now accept a DateTime64 timestamp instead of failing. Events are ordered at the column’s full precision, so two events within the same second are no longer ambiguous. #118427 (Alexey Milovidov).
• argMin and argMax now consistently return the value from the first row among rows with the same minimum or maximum within a block, matching argMinIf and argMaxIf. #122892 (Raúl Marín).
• Added the setting analyzer_compatibility_allow_cte_redefinition. When enabled, a CTE name may be defined more than once in a single WITH clause; a later definition shadows the earlier one as the query analysis before 24.3 did. MATERIALIZED CTEs and CTEs in WITH RECURSIVE cannot be redefined. #121623 (Dmitry Novik).
• Added settings semi_join_include_columns_from_both_sides and anti_join_include_columns_from_both_sides (enabled by default) for SEMI and ANTI joins in the analyzer. When enabled, behavior is unchanged: columns from both sides are accessible after the join. Disable either setting to enforce SQL-standard semantics, where only the preserved side’s columns are visible after the join. #96662 (Zhang Yifan).
• The setting aggregate_function_input_format with the values value and array now works in every input format, including RowBinary, Native, and Parquet, and for AggregateFunction nested in Array, Tuple, and Map. In the JSON formats the value is now a JSON value, for example {"x": [1, 2, 3]}, not a string containing it. #121120 (Alexey Milovidov).
• Compression codec names and their keyword arguments are now matched case-insensitively, so CODEC(zstd) works like CODEC(ZSTD) and ALP(std) like ALP(STD). Previously a codec whose name is mixed-case, such as Delta, DoubleDelta, Gorilla or Quantized, could not be used in the default_compression_codec, marks_compression_codec and primary_key_compression_codec settings, nor in the <compression> section of the server configuration, because those paths upper-cased the codec name. #119313 (Alexey Milovidov).
• Added aliases allow_suspicious_indexes, ignore_data_skipping_indexes, force_data_skipping_indexes, and secondary_indexes_enable_bulk_filtering for the settings spelled with indices, so both plural forms are accepted. #122565 (Mohammed Sohail).
• APPLY (f, 'prefix') now keeps the table qualifier in generated column names. #122702 (Groene AI).
• Async insert now supports eligible single-block local INSERT ... SELECT queries, with synchronous fallback for larger, multi-block, empty, transactional, or non-parallel-quorum inserts. #111663 (Mikhail f. Shiryaev).
• Added the opt-in optimize_row_order_if_no_order_by setting for MergeTree tables with an empty sorting key (ORDER BY () / ORDER BY tuple()). When enabled, it applies row-order optimization automatically to inserts into such tables. #104591 (Alexey Milovidov).
• WHATIF can now estimate hypothetical projections without reading the whole table: past max_rows_to_scan rows it estimates from a sample of granules and reports the resulting range in marks_span, max_rows_to_read lowers that budget, and it follows (prefer|force)_optimize_projection. It now also uses the real projection-optimizer weighing logic, so its verdicts, reasons, and pruned ranges track the optimizer more closely. #118233 (Yarik Briukhovetskyi). #121766 (Yarik Briukhovetskyi). #122060 (Yarik Briukhovetskyi). #123540 (Yarik Briukhovetskyi). #123587 (Yarik Briukhovetskyi).
• Added the setting parallel_replicas_for_queries_with_multiple_tables (default true) to control whether parallel replicas are used for queries joining multiple tables. When disabled, parallel replicas are not used for queries with JOIN, where the non-leftmost side is read in full on every replica. The setting does not affect a UNION query without a JOIN, where each branch is an independent single-table read, nor ARRAY JOIN. #76595 (Igor Nikonov).
• max_joined_block_size_rows and max_joined_block_size_bytes now also bound the result blocks of parallel_hash and grace_hash joins: squashing of the join output no longer merges blocks past these limits. Closes #124198. #124304 (Groene AI).
• Added the setting query_plan_filter_push_down_below_limit_by (default 1), controlling whether filters on LIMIT BY key columns are pushed below the LIMIT BY step. Setting it to 0 restores the plan shape used before 26.7, for queries where the pushed predicate makes index analysis costlier than the scan it saves. #121675 (Groene AI).
• Added the use_aggregation_memory_tracker setting to choose how GROUP BY measures memory when deciding on two-level aggregation. compatibility below 26.9 restores the measurement used before 26.7. #124439 (Hechem Selmi).
• system.query_log and system.processes now report the metrics of each task of a distributed query plan (make_distributed_plan = 1) in a row of its own: a task is a secondary query with its own query_id, correlated to the statement that started it by initial_query_id. The initiator’s row reports its own work instead of the volume its tasks read. #120949 (Groene AI).
• system.query_log now records the databases, tables, and columns read by a replica that executes a query plan shipped to it, which happens with parallel_replicas_plan_based. #122657 (Igor Nikonov).
• Queries with send_logs_level now also forward text logs from stateless make_distributed_plan workers through the coordinator, so worker-side exception context reaches the client. #118603 (Antonio Filipovic).
• The setting parallel_replicas_insert_select_local_pipeline is obsolete. It was a workaround for a performance problem fixed in 25.5, and it duplicated parallel_replicas_local_plan, which already decides whether the initiator reads its own share of a distributed INSERT SELECT. Use parallel_replicas_local_plan = 0 to leave all the reading to the remote replicas. #121653 (Igor Nikonov).
• Distributed SELECT queries now emit one OpenTelemetry span per remote shard read, with attributes such as clickhouse.cluster, clickhouse.shard_num, clickhouse.processed_stage, clickhouse.query_id, clickhouse.initial_query_id, and clickhouse.target_host, so slow or failed shards can be identified directly from system.opentelemetry_span_log. #115784 (Diego Gomes Tomé).
• Reads with parallel replicas are now covered by OpenTelemetry spans: a summary span with the totals of the coordination, one span per announcement and per read request handled by the coordinator, and clickhouse.replica_num on the per-replica spans, so the work distribution between replicas can be inspected in system.opentelemetry_span_log. #118570 (Diego Gomes Tomé).
• Added profile events for network activity: DNSRequests, DNSRequestMicroseconds, and DNSRequestError; DNSReverseRequests, DNSReverseRequestMicroseconds, and DNSReverseError; TLSHandshakes, TLSHandshakeMicroseconds, and TLSHandshakeErrors; TLSServerHandshakes, TLSServerHandshakeMicroseconds, and TLSServerHandshakeErrors; and HTTPServerConnectionsErrors. DNS now exposes timing as well as error counts, and TLS handshakes are now observable. #116136 (Alexey Milovidov).
• Added the AsyncInsertFlush profile event. Compare it with AsyncInsertQuery to see how many queries are batched into one asynchronous INSERT flush. #120187 (Sema Checherinda).
• Add the RuntimeFilterLookupsBeforeBuildFinished profile event, counting blocks that a JOIN runtime filter passed through unfiltered because its build side had not finished yet (for example with join_overflow_mode = 'break' and a finite max_rows_in_join / max_bytes_in_join). #107108 (Groene AI).
• Added last_error_symbols and last_error_lines to system.errors and system.error_log. Closes #74561. #100374 (Narasimha Pakeer).
• system.processors_profile_log now has entries for queries that fail with an exception, and new columns exception_code and exception. #121503 (Aleksandr Efimov).
• DROP TABLE and DROP DATABASE now write a RemovePart event to system.part_log for every part of the dropped table, as merges, DROP PARTITION, and TRUNCATE already do. #122093 (Garrett Thomas).
• Added the server setting min_allocation_size_to_log_stack_trace. When a single charge to the global memory tracker reaches it, the stack trace is written to the server log and to system.trace_log with trace type MemoryLargeAllocation. This helps attribute a global tracked total that has grown far beyond real memory usage, a state in which the server refuses every allocation while far below its limit. Default 0 keeps it off. #119527 (Groene AI).
• Show throttler and workload IO scheduler wait time in the client progress bar (ThrottlerSleepMicroseconds, SchedulerIOReadWaitMicroseconds and SchedulerIOWriteWaitMicroseconds are summed into the waited figure, and the stalled sections of the bar are colored yellow). #118536 (Azat Khuzhin).
• EXPLAIN ANALYZE time=1 shows more information about time and parallelism of every step. #114786 (Kirill Kopnev).
• Peak-memory reporting under EXPLAIN ANALYZE now accounts for aggregation correctly. #123324 (Azat Khuzhin).
• Removed nine unneeded ProfileEvents: the eight never-incremented DistributedIndexAnalysisParts, ParallelReplicasDeniedRequests, WasmMemoryAllocated, S3QueueSetFileProcessingMicroseconds, S3QueueSetFileProcessedMicroseconds, S3QueueSetFileFailedMicroseconds, FileSegmentRemoveMicroseconds, and ExternalDataSourceLocalCacheReadBytes, plus the duplicate CachedReadBufferPredownloadedBytes, which always matched CachedReadBufferPredownloadedFromSourceBytes. This reduces per-query memory and bookkeeping overhead; queries and dashboards should use CachedReadBufferPredownloadedFromSourceBytes. #123040 (Raúl Marín). #123042 (Raúl Marín).
• Killing a query that is loading skip-index marks from object storage now stops the read promptly: the remote read path checks query cancellation before each buffer fill, instead of reading on for minutes after the query was already cancelled. #113436 (Groene AI).
• Reduced cancellation latency for GROUP BY ... WITH TOTALS queries: KILL QUERY now interrupts totals accumulation within a single chunk instead of waiting for all aggregate states to be merged. #119470 (Roman Vasin).
• KILL QUERY and Ctrl+C in clickhouse-client now cancel long-running functions inside WHERE and HAVING filters instead of waiting for the function to finish. #112705 (Alexey Milovidov).
• KILL MUTATION, SYSTEM STOP MERGES, DETACH/DROP of a table, REPLACE/ATTACH PARTITION ... FROM and server shutdown now stop a running mutation within a fraction of a second when the mutation’s predicate has an IN (subquery) whose subquery works without reading new data (for example a cross join). Previously they waited until the whole set was built, and REPLACE/ATTACH PARTITION ... FROM gave up with TIMEOUT_EXCEEDED after 120 seconds. KILL MUTATION now also stops a mutation that is already writing its new part. #124122 (Groene AI).
• KILL QUERY and max_execution_time (with the default timeout_overflow_mode = 'throw') now interrupt writing the Parquet format in the middle of a block that is split into many row groups (for example with a small output_format_parquet_row_group_size), instead of only after the whole block is written. #125489 (Groene AI).
• DROP TABLE, DETACH TABLE, server shutdown, and ALTER TABLE ... MODIFY SETTING refresh_statistics_interval no longer wait for a background refresh of the statistics cache of a MergeTree table to read every part. On object storage with thousands of parts this wait could take minutes per table and delay a restart. Closes #125136. #125165 (Groene AI).
• DROP TABLE and other operations that cancel merges and mutations no longer wait for an in-progress vector_similarity index build to finish. #123062 (Raúl Marín). #123739 (Raúl Marín).
• KILL QUERY WHERE query_id = '<id>' now cancels the caller’s own query without requiring SELECT on system.processes. For a caller holding neither that grant nor KILL QUERY, an id that is not running as them cancels nothing and returns no rows. Any other WHERE condition keeps requiring the privileges it requires today. #117999 (Groene AI).
• Access to system.jemalloc_profile_text, system.jemalloc_sampled_allocations, system.jemalloc_bins, system.jemalloc_arena_bins, and system.jemalloc_stats now requires SYSTEM_JEMALLOC; the heap profiles flushed by system.jemalloc_profile_text and system.jemalloc_sampled_allocations are also removed after being returned. #121622 (Azat Khuzhin). #121634 (Azat Khuzhin).
• Added the access_control_improvements.readonly_restricts_set_role server setting, false by default. When enabled, SET ROLE is refused with READONLY in a session or query whose readonly setting is not zero. Previously SET ROLE was the only identity-changing statement that readonly did not refuse. #118021 (Groene AI).
• Secret arguments of table functions, table engines, database engines, backup engines, and dictionary sources are now declared next to each feature and are masked consistently in SHOW CREATE, system tables, query logs, and EXPLAIN AST, instead of relying on a central hardcoded parser list. #123364 (Raúl Marín).
• The PostgreSQL, MaterializedPostgreSQL, DataLakeCatalog, Cluster, Remote, and HDFS database engines key the exemption for the server’s replay of its own stored metadata on the metadata loader instead of on the query being internal. A HDFS database restored with RESTORE DATABASE is now checked against remote_url_allow_hosts. #120280 (Alexey Milovidov).
• Allow remote_url_allow_hosts to match the bucket name of an S3 URL. #124453 (Konstantin Vedernikov).
• Support proxy credentials in the http_proxy and https_proxy environment variables. Previously the userinfo part of the proxy URI was discarded, so ClickHouse could not use an authenticating proxy. Proxy credentials are sent only to the proxy, rather than to the target server inside a CONNECT tunnel. #123069 (Akanshaak). #124396 (Alexey Milovidov).
• Interserver authentication failures on a listener that uses the PROXY protocol now record the client address from the PROXY header in system.session_log when auth_use_forwarded_address is enabled, instead of the proxy address. #125326 (Groene AI).
• Added the tcp_with_proxy_allow_interserver_mode server setting, which allows disabling the native protocol’s interserver mode on tcp_with_proxy_port. #123188 (Pradeep Chhetri).
• Reject unexpected packets in the native protocol before deserializing their payload, avoiding unnecessary processing of data types and aggregate function states before returning UNEXPECTED_PACKET_FROM_CLIENT. #119368 (Alexey Milovidov).
• Added the setting kill_throw_if_noop, enabled by default, which makes local KILL QUERY and KILL MUTATION statements throw NO_QUERY_TO_KILL when their selection produces no eligible target. ON CLUSTER execution remains tolerant of empty matches. Set kill_throw_if_noop = false to preserve the previous behavior. Related #83348. #85023 (zwy991114).
• Merges and queries on S3 disks no longer hold one HTTP connection per open stream for their whole duration once connections are scarce. Previously this let a few merges of tables with many substreams, for example a JSON column, saturate disk_connections_hard_limit, stall each other, and fail every other request to the disks with HTTP_CONNECTION_LIMIT_REACHED. Once the Disk connection group is at or above disk_connections_soft_limit, reads from S3 disks fetch each buffer fill with a separate request and return the connection to the pool between fills; the new profile event ReadBufferFromS3RequestsCut counts such requests. #119730 (Alexey Milovidov). #124701 (Groene AI).
• The filesystem_cache_boundary_alignment setting can now be changed per query. For disk reads, 0 means to use the cache configuration’s boundary_alignment; for non-disk reads, 0 still disables alignment. #118791 (Kseniia Sumarokova).
• Added the filesystem cache setting use_real_disk_size (disabled by default), which accounts for cache reservations in filesystem allocation units instead of raw bytes. It makes FilesystemCacheSize, eviction-size metrics, and current_size in system.filesystem_cache_settings approximate the disk footprint. The accounting is block-aligned rather than exact and can include reserve-ahead bytes of in-flight reads. #79490 (codeworse).
• Custom HTTP headers configured for S3 are now handled case-insensitively by name, so a header spelled X-Amz-Meta-Foo reaches Google Cloud Storage translated to x-goog-meta-foo, as x-amz-meta-foo already did. #120014 (Sema Checherinda).
• Better error messages for S3 request retries: the HTTP response code is no longer printed as 18446744073709551615 when the request received no response at all, and the misleading message about a possible time skew is no longer printed on every retry. #118439 (Alexey Milovidov).
• Allow DESCRIBE FILESYSTEM CACHE on the introspection port. #115894 (Groene AI).
• Files on encrypted disks now support positional reads, so a table with a UNIQUE KEY on an encrypted disk can be loaded again after a server restart. #119991 (johnjing).
• The url table function and the URL table engine now expand glob patterns lazily instead of materializing every address before the query starts. glob_expansion_max_elements now limits how many addresses a query actually generates rather than how large the pattern is, so a query that stops early - for example under a LIMIT - can use a pattern that describes far more addresses than the limit. #115680 (Alexey Milovidov).
• Iceberg manifest writing now uses sequence-number inheritance: ADDED entries have NULL sequence_number and file_sequence_number unless compaction assigns an explicit sequence number. #120031 (Asya Shneerson).
• Iceberg manifest statistics are now written for boolean and fixed columns. #121548 (Konstantin Vedernikov).
• When a table is created through a catalog, CREATE TABLE catalog.table ENGINE = IcebergS3 can now omit explicit credentials and reuse the catalog credentials. #114477 (Konstantin Vedernikov).
• ALTER TABLE ... EXECUTE expire_snapshots now works for Iceberg tables in transactional catalogs such as Iceberg REST and Unity. #123248 (alesapin).
• A DataLakeCatalog database whose catalog cannot be built no longer reports its tolerated table-listing failure as an error of the query that enumerated it: the skipped database is now logged at debug instead of error, so a query such as SELECT ... FROM system.tables or system.iceberg_history no longer sends a server-side exception to a client whose send_logs_level includes error. Set show_data_lake_catalogs_in_system_tables = 1 to have the real exception raised instead. #121475 (Groene AI).
• When a DeltaLake write commit fails, ClickHouse now always reports the commit error rather than an error raised while removing the data files the failed INSERT had already uploaded, and every removal that fails during that cleanup is logged with the path it was left on. Previously the first failing removal replaced the commit error in the response, so the reason nothing was written was lost and no path was logged. Closes #122023. #122123 (Groene AI).
• Reading an Iceberg table whose position delete file has no pos or file_path column now fails with ICEBERG_SPECIFICATION_VIOLATION naming the file and the missing column, instead of a LOGICAL_ERROR (missing pos) or a misleading UNKNOWN_IDENTIFIER (missing file_path). #123583 (Groene AI).
• Paimon BOOLEAN columns are now read as Bool instead of Int8, matching Iceberg and DeltaLake. DESCRIBE and toTypeName now report Bool, and text formats render the values as false and true instead of 0 and 1. #119404 (JIaQi Tang).
• Avro files now include per-column size statistics in their metadata. #121392 (Konstantin Vedernikov).
• The Parquet reader now rejects a column chunk whose metadata has a negative num_values, total_uncompressed_size, or unencoded_byte_array_data_bytes with INCORRECT_DATA instead of failing later with a misleading memory-allocation error. It also caps memory estimates derived from these fields at the reader’s watermarks and ignores pre-allocations too large to ever succeed, so absurdly large values no longer cause an internal error. Closes #124224. #124330 (Groene AI).
• Bumped the bundled ORC library so that an ORC file declaring a compression codec the reader does not implement is rejected with an error naming the codec, instead of an internal error that debug and sanitizer builds treat as a failed assertion. #115889 (Groene AI).
• INSERT ... VALUES over a stream that fails mid-row now reports that read failure directly, instead of retrying the failed read while recovering from a parse error. This affects an HTTP request body that is shorter than its declared Content-Length and is streamed to the input format. #121916 (Groene AI).
• ClickHouse Keeper no longer drops the client session when the outgoing request queue stays full for longer than operation_timeout_ms. The enqueue wait is now three times session_timeout_ms, so a short burst of backpressure no longer invalidates the session’s ephemeral nodes and watches. #121559 (Kseniia Sumarokova).
• ClickHouse Keeper now reports orphaned nodes in a snapshot as CORRUPTED_DATA instead of an opaque internal error. With keeper_server.remove_orphaned_nodes_on_startup = true and keeper_server.digest_enabled = false, it can also remove those nodes during startup so the snapshot still loads. #95761 (Seva Potapov).
• The Keeper rcfg command now honors its retry count per action, stops once max_total_wait_time_ms is exhausted, and rejects a negative retry before running any action. #122758 (Groene AI).
• Updated NuRaft to fix handling of abandoned peers when a removed server rejoins. #120736 (Antonio Andelic).
• Added the setting get_zookeeper_lock_acquire_timeout_ms so Keeper-backed queries such as reads from system.zookeeper, zookeeperSessionUptime, and SYSTEM RECONNECT ZOOKEEPER can fail fast while another thread is reconnecting Keeper. #94836 (c-end).
• Activating a replica whose active or is_active node in Keeper is still held by another session is now reported as REPLICA_ALREADY_EXISTS instead of LOGICAL_ERROR, so it no longer aborts the server in debug and sanitizer builds. The same applies when that node is rewritten while it is being removed. #114408 (Groene AI).
• The ZooKeeper client, the Graphite metrics writer, the Redis source, system.zookeeper_info, the distributed query exchange, and host names in listen_host and bind_host now resolve through the ClickHouse DNS cache, so these lookups are cached, visible in system.dns_cache and subject to dns_allow_resolve_names_to_ipv4 and dns_allow_resolve_names_to_ipv6. Previously they resolved names on their own, bypassing the cache. #120831 (Alexey Milovidov).
• Added the http_x_clickhouse_format_overrides_output_format setting (default true). When disabled, the X-ClickHouse-Format HTTP header only sets default_format, as it did before 26.8, instead of overriding a query’s FORMAT clause via output_format. #121132 (Alexey Milovidov).
• The Prometheus HTTP API accepts RFC3339 times with Z or a +hh:mm offset and fractional seconds in time, start, and end, as Prometheus does, so clients like vmalert work. String time arguments of prometheusQuery, timeSeriesSelector, and the timeSeries*ToGrid functions accept RFC3339 too. #122808 (Valery Petrov).
• PromQL regex matchers that list plain literals, such as {job=~"api|web"}, now use the primary key and return exactly the listed values, as in Prometheus. Before, only the first alternative was anchored at the start and only the last at the end, so such a matcher also returned values like api-server or myweb. #122796 (Valery Petrov).
• PromQL instant queries now return series in a deterministic order: topk and bottomk at the root are sorted by value, as Prometheus returns them, and other instant queries are sorted by their labels, like range queries, so repeated requests list the series in the same order. #122347 (James Cunningham). #122976 (Valery Petrov).
• /api/v1/format_query and PromQL serialization now print durations in Prometheus units, so sum(rate(x[5m])) formats as itself rather than as sum(rate(x[300])). #120939 (Valery Petrov).
• PromQL now validates label names in label_join and label_replace. #118190 (Minh Vu).
• A Prometheus remote-write or remote-read request whose body cannot be decompressed is now answered with 400 instead of 500, and the error names the size of the block and the uncompressed length it declares. #120938 (Valery Petrov).
• Added the merge_selector_min_age_to_disable_right_tail_heuristic MergeTree setting. It lets the Simple and StochasticSimple merge selectors include small parts at the right end of a range in merge candidates once every part in the range is old enough. #122638 (Alexey Milovidov).
• Added the MergeTree settings merge_selector_small_parts_min_count, merge_selector_small_parts_threshold, and merge_selector_small_parts_max_age to reduce the number of merge operations under rapid small-part insertion by requiring a minimum batch size before merging small fresh parts. #100436 (Amos Bird).
• Calculate projections using the query analyzer and reject missing projection columns rather than writing incomplete projection parts. Row policies and additional_table_filters for system.one no longer affect projection calculation. #118375 (Alexey Milovidov).
• ALTER TABLE ... ADD COLUMN ..., MODIFY ORDER BY ... now rejects a new sorting key that uses a subcolumn of the added DEFAULT or MATERIALIZED column, and its error message names whether the expression kind is DEFAULT or MATERIALIZED instead of always saying default expression. #121873 (GON).
• An unresolved name in a column DEFAULT, MATERIALIZED, or ALIAS expression is now reported by the name the user wrote, instead of an internal <column>_tmp_alter<random> alias that appears nowhere in the query and also broke the “Maybe you meant” hint. A sorting, primary, or partition key that contains a nullable column or a constant now names that column or constant. #116871 (Alexey Milovidov).
• Web UI: save and load the list of tabs, their queries, and query parameters as a Markdown file. The tab bar is now always shown, and the connection settings always live behind the key button in its corner. #121234 (Alexey Milovidov).
• Web UI: added a Format button to the query editor, which formats the query with formatQuery. #121652 (Alexey Milovidov). #122477 (Alexey Milovidov).
• Web UI: clicking a cell of a JSON column pretty-prints the value with two-space indentation and syntax highlighting. #118640 (Alexey Milovidov).
• The graph view of the Web UI for EXPLAIN ... graph = 1 now follows the light/dark theme instead of always drawing a white canvas with black lines, and its labels use a sans-serif font instead of the Graphviz default. Rendering a graph also works when play.html is opened as a local file. #116208 (Alexey Milovidov).
• In the Web UI, the digits of a long number in the result table are now grouped by a thin gap, as in 1 234 567, instead of an underline. The gap is an empty element, so selecting or copying the number still yields the digits alone. #119839 (Alexey Milovidov).
• Show a checkmark on the Play UI result-table copy button for 1.2 s after a cell value is copied to the clipboard. #120617 (yanglongwei).
• In the Web UI, the browser window title of a renamed query tab is now exactly the tab’s name. #122067 (Alexey Milovidov).
• Web UI: when a login last succeeded with an empty password, a stale password autofilled by the browser’s password manager for that server and user is no longer used. #118638 (Alexey Milovidov).
• The schema visualizer now supports hiding and showing table categories by clicking their legend labels, with a Hide all / Show all control. #124963 (Mark Needham).
• The command-line client no longer shows as-you-type hints on a query recalled from the history or otherwise displayed at once, so pressing Up/Down keeps navigating the history until the line is edited. #115065 (Alexey Milovidov).
• The progress bar in clickhouse-client and clickhouse-local no longer flickers when the query result is redirected to a pipe or a file. #123549 (Alexey Milovidov).
• A bare table name can now use the archive path syntax in clickhouse-local, so SELECT * FROM 'archive.tar.zst::data.native' reads a file inside an archive, like SELECT * FROM file('archive.tar.zst::data.native'). The same applies to tables of a URL database with a file:// target and of a Filesystem database. #118170 (Alexey Milovidov).
• Clean up client output when a web terminal or SSH session is closed during a running query, avoiding spurious write-buffer errors. #125558 (Groene AI).
• A musl-built ClickHouse binary is now self-contained enough to run in a FROM scratch container image with no extra files: CA certificates are embedded, localhost resolves to loopback even without /etc/hosts, and DNS no longer falls back to the unreachable localhost when /etc/resolv.conf is missing. #115389 (Alexey Milovidov).
• Hedged requests are now supported on macOS. Previously use_hedged_requests was silently ignored on macOS and distributed queries always used non-hedged connections. #119915 (Raúl Marín).
• Determine the number of CPUs from the kernel’s sysfs lists instead of the libc, so per-CPU structures and auto thread counts do not depend on the libc’s interpretation of the affinity mask. #118912 (Konstantin Bogdanov).
• Workloads created without a parent workload are now allowed; they become children of an implicit root workload, so several workload trees can coexist. #118578 (Sergei Trifonov).
• RANDOMIZE FOR is now applied on each replica of a refreshable materialized view independently, so a random replica performs each refresh instead of the same one every time. Replicas no longer agree on next_refresh_time, so the value reported by system.view_refreshes on one replica is the time that replica would refresh at, not necessarily when the refresh happens. Closes #78410. #112979 (Groene AI).
• FileLog now reads a file again from the beginning if, at the next read, it is shorter than the offset already read from it (for example after logrotate with copytruncate), instead of failing every read of the table with CANNOT_READ_ALL_DATA until the file grows back to its old size. #123390 (Groene AI).
• A relative path in the FileLog table engine is now resolved against user_files_path, like in the file table function, so FileLog('nginx/', 'JSONEachRow') reads <user_files_path>/nginx/. A relative path that already points inside user_files_path from the server’s working directory keeps its meaning. #123397 (Groene AI).
• Connections from the MongoDB table engine, table function, and dictionary source now identify themselves to the MongoDB server with the ClickHouse name and version in the driver handshake metadata. MongoDB operators can tell ClickHouse connections apart from other clients in server logs, currentOp output, and monitoring tools. #122205 (bisht2050).
• Lookups of YTsaurus dictionaries can now be throttled. #100044 (MikhailBurdukov).
• AI functions using the openai, huggingface, or tei provider send the query ID and the initial query ID in the X-ClickHouse-Query-Id and X-ClickHouse-Initial-Query-Id HTTP headers. #123622 (George Larionov).
• aiFilter now requests structured JSON Schema output. #121526 (Andrew Wilkins).
• WebAssembly UDF calls with ABI BUFFERED_V1 now measure the whole input batch instead of one serialization per row, so usage is priced by the stream the guest actually receives for formats whose cost is not additive across rows. #118989 (Vasily Chekalkin).
• Allow the ON CLUSTER clause for RENAME TABLE and EXCHANGE TABLES on a Replicated or Cloud Shared database when the cluster name matches the database’s automatic cluster name. The clause is treated as a no-op and the database engine propagates the DDL as usual, which lets scripts written for non-Replicated databases be ported without rewriting ON CLUSTER clauses. #106192 (Groene AI).
• Stored HTTP handler queries no longer break when a request supplies max_parser_depth, max_parser_backtracks, or max_query_size, and the CREATE HANDLER, ALTER HANDLER, and DROP HANDLER documentation is now published. #114883 (Alexey Milovidov).
• Improve memory tracking for query sources, including recursive CTEs and external data sources, so their container allocations count toward memory limits. #124952 (Nikita Taranov).
• A subquery with its own SETTINGS clause now plans GROUP BY using the settings after constraints and max_threads clamping are applied, instead of reapplying the clause as written. This prevents nested SETTINGS max_threads = ... from bypassing settings constraints or expanding the aggregation pipeline to an excessive number of streams. #123156 (Groene AI).
• Improved the error message of parseDateTime. #120489 (Robert Schulze).
• Restoring an invalid data part from backup now reports BACKUP_DAMAGED instead of CORRUPTED_DATA. #113085 (Alexander Tokmakov).
• Deserializing an aggregate function state that declares an oversized element count now reports TOO_LARGE_ARRAY_SIZE and names both the offending count and the limit, instead of a bare POCO_EXCEPTION carrying no numbers. #121381 (Groene AI).
• A setting given a value of the wrong type or out of range now names the setting and the value in the error message. #116882 (Alexey Milovidov).
• When a dictionary with a PostgreSQL source and a custom QUERY returns its columns in an order that does not match the dictionary structure, the error now names the destination dictionary column and the column order the query must return, instead of only reporting that a value could not be parsed. #114212 (Groene AI).
• A PostgreSQL dictionary whose QUERY returns fewer columns than the dictionary structure now fails to load with a NUMBER_OF_COLUMNS_DOESNT_MATCH error that names the column order the QUERY must use, instead of a LOGICAL_ERROR. A QUERY that returns too many columns gets the same column-order hint. #125397 (Groene AI).
• A column matcher (*, t.*, COLUMNS('...')) inside additional_table_filters, additional_result_filter or a row policy is now rejected with a clear BAD_ARGUMENTS diagnostic instead of the obscure Unqualified matcher * cannot be resolved. There are no table sources. CREATE / ALTER ROW POLICY with a matcher now fails immediately. A matcher inside a subquery of such a filter, for example x IN (SELECT * FROM allowed), keeps working. #119250 (Alexey Milovidov).
• A matcher (*, t.*, COLUMNS(...)) used as a whole LIMIT AFTER or LIMIT UNTIL boundary is now rejected with UNEXPECTED_EXPRESSION, as in WHERE, instead of failing with std::out_of_range or silently using only its first column. #123309 (Groene AI).
• CREATE FUNCTION now rejects a SQL user-defined function whose lambda arguments cannot be evaluated, instead of storing a function that later fails with BAD_ARGUMENTS on every call. Closes #121197. #121370 (Groene AI).
• A correlated subquery in HAVING of a WITH TOTALS query that also uses max_rows_to_group_by with group_by_overflow_mode = 'any' (and totals_mode other than after_having_exclusive) now fails with NOT_IMPLEMENTED instead of a logical error. #122601 (Alexey Milovidov).
• The mergeTreeAnalyzeIndexes and mergeTreeAnalyzeIndexesUUID table functions now reject a vector_search_index_analysis limit above max_limit_for_vector_search_queries, the bound under which the plan optimization rewrites a vector search query. A larger limit previously made the vector similarity index reserve memory proportional to it and exhaust the server’s memory. #122154 (Groene AI).
• An unconditioned CROSS JOIN is no longer reported as INNER in EXPLAIN and in the join columns of system.query_log when the dpsub join order algorithm runs with a conflict detector enabled. Query results were not affected. #122236 (Fisnik Kastrati).
• Formatting a query where an IN is the left operand of a tuple element ((x IN (1)).1) inside a multi-argument function call is now stable across a format-parse-format round trip. #124325 (Groene AI).
• Exchanging -Merge aggregate states with older servers now stays compatible: the state version of a -Merge function follows the state version of the function it merges only for peers that announce the new protocol revision 54494. #123070 (Groene AI).
• An aggregate function parameter whose value has no unambiguous bare literal form, such as a Decimal, a wide integer, a UUID, an IPv4 or IPv6, or a non-negative Int64, is now always printed with an explicit ::Type suffix, and inside an array or tuple parameter the suffix goes on the elements. Such a parameter therefore reads back as the same value for every family, instead of only for the ones that had opted in. The typed form also travels over the Native protocol, so an insert into a column whose parameter was spelled without the type is now rejected there too, as it already was locally. #121725 (Groene AI).
• Diagnostics about slowly acquired locks are now logged at Information level instead of Warning, so they no longer reach clients that use send_logs_level = 'warning' while still being written to the server log with a stack trace. #123797 (Alexey Milovidov).
• Adaptive aggregation now logs less noisily by reducing the severity of its messages. #123312 (Azat Khuzhin).
• kqlRangeCount now rejects the Int256 minimum span with step -1 instead of incorrectly reporting an empty range. #117212 (Alexey Milovidov).
• A distributed query whose shards plan aggregation with the experimental enable_cascades_optimizer no longer hits the logical error GroupingAggregatedTransform pushed bucket N twice on the initiator. #119814 (Alexey Milovidov).
• Local execution of distributed query plans (distributed_plan_execute_locally = 1) now materializes constant columns in its in-memory exchange like serialized exchanges do, fixing a query that could fail with a LOGICAL_ERROR when it sorted by a constant expression. #124097 (Groene AI).
• The internal function __bitSwapLastTwo now raises the same BAD_ARGUMENTS error for unsupported argument types when the expression is JIT-compiled as when it is interpreted. #117613 (Groene AI).
• Fix a UBSan failure during parallel parsing of TabSeparated input with an empty buffer. #122636 (Ray Winkelman).
• The ErrorCode registry is now sparse, so ClickHouse can reserve gaps between error codes without preallocating counters for them. #111796 (Maxim Orlovsky).
• Fixed a signed integer overflow when a negative or huge lock_acquire_timeout value is used as the timeout for scheduling a background task. #117898 (Alexey Milovidov).
• Cancelled marks-loading background tasks are no longer counted as errors in system.errors and system.error_log. Previously ASYNC_LOAD_CANCELED was recorded during normal operation whenever a reader was dropped before its asynchronously loaded marks were needed. #118928 (Alexey Milovidov).
• ALTER TABLE ... EXECUTE expire_snapshots for Iceberg tables now counts a file as deleted only when its removal succeeds, reports the removals that return an error in a new failed_deletions_count metric like remove_orphan_files, and counts a file that several expired manifests list only once. Closes #125332. #125379 (Groene AI).
• Reading an archive that libarchive fails to open without giving a reason (for example a .tar.gz with a truncated gzip trailer) now throws CANNOT_UNPACK_ARCHIVE instead of STD_EXCEPTION (fmt::format_error: string pointer is null). Closes #125620. #125650 (Groene AI).
• RESTORE from a tar backup with a non-empty password setting now fails with NOT_IMPLEMENTED, like BACKUP to a tar archive, instead of LOGICAL_ERROR. Closes #125648. #125666 (Groene AI).

​

Bug Fix (user-visible misbehavior in an official stable release)

• Fixed silent data loss when writing to an S3 key that already holds an object, for example INSERT INTO FUNCTION s3(...) with s3_truncate_on_insert = 1. Closes #114348. #118829 (Sema Checherinda).
• Fix possible data loss when a part’s checksums.txt is repaired (recalculated and written on load, or by CHECK TABLE) and the server loses power afterwards. The metadata rewrite is now durable and atomic (write to a temporary file, fsync, atomically rename), and an empty checksums.txt is now recalculated on load instead of detaching the whole part as broken. #111414 (Groene AI).
• Fix a part being detached as broken, losing its rows, when columns.txt or metadata_version.txt survives a power loss at length zero. Such a file is now treated like an absent one, which the load path already handles safely. #119582 (Alexey Milovidov).
• Fix data loss when loading a part with a broken projection: a marks file damaged by an interrupted write could be accepted as an empty part instead of being detached as broken. #119586 (Alexey Milovidov).
• Fixed replica recovery of a Replicated database deleting the data of a KeeperMap table, for example after the server was killed during an ALTER of that table. A KeeperMap table that still exists in Keeper is now kept or renamed with its data, instead of being dropped and re-created empty. Closes #111957. #122917 (Groene AI).
• DROP DATABASE of a Replicated database now drops a materialized view with ENGINE = TimeSeries instead of failing with It's not initial query and leaving the database half-dropped. #123663 (Groene AI).
• Fixed data loss when a single ALTER TABLE renamed a Nested member, or another column with a dot in its name, and modified the renamed column. Such an ALTER is now rejected with NOT_IMPLEMENTED, as it already was for other columns. #122794 (Groene AI).
• Fixed merges deleting all rows of a part whose rows TTL was never calculated, although none of its rows had expired: a part attached with ATTACH PARTITION ... FROM from a table without that rows TTL, or written before MODIFY TTL with materialize_ttl_after_modify = 0 added it, for example after MODIFY COLUMN ... REMOVE TTL. The rows were lost in background TTL merges and in merges with fully expired parts. Closes #121124. #123068 (Groene AI).
• Fixed a race in MergeTree where an ALTER TABLE ... UPDATE or DELETE mutation submitted concurrently with another mutation could be silently skipped on some parts while being reported as done. #124167 (Murphy).
• A mutation on a Replicated* table whose predicate contains a lambda can no longer prune to too few partitions when a lambda parameter name matches an ALIAS or EPHEMERAL column definition. Previously some matching rows were left unmutated while the mutation still finished successfully. #121193 (Groene AI).
• Fix background merges with TTL silently stopping for the whole server. A merge with TTL that was selected and then dropped before it ran — for example because its table was dropped right after the expired rows were inserted — leaked one of the max_number_of_merges_with_ttl_in_pool slots, which are shared by all tables. After two such merges no table would delete expired rows any more until the server was restarted. #121236 (Alexey Milovidov).
• Make the asynchronous startup of a ReplicatedMergeTree table robust against a concurrent DETACH, which could previously hang the startup forever or leave background tasks running on a table that was already shut down. #121672 (Alexey Milovidov).
• Fix corrupted shared data of nested JSON objects after a failed attempt to parse a value during insert. It could lead to errors like Unexpected version of JSON type binary encoding or Unknown type code when reading the data after merges or when serializing it in binary formats. #123898 (Rory Shanks).
• The server no longer terminates when a transient error such as MEMORY_LIMIT_EXCEEDED, a network error, or concurrent SYSTEM RESTART REPLICA happens while outdated data parts are loading in the background. Such errors are now retried later instead of aborting the server. #117981 (Alexey Milovidov).
• Fixed silent row loss when a MergeTree table with a projection is read with a WHERE in which an AND is used with negative polarity, under a NOT, as an operand of a comparison, or as the branch condition of if/multiIf/CASE, and one of that AND’s conjuncts references a column the projection does not store. #120286 (Groene AI).
• Fixed some SELECT queries silently returning fewer rows when the table has a projection that exists only in some of its parts, for example one added with ALTER TABLE ... ADD PROJECTION after data was inserted, and query_plan_remove_unused_columns is 0, which compatibility up to 25.11 sets. #124022 (Groene AI).
• Fixed wrong query results with skip indexes on SummingMergeTree, AggregatingMergeTree, CoalescingMergeTree, and GraphiteMergeTree tables: with optimize_on_insert, the default, an INSERT built every skip index whose expression is not a plain column from the rows before they were merged, so queries using the index could skip granules with matching rows. SummingMergeTree also no longer keeps rows whose summed columns total zero because of such an index. #123291 (Groene AI).
• Fixed wrong results when a WHERE clause mixing AND and OR was evaluated with skip indexes (use_skip_indexes_for_disjunctions, enabled by default). If a minmax or set index could not evaluate one atom of the predicate for a granule, the per-atom verdicts recorded for that granule were shifted, so one atom’s verdict was applied to another and granules holding matching rows were skipped. #124303 (Groene AI).
• Fixed a query silently returning fewer rows after an unrelated JOIN had run: a join runtime filter moved into a read’s PREWHERE made the read record a false “no matching rows” verdict in the query condition cache. #119823 (Groene AI).
• Fixed the query condition cache serving one session’s granule verdicts to another when the two differ only in a setting a conversion function captured, such as precise_float_parsing, which silently dropped matching rows from the second session’s result. #119168 (Alexey Milovidov).
• Fixed LEFT, RIGHT, FULL, and ANY outer JOIN silently dropping non-matched rows when the WHERE filter reads a Date32 or Enum column of the non-preserved side. Closes #121115. #121181 (Groene AI).
• Fixed a join runtime filter silently dropping rows whose join key is a float. When the build side had a single distinct value the filter replaced set membership with equals, but a hash join compares float keys bitwise, so NaN did not match itself and 0.0 matched -0.0. #119720 (Groene AI).
• Fix wrong values after lazy materialization expands a sparse column during a JOIN: rows could receive values from other rows. #124442 (Nikolai Kochetov).
• ALTER TABLE ... EXECUTE expire_snapshots on an Iceberg table now always keeps the head snapshot of every branch, even when the table property history.expire.min-snapshots-to-keep is 0. Previously such a table could lose all its data files. Closes #124333. #125365 (Groene AI).
• Fixed an ssl_certificate authentication bypass where an embedded NUL byte in a certificate’s Common Name (for example john\0.evil.com) was truncated server-side to john, letting a holder of such a certificate authenticate as a different user. #117403 (Konstantin Bogdanov).
• Fixed one failed write to a non-replicated MergeTree table’s deduplication log (with non_replicated_deduplication_window greater than zero) making every following INSERT into that table fail with Cannot write to canceled buffer until the server was restarted. The log now replaces a writer that a failed flush left unusable, no longer finalizes such a writer while rotating or shutting down, and no longer stops removing its outdated files, which could also make ALTER TABLE ... MODIFY SETTING non_replicated_deduplication_window fail. #118350 (Groene AI).
• Fixed wrong results after ALTER TABLE ... MODIFY COLUMN <column> <type> AFTER <other> on a column used in PARTITION BY. Closes #92834. #120862 (Groene AI).
• Fixed a merge or a mutation making a table unreadable (Stream ... is not found) when a column that a part does not hold is dropped and its name is also the name of a subcolumn of another column, for example a.size0 next to an Array column a. #118246 (Alexey Milovidov).
• Fix wrong rows returned by ORDER BY ... LIMIT over a MergeTree table read in the order of a function of its sorting key, for example ORDER BY toStartOfMonth(date) DESC, i, when the sorting-key column uses sparse serialization. #109424 (Groene AI).
• ALTER TABLE ... DROP PARTITION and other partition commands with an invalid or out-of-range Date or DateTime literal, such as '2024-02-30' or '9999-01-01', no longer silently act on a different partition. They now raise an error. #122419 (DeviousCardi).
• Fixed wrong query results and unsorted parts in MergeTree tables whose sorting, primary or partition key reads a subcolumn of a column (for example PRIMARY KEY cityHash64(`c1.1`) or ORDER BY t.a) when a column TTL or an UPDATE changes that column. A column TTL on such a column is now refused as on any other key column, UPDATE of a column that a MATERIALIZED key column of this kind is computed from is refused, and TTL merges of a table already stored with such a TTL fail instead of writing unsorted parts. #123628 (Groene AI).
• After a canary OOM, only the merges and mutations that are already running are cancelled. Newly scheduled ones now run as usual instead of all later merges and mutations staying cancelled until server restart. Closes #121748. #123870 (Javier G. Sogo).
• Fixed a spurious Cannot schedule a task: no free thread ... (CANNOT_SCHEDULE_TASK) error that rejected a valid SELECT over a MergeTree table when it had more parts to analyse than the query had threads and their analysis took longer than lock_acquire_timeout. Index analysis now runs one pool job per thread instead of one per part, which also makes it faster for queries over many parts. #116822 (Groene AI).
• Fixed wrong results and CANNOT_READ_ALL_DATA or INCORRECT_DATA errors when reading MergeTree parts in packed storage (min_bytes_for_full_part_storage), for example with parallel replicas, with a patch part of a lightweight update or delete, or when reading a JSON subcolumn. #116920 (Groene AI).
• Fixed ILLEGAL_COLUMN (Illegal column ... of first argument of function to...) raised by every query touching a column that was dropped and re-added with a different type, for as long as a part still carried the old column. The re-added column now reads as its default, as it does once the drop’s mutation rewrites the part. Queries served from a projection no longer return the old values of such a column. #118585 (Alexey Milovidov).
• Fixed reading stale data, or a Sizes of nested column and null map of Nullable column are not equal / Got empty stream for SerializationLowCardinality keys error, from a column that was turned into ALIAS or EPHEMERAL (or dropped while its partition was detached) and later made physical again, when a mutation that rewrites only some columns (for example MATERIALIZE INDEX) ran on a Wide part in between. #124516 (Groene AI).
• Fixed a use-after-free (on some versions a crash with Pure virtual function called!) on INSERT, merge or SELECT ... FINAL of a SummingMergeTree table with a Nested column whose name ends in Map and whose element type has LowCardinality inside an Array or Map, for example Array(LowCardinality(String)). #125533 (Groene AI).
• Fix a MergeTree table with a rows TTL and a text index never dropping its expired parts: the TTLDrop merge failed with Text index transform for index ... not found and retried forever. #113385 (Groene AI).
• Fixed dropping whole parts with TTL (ttl_only_drop_parts = 1) still reading all data from large, fully expired parts. #108626 (Michael Jarrett).
• Fixed a server crash on CREATE TABLE or ALTER TABLE ... MODIFY TTL when a TTL expression uses the result of arrayMap with a constant lambda, such as toString(arrayMap(x -> CAST(1, 'Dynamic'), arr)). #122868 (Groene AI).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK (Not found column _row_exists in block) that permanently blocked all further mutations of a MergeTree table after a lightweight DELETE run with lightweight_delete_mode = 'lightweight_update' was followed by a heavyweight ALTER TABLE ... UPDATE and then by another mutation rewriting all columns. Closes #115570. #115665 (Groene AI).
• Fixed ALTER DELETE and ALTER UPDATE silently skipping rows when the predicate references an ALIAS column whose definition contains a non-deterministic function such as now. #117273 (Alexey Milovidov).
• Fixed a patch part whose source_parts.dat lost its content to a zeroed block loading as an empty index, which silently unapplied an acknowledged lightweight UPDATE and then let background cleanup delete the only copy of it. Such a part is now detached as broken. #119585 (Alexey Milovidov).
• Fixed ALTER TABLE ... DELETE and ALTER TABLE ... UPDATE skipping a whole partition when the parts the initiating replica holds in it are all empty: rows another replica had acknowledged in that partition survived the mutation on every replica. #119149 (Alexey Milovidov).
• Fixed a pending lightweight DELETE read with apply_mutations_on_fly = 1 evaluating its predicate over patch parts created after it, which made rows disappear from the result and reappear after the mutation materialized. Closes #118049. #118919 (Alexey Milovidov). #119306 (Alexey Milovidov).
• Fixed lightweight UPDATE for parts adopted from another table via REPLACE PARTITION or ATTACH PARTITION FROM: after such a part was cloned again (by a mutation that does not modify it, by a clone instead of a fetch of a replicated mutation, or by a move between disks), applying patch parts could silently drop updated rows in release builds and raise the current_version != existing_version exception in debug and sanitizer builds. #115979 (Alexey Milovidov).
• Fixed lightweight UPDATE ... ON CLUSTER on ReplicatedMergeTree tables running on every replica of a shard instead of once per shard, which applied assignments such as SET v = v + 1 once per replica. A plain UPDATE in a Replicated database with more than one shard had the same problem. #122393 (Pratham Nayak).
• Fixed a concurrent heavyweight ALTER TABLE ... UPDATE/DELETE silently dropping an acknowledged lightweight UPDATE on a plain MergeTree table. #119155 (Alexey Milovidov).
• Fixed mutations such as ALTER TABLE ... UPDATE failing with CANNOT_READ_ALL_DATA on every attempt, and never completing, when they read a column together with its LowCardinality subcolumn in a Wide part: a Dynamic column and its LowCardinality(String) subcolumn, or a Map with map_serialization_version = 'with_buckets' and its keys or values subcolumn. The problem appeared in 26.6. #123160 (Groene AI).
• Fixed an acknowledged lightweight UPDATE being invisible on ReplicatedMergeTree when it targets a column whose RENAME COLUMN mutation has not materialized yet. #119144 (Alexey Milovidov).
• Fixed a lightweight UPDATE made before ALTER TABLE ... RENAME COLUMN on a ReplicatedMergeTree table being lost when a merge ran before the rename’s mutation (for example with apply_patches_on_merge = 0): queries returned the old value, and the mutation then dropped the update from disk. #124653 (Groene AI).
• Fixed a wrong result of a mutation of a MergeTree table when several pending mutations are executed as a single task and a patch part, created by a lightweight UPDATE, has a version in between: the patch was applied before the commands that precede it, so they were evaluated over data they must not see. #119127 (Alexey Milovidov).
• Fixed ReplicatedMergeTree TTL processing stalling indefinitely when a TTL merge that drops fully expired parts was larger than half the free disk space. Such a merge deletes every row and therefore needs no room for its source parts, but it was postponed on their size, and the postponed entry kept occupying the only max_replicated_merges_with_ttl_in_queue slot, so no TTL merge ran and no space was ever freed. It is now admitted when the table’s only TTL is an unconditional rows TTL. Closes #119775. #119983 (Groene AI).
• Fixed ATTACH PARTITION on ReplicatedMergeTree failing to rename a broken part in detached/ aside, so every later ATTACH PARTITION for that partition kept failing on it. #116734 (Groene AI).
• Fixed the replication queue never draining when a part is present locally but its node in ZooKeeper is missing: the GET_PART entry for such a part re-downloaded it and failed with DUPLICATE_DATA_PART on every retry. #119189 (Alexey Milovidov).
• Writing the empty covering part of a replicated table now happens before the covered parts are removed from the working set, so an exception during the operation no longer leaves the table with unexpected data parts and readonly after reload. #122610 (Alexey Milovidov).
• Background Keeper users now retry a get_zookeeper_lock_acquire_timeout_ms timeout like other Keeper errors, so a ReplicatedMergeTree table attached during a transient Keeper outage no longer stays read-only. #124327 (Groene AI).
• A table read by a scalar subquery inside a CONSTRAINT is now a loading dependency of the table that declares the constraint. Previously such a table could be dropped, after which the server did not start, because attaching the dependent table executes the subquery. #114923 (Alexey Milovidov).
• Fixed a server startup failure after creating a table with a quantile* state column whose level is a Decimal, a wide integer, a UUID, an IPv4, or an IPv6. Such a level was persisted into the metadata as a quoted string, which reparsed as a String and was rejected, so metadata loading failed with CANNOT_CONVERT_TYPE and the server did not start. The same CREATE also failed outright on a Replicated database. #119810 (Groene AI).
• A table whose stored definition carries a column TTL inherited from its source table now loads again on engines that do not support TTL (Merge, Buffer, Distributed, Remote). Previously such a table could not be loaded, which failed the load of the whole database, and with async_load_databases = 0 the server refused to start. The TTL stays in the definition and has no effect; ALTER TABLE ... MODIFY COLUMN ... REMOVE TTL removes it. #122273 (Mohammed Sohail).
• Fix a Remote or Cluster database without a local shard making SELECT count() through the merge table function return 0, and making the server fail to start when a dictionary reads one of its tables. Closes #122612. #122661 (Groene AI).
• Fix Part ... intersects previous part ... when attaching a MergeTree table after REPLACE PARTITION, if parts outdated by an earlier DROP PART were still on disk. The server could fail to start because of it. #120980 (Alexey Milovidov).
• Fix Part ... intersects next part ... when loading a MergeTree table after REPLACE PARTITION into it ran concurrently with MOVE PARTITION from it. The server could fail to start because of it. #123780 (Alexey Milovidov).
• Fixed S3, HDFS, AzureBlobStorage, and URL tables over a named collection with format = 'auto' NOT OVERRIDABLE failing to load with Override not allowed for 'format', which could stop the server from starting. The format that ClickHouse inferred and stored at CREATE time is now accepted when the table is loaded again; a format written by the user over the locked value is still refused. #124215 (Groene AI).
• Restored compatibility with Azure Blob Storage endpoints that omit the ETag header. #120419 (alesapin).
• Fix Part ... intersects previous part ... when loading a MergeTree table after a TRUNCATE, DROP PARTITION or MOVE PARTITION TO TABLE failed to commit (e.g. because of a memory limit or a disk error) and a merge ran before the restart. The server could fail to start because of it. #122375 (Alexey Milovidov).
• A PROJECTION that could not be analyzed when the table was loaded at server startup is no longer deleted from the table’s persisted metadata by the next ALTER, which also orphaned its already-materialized data with no error at any step. While such a declaration is present, an ALTER that rewrites the table metadata is now refused (ALTER TABLE ... DROP PROJECTION and CLEAR PROJECTION still work), so it survives until the cause is removed or the user drops it. #118378 (Groene AI).
• Fixed non-deterministic conditions being moved to PREWHERE in SELECT ... FINAL when optimize_move_to_prewhere is enabled. #109538 (Yarik Briukhovetskyi).
• Fixes wrong results with query_plan_optimize_lazy_final when the primary key is nullable and holds a NULL: the NULL-keyed row was silently missing from the result of a FINAL query. #112857 (Groene AI).
• Fixed rows silently dropped by index analysis when a DateTime64 column is compared with a constant of another scale and the sorting or partition key contains a CAST of that column, such as ORDER BY d::String. #118688 (Alexey Milovidov).
• Fix rows silently dropped by partition, primary-key, and skip-index pruning, and by Iceberg partition pruning, when a string constant is compared with a DateTime or DateTime64 column whose type name carries no time zone and the effective session_timezone differs from the one that column’s type was built with. #120114 (Groene AI).
• Fixed a minmax skip index, a set skip index, part-level minmax pruning, and statistics-based part pruning dropping rows that hold a NaN when the query negates a float comparison, for example WHERE NOT (val BETWEEN 0 AND 3). Closes #106533. Closes #106948. #107074 (Groene AI).
• Fixed silently missing rows for a range predicate on toDayOfWeek(key, mode) with the Sunday-first modes 2 and 3 over a date primary key. #117683 (Alexey Milovidov).
• Fixed ngrambf_v1, tokenbf_v1, and sparse_grams skip indexes silently dropping matching rows when an IPv6 column is compared with a string literal, as in WHERE ip = '2001:db8::'. #113157 (Groene AI).
• Fixed a set skip index whose expression has type Dynamic or Variant, which includes an undeclared JSON path such as j.a, silently returning no rows for != and IS NULL. #120780 (Groene AI).
• Fixed wrong results when a query converts a Date or Date32 column to DateTime and the column’s values leave the range of day numbers that fits into 32-bit seconds. Key analysis, statistics part pruning, and reading in sorting-key order assumed the conversion preserves order, so a filter could skip parts holding matching rows or answer count from metadata without filtering, and ORDER BY could return rows out of order. Affects toDateTime, toDateTime32, CAST(... AS DateTime), and accurateCast. #117031 (Groene AI).
• Fix rows silently dropped by primary-key pruning when toDayOfWeek is applied to a DateTime column whose time zone differs from the session one, or over a Date32/DateTime64 key that leaves the Date range. toDayOfWeek with a Sunday-first week_mode can use the primary key again. #118655 (Alexey Milovidov).
• Fixed rows silently missing from a query over a table whose primary key is a Tuple with a Nullable element holding NULL: primary-key analysis dropped whole granules of matching rows. #118945 (Alexey Milovidov).
• Fix a JSONAllPaths skip index silently pruning every granule for a filter on a subcolumn inside a declared typed path of a JSON column (for example json.a.b for JSON(a JSON) or JSON(a Tuple(b Int64))). Such a filter no longer uses the index. #119321 (Alexey Milovidov).
• Fix rows silently dropped by primary key and partition pruning when a MergeTree key expression applies a timezone-observing function such as toYYYYMM, toString or a CAST to String to a DateTime/DateTime64 column and a compared constant declares a different timezone. Both a scalar comparison constant and an IN-set element are affected, including when the differing timezone sits under an Array, Map, Tuple, Nullable, LowCardinality or Variant wrapper that only one side carries, when a custom type name such as SimpleAggregateFunction wraps that composite on one side only, and when a sibling leaf differs only by a custom type name (Bool against UInt8). Also fix the same loss when a compared Tuple and the key column’s disagree on explicit element names, which toJSONString renders as an object against an array; no timezone difference is needed. #113781 (Groene AI).
• Fixed rows silently dropped, or counted without being filtered, when a DateTime64 constant of a finer scale than the column is compared with !=, NOT IN, or IN: index analysis truncated the constant and treated the truncated value as the exact point the query asked for, pruning the granule or partition that held a matching row. #118957 (Alexey Milovidov).
• Fixed index analysis for a condition that wraps the primary key in a widening integer cast such as toInt64 and then applies another monotonic function such as negate or abs. Such a condition could silently drop matching rows, or fail with LOGICAL_ERROR: abs's argument does not match the expected data type. Present since 26.5. #116699 (Groene AI).
• Fixed wrong results from primary-key analysis for predicates such as toDate(dt64_key) >= ... and toDateTime(date32_key) >= ...: at the default date_time_overflow_behavior = 'ignore' the conversion wraps for values the result type cannot hold, so matching rows silently disappeared and count() could overcount. #117394 (Alexey Milovidov).
• Fixed partition pruning skipping a partition whose rows match a filter like NOT (c <=> f(partition_key)) when f returns NULL for that partition’s key value (for example moduloOrNull or nullIf). Such queries returned fewer rows than with use_partition_pruning = 0. #124041 (Groene AI).
• Fixed use_skip_indexes = 0 also disabling part pruning by the partition-key min-max index since 26.5, which made such queries read every part of the table. #123403 (Ashish Kumar Choubey).
• Fixed missing rows in queries mixing AND and OR when a minmax or set skip index over an expression with mortonEncode or hilbertEncode is combined with other skip indexes (use_skip_indexes_for_disjunctions, enabled by default). Closes #124452. #124784 (Groene AI).
• Fix a query that compared a Dynamic or JSON column with a constant, such as d = '2024-01-05' or d = toDate('2024-01-05'), and returned fewer rows when the table was ordered by a transform of that column, such as ORDER BY d::String. Such a comparison no longer uses the primary key; compare the key expression itself, as in d::String = '2024-01-05', to keep using it. #124476 (Ashish Kumar Choubey).
• Fixed a text index dropping matching rows when a FixedString is involved in the comparison. #118972 (Alexey Milovidov).
• Fixed wrong results of hasPhrase, hasToken, hasAnyTokens, and hasAllTokens over a text index with a postprocessor or preprocessor when direct read from the text index could not answer the function, for example NOT hasPhrase with a phrase whose tokens are all removed by a stop-word postprocessor. #125118 (Groene AI).
• Fixed text-index search functions returning wrong results, including silently dropped rows, on a column whose text index declares a preprocessor. #117774 (Elmi Ahmadov).
• Fixed wrong results after a mutation of a wide data part that does not store a column added by a metadata-only ALTER TABLE ... ADD COLUMN: the mutation wrote a column statistic for that column, so statistics-based part pruning could drop a part holding matching rows and min and max could be answered with the pre-ALTER default. #121936 (Groene AI).
• Fixed wrong results for min, max, and filtered count on a MergeTree column that carries a column TTL. When the TTL emptied the column, the part kept a statistic for it and queries were answered from that stale statistic instead of from the rows the part stores. #122058 (Groene AI).
• Fixed wrong results from use_query_condition_cache_for_top_k: SELECT ... SAMPLE ... ORDER BY ... LIMIT n no longer reuses cache entries written by the same query without SAMPLE, and Top-K query condition cache entries are no longer written for reads with on-the-fly mutations or patch parts. Closes #122564. #122538 (Nikita Fomichev). #122617 (Nikita Fomichev).
• Fix an exception NOT_FOUND_COLUMN_IN_BLOCK and, in other queries, wrong results when a lambda argument has the same name as an identifier of an expression that is bound to an alias outside of that lambda, for example SELECT number + 1 AS n, arrayMap(number -> number + n, [1, 2]) FROM numbers(2). #115544 (Alexey Milovidov).
• Fixed wrong results when a GROUP BY, DISTINCT, LIMIT BY, or window PARTITION BY key was plus or minus of a partition column with a non-integer constant. #120182 (Groene AI).
• Fixed a regression in 26.8 where columns of a parameterized view could not be qualified by the view name (SELECT pv.col FROM pv(p = 1) failed with UNKNOWN_IDENTIFIER) and a parameterized view used as a JOIN operand required an explicit alias. Closes #119837. #112481 (Groene AI). #120698 (Dmitry Novik).
• Fixed CANNOT_INSERT_NULL_IN_ORDINARY_COLUMN and CANNOT_PARSE_TEXT when a PREWHERE guard predicate and a throwing expression, such as a CAST or toUInt64(substring(...)), end up in the same read step: either because they read subcolumns of the same column, for example a JSON or Map column, or because a JOIN runtime filter’s key computation was merged into the guarded subquery. Such a query used to fail because the throwing expression was evaluated on rows the guard had rejected. #111954 (Groene AI).
• Fix wrong window function results when a QUALIFY condition, or a comparison derived from it by optimize_and_compare_chain, turns into an indexHint that prunes rows below the window. Closes #121177. #122975 (Yarik Briukhovetskyi).
• Fixed wrong aggregate values with optimize_aggregation_in_order_limit when the ORDER BY is a strict prefix of the GROUP BY key and a group’s rows span multiple parts. Closes #116849. #117573 (Konstantin Bogdanov).
• Fixed wrong results of sum, min, max, and avg over arithmetic with a Decimal operand and a constant, for example min greater than max, a changed sum of a division, or a missing DECIMAL_OVERFLOW: optimize_arithmetic_operations_in_aggregate_functions no longer moves such arithmetic out of the aggregate. #117779 (Alexey Milovidov).
• Fix <agg>IfOrNullIf and <agg>IfOrDefaultIf ignoring the outer condition when used without GROUP BY keys or as a window function with non-Nullable arguments. #122908 (Raúl Marín).
• Fixed a possible hang of a distributed GROUP BY query with uniqExact and distributed_aggregation_memory_efficient = 0 when merging large states of two-level aggregation. #123352 (Alexey Milovidov).
• Fixed missing rows of an empty grouping set () in GROUP BY GROUPING SETS queries without aggregate functions. For example, SELECT number FROM numbers(3) GROUP BY GROUPING SETS ((number), ()) returned 3 rows instead of 4, and GROUPING SETS (()) returned one row per input row. Closes #61461. #123539 (Groene AI).
• The uniq aggregate function returned garbage (for example 18446743978444128518) for cardinalities above roughly thirty billion, because its state kept only 32 bits of each hash and the correction formula overflowed. The state now additionally keeps a small sample of full 64-bit hashes (state version 1) which the estimate switches to at about two billion elements: SELECT uniq(number) FROM numbers_mt(100000000000) returns 99360630405 (−0.64%), and the estimate stays accurate (about 1%) up to trillions. Below two billion nothing changes: the estimates, the memory usage, and the state format are exactly as before, and states written by older versions remain readable and merge correctly in mixed-version clusters. #115393 (Alexey Milovidov).
• Fix wrong results for uniq*, GROUP BY, LIMIT n BY, ORDER BY after GROUP BY, and window PARTITION BY over toString of a DateTime/DateTime64 in a time zone with a daylight-saving transition, and over toFixedString of a String. Both functions claimed to be injective while mapping distinct arguments onto the same result, and the default-on optimizations that take that claim as a proof then eliminated the wrapper. #117238 (Alexey Milovidov).
• Fixed optimize_rewrite_aggregate_function_with_if (enabled by default) turning a short-circuit-guarded aggregate such as sum(if(number = 0, 0, intDiv(42, number))) into a query failure: the rewritten sumIf(intDiv(42, number), number != 0) evaluates its argument for every row, so the guarded expression ran on the rows the condition excluded and threw. The rewrite is now refused when the lifted branch relies on short-circuit evaluation. Closes #85784. #118697 (Groene AI).
• Fixed wrong results and errors of min, max, sum, and avg over an arithmetic operation with a constant with optimize_arithmetic_operations_in_aggregate_functions enabled (the default): on an Array or Tuple operand (min(arr * -1)), on an IPv4 or IPv6 operand (sum(ip * 1) threw ILLEGAL_TYPE_OF_ARGUMENT), for avg of a date, time, or interval (avg(t + INTERVAL 500 MILLISECOND)), for multiplication or division by zero, infinity, or NULL (min(x / -0.0)), and for a day or longer added to a DateTime or a Time64 across a DST change or a month end. Closes #122647. #123908 (Groene AI).
• Fixed hasAll, hasAny, hasSubstr, startsWith, and endsWith returning wrong results on Array(Nullable(T)). A NULL element was compared as whatever value was physically stored beneath it, so these functions disagreed with has on identical input: hasAll([1, 2, NULL, 4], [0]) returned 1 while has([1, 2, NULL, 4], 0) returned 0. Closes #68310. #116239 (Groene AI).
• Fixed wrong results from JIT-compiled conversions between integers and Decimal. With the default compile_expressions = 1, a conversion the destination could not hold silently wrapped instead of raising DECIMAL_OVERFLOW once the expression crossed min_count_to_compile_expression. Closes #118958. #118994 (Groene AI).
• Fixed if, multiIf, and CASE returning a wrapped value instead of raising DECIMAL_OVERFLOW once JIT-compiled, when a Decimal, DateTime64, or Time64 branch is lifted to a larger scale (with default settings, after the query had run a few times). #123186 (Groene AI).
• Fixed wrong results of extract and replaceRegexpAll (and its alias regexp_replace) under the default-enabled optimize_rewrite_regexp_functions: extract could return a different occurrence when a leading greedy ^.* was stripped, or a truncated result for a pattern ending with .*$ whose only parentheses are an inline flag group, such as (?i)b.*$; replaceRegexpAll could silently lose replacements when a pattern that can match more than once (with an inline (?m) flag, or a tail that can match the empty string, such as o*$) was rewritten to replaceRegexpOne; and both mishandled a trailing .*$ or $ that is literal text inside an unterminated \Q. #117228 (Alexey Milovidov). #117227 (Alexey Milovidov). #118361 (Alexey Milovidov).
• Fixed rows silently missing from the result of match with a pattern that has a zero-allowing quantifier behind an inline flag group or an empty \Q\E quote. #118351 (Alexey Milovidov).
• Fixed the best-effort date/time parsers reading one byte past the end of the value. The parsed result of a row of a String column could depend on the text of the following row: a valid value was rejected as NULL or zero, or addMonths and the rest of the add-interval family returned date and time fields taken from the neighbouring row. #122246 (Groene AI).
• Fixed toStartOfMinute, toStartOfHour, toStartOfDay, toStartOfFiveMinutes, toStartOfTenMinutes, toStartOfFifteenMinutes, timeSlot, toTimeWithFixedDate, toStartOfInterval and dateTrunc returning a value one second later, possibly in the next interval, for a DateTime64 before 1970 with a fractional part. For example, toStartOfMinute(toDateTime64('1969-12-31 23:58:59.500', 3)) returned 23:59:00. Also fixed toStartOfInterval with an origin and a MONTH, QUARTER or YEAR interval returning a value after its argument when the argument was earlier in the day than the origin, and rounding a bucket start clipped to the end of a shorter month to the previous bucket. #122138 (Alexey Milovidov).
• Fixed accepting a column whose two serialization streams map to the same file name (for example Nullable(JSON(`null` Int64)) or Tuple(`a` Tuple(`b` UInt64), `a.b` UInt64)). Such a table was created successfully, but its parts were written with both streams in one file and could not be read, merged, or checked. Creating such a column now fails with BAD_ARGUMENTS, as a collision between two different columns already did. Existing tables still attach, with the collision logged. #115632 (Pavel Kruglov).
• Fix a valid JOIN ON query rejected with TYPE_MISMATCH by optimize_redundant_comparisons, which is enabled by default since 26.7. It affects a join expression whose AND has a Nothing-typed operand, for example a join key compared against an ARRAY JOIN [] alias. #117476 (Groene AI).
• Fixed wrong results, or an ILLEGAL_COLUMN error, of a JOIN whose condition equates both a column and an expression of that column (for example ON t1.k = t2.k AND intDiv(t1.k, 2) = intDiv(t2.k, 2)) when a filter on that expression is pushed down to the other side of the join: the pushed-down copy applied the expression twice. Closes #122613. #122686 (Groene AI).
• Fixed SAMPLE on the first table of a JOIN being applied to the other joined tables too, which returned too few rows, or failed with SAMPLING_NOT_SUPPORTED when a joined table had no sampling key. #124767 (awss).
• Fixed missing rows from a Distributed or remote query whose WHERE has a NOT over an AND containing a conjunct that cannot be converted to an AST, such as a lambda or dictGet: the pushed-down predicate became stronger than the original one. #117620 (Alexey Milovidov).
• Fix wrong results when a query selects an ALIAS column together with a window function with enable_analyzer = 1, over a Distributed table (or remote) with two or more shards, or over a table read with parallel replicas. Values could be returned in the wrong columns with no error, or the query could fail with NUMBER_OF_COLUMNS_DOESNT_MATCH or CANNOT_PARSE_DATETIME. #116498 (Groene AI).
• Fixed a distributed SELECT silently returning a partial result when connection_pool_max_wait_ms is set to a finite value and skip_unavailable_shards = 1. A shard whose replica the initiator had no free connection slot for was reported as an unavailable shard and dropped, so a two-shard query could return one shard’s rows with exit code 0. Such a shard now fails the query with ALL_CONNECTION_TRIES_FAILED, the same error the default skip_unavailable_shards = 0 already produces. Closes #120832. #120969 (Groene AI).
• Fixed asynchronous TLS connections occasionally failing with an unrelated OpenSSL error left by another connection on the same OS thread. This could terminate distributed queries with a misleading network exception. #116692 (Dmitriy Borisenko).
• Fixed a query over a Distributed table silently returning too few rows when its WHERE calls joinGet: the condition was sent to the shards, and each one answered the call against its own local Join table rather than the initiator’s. joinGet is now treated as non-deterministic and stays on the initiator. #121822 (Nikita Taranov).
• Fix THERE_IS_NO_COLUMN for a Distributed query that has an IN over the sharding key outside its WHERE, in the SELECT projection, in GROUP BY, in ORDER BY, or in LIMIT BY. The default-enabled optimize_skip_unused_shards_rewrite_in rewrote that IN per shard, which changed the result column’s name, so the initiator could not bind the column the shard returned. #117252 (Alexey Milovidov).
• Fixed cancelling a distributed, s3Cluster or parallel-replicas SELECT taking up to receive_timeout (300 s by default) to take effect, and never reaching some of the remote sources, when one of them was draining the packets left over after a LIMIT was satisfied. #118696 (Groene AI).
• Fixed a server abort (vector[] index out of bounds) when reading a Distributed table with prefer_global_in_and_join = 1 and an IN expression that the analyzer shares between two parent expressions. The local to GLOBAL IN rewrite now rewrites the IN function in place, so every reference to a shared node is promoted. #120681 (Groene AI).
• Fixed INSERT INTO <replicated table> SELECT ... FROM url(...), and the same shape with s3 or azureBlobStorage, with parallel replicas enabled inserting the data once per replica. Closes #120485. #120488 (Alexey Milovidov).
• Fix INSERT INTO <Distributed table> SELECT ... FROM url(...) (and s3, azureBlobStorage, …) with parallel replicas enabled inserting the data once per shard of the destination cluster: the query forwarded to the shards now names the *Cluster table function with the cluster of the Distributed table, so every shard reads only its share of the files. Such an INSERT also no longer writes into the shards’ local tables when the destination has a deterministic sharding key, because the rows a cluster table function hands to a shard cannot satisfy that key and the resulting placement made optimize_skip_unused_shards return partial results; it falls back to the ordinary INSERT ... SELECT instead, which places the rows through the Distributed sink. #121070 (Igor Nikonov).
• Fixed duplicated rows returned by ANY RIGHT JOIN with any_join_distinct_right_table_keys = 1 when parallel replicas are enabled. #123683 (Groene AI).
• Fixed CREATE MATERIALIZED VIEW ... POPULATE and CREATE TABLE ... AS SELECT in a Replicated database with parallel replicas enabled: each replica sent its population to other replicas, so its rows landed in other replicas’ tables (duplicated there, missing on the replica itself), or the query failed with UNKNOWN_TABLE. The population now ignores the query’s parallel replicas and parallel_distributed_insert_select settings, so each replica fills only its own table. Closes #124446. #124477 (Groene AI).
• Fix S3 client refresh overwriting static authentication settings and breaking subsequent queries. #120433 (alesapin).
• Fix a server abort when an Azure Blob Storage endpoint reports an asynchronous blob copy as completed without the optional x-ms-copy-source or x-ms-copy-status-description header. #120935 (Alexey Milovidov).
• Reads from Azure Blob Storage now verify that the returned range starts at the requested offset, so an endpoint or proxy that ignores the request Range fails the read with HTTP_RANGE_NOT_SATISFIABLE instead of delivering the wrong bytes. #120955 (Alexey Milovidov).
• A bounded read from Azure Blob Storage whose response ends before the requested right bound is now reopened and, if the endpoint keeps answering short, fails with UNEXPECTED_END_OF_FILE instead of silently returning a truncated file. #120963 (Alexey Milovidov).
• Fix a buffer over-read on the sequential read path from Azure Blob Storage: an endpoint that answers a ranged request with more data than was requested could deliver bytes from outside the requested range to the caller. The end of a read is now decided by the locally set right bound, not by the Content-Length of the response, and a read is pinned to the listed generation of the blob with If-Match, so an object replaced during a read fails with AZURE_OBJECT_CHANGED_DURING_READ instead of being truncated (new setting azure_validate_etag_on_read, enabled by default). #120910 (Alexey Milovidov).
• Fixed a disk with object_storage_type = 'local' and a relative path removing its own root directory once the last blob in it was deleted. Anything else stored in that directory was lost with it; in particular, a disk whose path pointed at the directory of its own local metadata became unusable. #119119 (Alexey Milovidov).
• Fixed S3 requests that keep failing with a retryable error being attempted (s3_retry_attempts + 1)^2 times instead of s3_retry_attempts + 1. Fixed an error that a plain-HTTP S3 endpoint sends before reading the whole request body (for example NoSuchUpload for an expired multipart upload) being retried as a Broken pipe network error instead of being reported. Closes #124326. #124557 (Groene AI).
• Fixed a heap use-after-free in the s3, azureBlobStorage, hdfs, and similar object-storage table functions and engines when an exception such as MEMORY_LIMIT_EXCEEDED was thrown while creating the reader of a file in a row input format such as CSV. #125090 (Groene AI).
• Fixed exceptions such as UNKNOWN_CODEC when reading MergeTree tables through the userspace page cache (use_page_cache_for_disks_without_file_cache = 1 with remote_filesystem_read_method = 'read', or use_page_cache_for_local_disks = 1): files kept inside packed archives (statistics, skip indices, parts in packed storage) were read from the wrong offset. Closes #122157. #122949 (Groene AI).
• Fix Parquet time32/time64 columns being read as DateTime64 and then silently shifted by session_timezone when inserted into a Time64 target column. Parquet TIME logical types now map directly to Time64, and the missing Time64 -> Time cast is added. #106019 (Shaohua Wang).
• Fixed a Parquet reader bug where a filter on an integer column read with a type hint that reorders the file’s statistics (for example UInt32 over a plain INT32 column, Enum8 over a UINT_8 one, or IPv4 over a UINT_64 one) could silently return fewer rows than the file holds. Such statistics are no longer used for row group or page pruning, and a DATE value outside the requested type’s range now raises the same error with pruning as without it. Closes #118376. #121033 (Groene AI).
• Fixed an infinite loop in the Parquet reader on a ZSTD or Brotli compressed page that decompresses to fewer bytes than its header declares. Such a query could not be stopped by max_execution_time or KILL QUERY; it now fails with INCORRECT_DATA. Closes #123274. #123298 (Groene AI).
• Fix Unexpected number of rows in column subchunk (LOGICAL_ERROR) when filtering a Parquet file whose Bool column is RLE-encoded (pyarrow’s default for data page v2). The filter is now applied while decoding RLE-encoded and dictionary-encoded pages, which also makes PREWHERE over dictionary-encoded columns of Parquet files without a page index faster (ClickBench q31 1.92 s to 1.21 s). Closes #123483. #123618 (Groene AI).
• Fixed wrong results when reading a Parquet file with a structure hint whose type differs from the stored one. Row-group and page pruning, bloom filters, and dictionary filters now hash and interpret the query constant in the stored physical type rather than the hinted one, which fixes cases such as narrowing, signedness flips, Enum, Date/Date32/Time, length-mismatched String/FixedString, and Bool over an integer column. Reading a UINT_8 column with a Bool hint now also treats every nonzero value as true, so x = true, x IN (true), and WHERE x agree. #123864 (Groene AI). #121077 (Groene AI). #124285 (Groene AI).
• Fixed Parquet output failing with Capacity error: array cannot contain more than 2147483646 bytes or Uncompressed page is too big when a batch of rows held more than 2 GiB of String or FixedString data. A batch is now bounded by bytes as well as by rows, which also lowers the peak memory of the write. #122482 (Alexey Milovidov).
• Fixed SELECT ... FORMAT Arrow failing with Method getDataAt is not supported for JSON, Dynamic, and QBit columns, and silently exporting heap addresses instead of the aggregate state for AggregateFunction columns. Added the output_format_arrow_unsupported_types setting (throw / text / binary, default binary), which supersedes output_format_arrow_unsupported_types_as_binary. Such columns can now also be read back into their original type, which previously failed or silently corrupted the value. INSERT INTO ... ENGINE = ArrowFlight now follows this setting too; it previously rejected such columns unconditionally, so an insert that used to fail may now write an opaque column. Set output_format_arrow_unsupported_types = 'throw' to keep rejecting it. #115250 (Yakov Olkhovskiy).
• Fix O(n^2) decompression of gzip/zlib/deflate streams whose single DEFLATE block spans the whole stream, the shape zlib-ng produces at compression level 1 (the default of the official .NET SDK): a 24 MiB gzip HTTP body that took 11 s to ingest in 26.7 now takes 0.12 s, identical to a multi-block body, with memory bounded by the buffer size instead of the block size. #114059 (Alexey Milovidov).
• Fixed a segmentation fault or std::length_error when reading TSV formats with input_format_tsv_crlf_end_of_line = 1 if an escape sequence ended exactly at the end of a read buffer. #123861 (Groene AI).
• Fixed lost rows when a query over a MySQL, PostgreSQL, SQLite, or ODBC/JDBC table compares a column, string, number, or Decimal, with an Enum constant, also inside a tuple, an IN list, Nullable, Variant, or Dynamic, at any depth. The filter sent to the external database used the Enum’s number, while ClickHouse compares a string column with the Enum’s name, and PostgreSQL rejected such a query. An IN list or row of Variant or Decimal constants is no longer sent as one quoted string. #123843 (Groene AI).
• Fixed wrong results of a negated IN with a NULL in the list, e.g. x NOT IN (1, NULL) or NOT (x IN (1, NULL)), over MySQL, PostgreSQL, SQLite, and ODBC tables: the filter pushed to the external database now leaves the NULL members out of the list, the way ClickHouse evaluates the set. #124339 (Groene AI).
• Fixed silent data loss when an EmbeddedRocksDB table is renamed. RENAME TABLE now moves the table’s RocksDB directory to the table’s new location, so the rows are still there after a server restart. Previously the directory was left behind, and after a restart the table opened a fresh empty RocksDB while the real rows stayed orphaned under the old path. #112224 (Groene AI).
• optimize_inverse_dictionary_lookup (enabled by default) no longer rewrites dictGet(dict, attr, key) = const when a Variant, Dynamic, or JSON appears in the type of key or in the dictionary’s declared key type. The rewritten predicate is not equivalent there, so the query could fail with CANNOT_INSERT_NULL_IN_ORDINARY_COLUMN, ILLEGAL_TYPE_OF_ARGUMENT, or NO_COMMON_TYPE, or return wrong results. #120888 (Groene AI).
• Fixed SQL injection through identifiers in the PostgreSQL integration: relation, schema, column, publication, and replication-slot names used by the PostgreSQL table engine, the postgresql table function, PostgreSQL dictionary sources, and MaterializedPostgreSQL were escaped with backslashes, which PostgreSQL does not honour inside a quoted identifier, so a remote name containing a double quote could end the identifier and let the rest execute on the PostgreSQL server. Relations and columns whose names contain a double quote or a backslash can now be read and replicated, including those listed in materialized_postgresql_tables_list; an empty element in that setting is now rejected. #119216 (Groene AI). #119103 (Groene AI).
• Fixed PARALLEL WITH, EXECUTE AS, and the QueryRunner table engine skipping the access checks of the CREATE statements they execute, which allowed creating tables and databases without the corresponding privileges and defeated the readonly and allow_ddl settings. #119848 (Alexey Milovidov).
• Fixed an HDFS database serving a table from its internal cache without the READ ON HDFS grant. The grant and the remote_url_allow_hosts filter were checked only while a table name was resolved for the first time, so once any user holding the grant had read a name, every user with SELECT on the database was served the contents of the file, and that table kept being served after its host was removed from the allow-list. EXISTS TABLE, which requires only SHOW TABLES, answered from the same cache and so reported which names other users had resolved. Both checks now run before the cache is consulted, and the existence probe no longer consults it. #119523 (Groene AI).
• Fixed filesystem containment checks accepting a path with an embedded NUL byte. A DeltaLakeLocal or IcebergLocal table could be pointed outside of user_files this way. #119846 (Alexey Milovidov).
• Fixed several engines disclosing an Alias table’s target schema to a user who holds no privilege on that target. Reading the metadata of an Alias through any of them now requires table-level SHOW COLUMNS on its target, which is what a DESCRIBE of the target already requires. #117572 (Groene AI).
• remote(<cluster name>, 'db', 'table', 'user', 'password'), and the same form of remoteSecure and of the Remote and RemoteSecure table engines, no longer write the password into system.query_log, system.text_log, and the server log in cleartext. #121474 (Groene AI).
• Fixed SQL injection through identifiers when inserting into a table with the SQLite engine. It is now also possible to insert into a SQLite table whose name, or column name, legitimately contains a double quote or a backslash. #117421 (Groene AI).
• Fixed credential masking for nats_url, rabbitmq_address, the jdbc and odbc connection string, and the format_avro_schema_registry_url, url_base, and s3_base settings in SHOW CREATE, system.tables, and the query text in system.query_log. Closes #121257. #121425 (Groene AI).
• Fixed a Backup database engine printing the credentials of its destination in plaintext. SHOW CREATE DATABASE, system.databases.engine_full, system.query_log, and the messages that reject a malformed locator no longer expose an AzureBlobStorage account_key, a connection string, or a presigned URL regardless of format_display_secrets_in_show_and_select. #117429 (Groene AI).
• KILL TRANSACTION ... ON CLUSTER, KILL PART_MOVE_TO_SHARD ... ON CLUSTER, SYSTEM STOP THREAD FUZZER ON CLUSTER, SYSTEM START THREAD FUZZER ON CLUSTER, and SYSTEM RESET COVERAGE ON CLUSTER no longer run without checking any privilege. Each now requires what its non-ON CLUSTER spelling has always required. #118360 (Groene AI).
• Fixed metadata disclosure through a chain of Alias tables. DESCRIBE, hasColumnInTable, system.columns, system.tables, system.projections, system.constraints, and system.data_skipping_indices now require the privilege on every table an Alias resolves through, as reading its rows already did. #118967 (Groene AI).
• Fixed a bug where REVOKE GRANT OPTION FOR was accepted but then omitted from SHOW GRANTS, from system.grants, and from the persisted access metadata whenever another revoked database, table, column, or named-collection name shared a prefix with it. The grant option therefore came back after a server restart, and the user could pass it on with GRANT CURRENT GRANTS. #122078 (Groene AI).
• Fixed settings constraints from a settings profile not being enforced against the other setting assignments in the same statement. SET profile = 'p', <setting> = <value> could override a CONST, MIN/MAX, or readonly constraint that profile p installs. The same escape existed through an in-query SETTINGS clause, HTTP URL parameters, and Arrow Flight session options. #120376 (Raúl Marín).
• Fixed row policies on Memory tables being ignored when a read is executed from a serialized query plan, such as a distributed plan or a parallel-replicas plan. Since Memory gained PREWHERE support, such reads could return rows the policy excludes. #122503 (Alexey Milovidov).
• Fixed an integer underflow in the MySQL protocol handshake handler that could cause the server to buffer an unbounded amount of data from an unauthenticated client. #115776 (Dmitrii Tishkin).
• Statistics-based part pruning is now disabled for tables with row policies. #122035 (alesapin).
• A row policy on a Set-engine table was accepted but never applied: the right side of IN read the prebuilt set directly and returned the rows the policy hides. Such a query now fails with ACCESS_DENIED, because the Set engine has no read path that could filter those rows, and IN set_table now requires SELECT on all columns of the table. On Replicated databases, ALTER ... DELETE, ALTER ... UPDATE, DELETE, and UPDATE now run the initiator’s checks as the submitting user, so the same row-policy and access checks apply there too; with ON CLUSTER, run the query on a host that has the right-hand Set table, or enable distributed_ddl_use_initial_user_and_roles on every host and set distributed_ddl_entry_format_version to at least 8. Queries that used a Set table covered by a database-wide row policy ran unfiltered before and now throw. #120312 (Harikrishnan Prabakaran). #122141 (Alexey Milovidov).
• Fixed the Kafka and NATS table engines ignoring remote_url_allow_hosts: broker addresses from kafka_broker_list, nats_url, and nats_server_list are now checked against the remote host filter. A forbidden address is rejected at CREATE TABLE with UNACCEPTABLE_URL, as the RabbitMQ engine already does. #121825 (János Benjamin Antal).
• HTTP headers in HTTP dictionary sources are now hidden in SHOW CREATE DICTIONARY, system.tables, and query logs, like passwords. They can still be shown with display_secrets_in_show_and_select. #122673 (Matheus Nerone).
• Fixed secret settings of S3Queue and AzureQueue tables written with the legacy s3queue_ prefix, such as s3queue_after_processing_move_secret_access_key, being shown in cleartext in SHOW CREATE TABLE, system.tables, system.query_log, and the server log. They are now hidden as they are without the prefix. #123060 (Groene AI).
• Column-level SELECT privileges are now enforced for columns used only inside indexHint and for ALIAS columns referenced in PREWHERE. As in WHERE, referencing an ALIAS column in PREWHERE requires the grant on the alias name; a grant on its source columns is no longer sufficient. A metadata-only ALTER that changes a subcolumn read by a MATERIALIZED column through an ALIAS is now refused like a direct reference. #120516 (Shankar Iyer).
• Fixed a row policy bypass: a direct JOIN with an EmbeddedRocksDB, KeeperMap, or Redis table in the legacy analyzer (enable_analyzer = 0) ignored the row policy of the right table and returned hidden rows, and joinGet on a Join table ignored its row policy in both analyzers. Such a table is now joined through its query plan, which keeps the filter, and JOIN or joinGet against a Join table under a row policy throws ACCESS_DENIED. #121398 (Vladimir Cherkasov).
• CREATE OR REPLACE in a Replicated database or with ON CLUSTER now requires the drop privilege for the kind of the table it replaces, so a user without DROP DICTIONARY can no longer replace a dictionary. #122272 (Nikolay Degterinsky).
• A view with SQL SECURITY DEFINER or SQL SECURITY NONE that can hide rows is now read as a whole and is opaque to the optimizations of the query that reads it; previously an expression of the outer query could be evaluated on the hidden rows and reveal them, for example through an exception message. EXPLAIN shows the plan of such a view only to users who could have created it or who may display secrets, and seeing the plan of such a sealed view now also requires the CREATE VIEW privilege. #122330 (Alexey Milovidov). #123325 (Alexey Milovidov).
• Fix a credential written in cleartext to system.query_log and the server log when a key = value argument precedes the positional arguments of the mongodb, redis, remote, arrowFlight, and ytsaurus table functions or the MongoDB, Redis, Remote, ArrowFlight, ExternalDistributed, and YTsaurus table engines. A positional argument after a named one is now always hidden, and the value of a named-collection override whose key is written as an expression is hidden for every such engine. #122770 (Raúl Marín).
• Rewriting the body of a view with ALTER TABLE ... MODIFY QUERY now needs the same grants as creating it with its SQL security: SET DEFINER on the definer of a SQL SECURITY DEFINER view, or ALLOW SQL SECURITY NONE for a SQL SECURITY NONE view. An ON CLUSTER MODIFY QUERY fails with UNKNOWN_TABLE unless it runs on a host that has the view. #121629 (Sema Checherinda).
• Fixed GRANT ... ON db*.table being silently rewritten to db.table*, which could cause incorrect privilege grants. The pattern db*.table is now rejected as a syntax error. CREATE / ALTER / DROP ROW POLICY, SHOW POLICIES, and SHOW CREATE POLICY now also reject prefix wildcards (db*.*, db.table*, table*), which were previously silently narrowed to the exact name (db.*, db.table, table). #108063 (Aditya Kumar).
• Update mongo-c-driver to 2.5.6 and mongo-cxx-driver to r4.6.0, fixing an out-of-bounds write during MongoDB connection monitoring reconnects (CVE-2026-96746). #119346 (Actuele AI). #124282 (Actuele AI). #124648 (Actuele AI).
• Fixed access entities (users, roles, row policies, and others) of the local_directory access storage being lost after a restart, or a dropped row policy making every query fail with FILE_DOESNT_EXIST. This happened to entities created or dropped within a minute after SYSTEM RELOAD USERS if the server was then killed, or at any time after SYSTEM RELOAD USERS failed to write the storage’s .list files. #125058 (Groene AI).
• The web terminal and the SSH interface no longer reset the settings of the user (from settings profiles, roles, and CREATE USER ... SETTINGS) to default values. Previously, restrictions such as readonly, resource limits, and additional_table_filters were not applied in these interfaces. #125515 (Alexey Milovidov).
• Update expat to 2.9.0, fixing a denial of service through a crafted backup manifest during RESTORE (CVE-2026-66046) and the other 2.8.3–2.9.0 security fixes. #124255 (Actuele AI).
• Fixed failures during pipeline expansion, such as external sorting or the merge stage of a GROUP BY, when the memory limit is exceeded. The query now reports the original exception instead of an inconsistent-pipeline error, a segmentation fault, or a read of freed memory. Closes #125251. #121502 (Alexey Milovidov). #125283 (Groene AI).
• Fixed a stored query combining UNION with a parenthesized INTERSECT/EXCEPT group losing its parentheses when serialized, so X UNION ALL (Y EXCEPT Z) was read back as (X UNION ALL Y) EXCEPT Z. Views returned wrong results after a server restart, and a materialized view could end up with metadata the server refuses to load, which prevented it from starting. #111931 (Groene AI).
• Fixed the compatibility setting: compatibility = '26.8' no longer reverts five defaults that 26.8 shipped or forces MergeTree patch_parts_version back to v1, and compatibility = '26.1' no longer disables use_skip_indexes_on_data_read, which 26.1 shipped enabled. #118666 (Alexey Milovidov).
• Fixed compatibility = '25.11' incorrectly resolving query_plan_max_limit_for_lazy_materialization to 10 instead of 100. #122768 (Raúl Marín).
• Setting the obsolete enable_analyzer (allow_experimental_analyzer) to 0 is accepted again and replaced with 1, instead of failing with SETTING_CONSTRAINT_VIOLATION or INCORRECT_QUERY. Queries, sessions, settings profiles, and client configurations that still set it keep working after an upgrade, always with the analyzer, including configurations that spell the value as a Float64 literal. #122401 (Nikita Fomichev). #122763 (Raúl Marín).
• Fixed a server crash on AArch64 that could happen when the query profiler sampled a thread while it was connecting to a remote server, for example with remote or remoteSecure. Capturing stack traces, used for exceptions and by the query profiler, is also several times faster. #122834 (Raúl Marín).
• Fixed listen_host :: on hosts without a global IPv6 address, such as a container on an IPv4-only Docker network: it failed with a DNS error, so the official Docker image listened on IPv4 only and localhost over ::1 was refused. The server now listens on the dual-stack wildcard there. gRPC and Arrow Flight use one listener when :: is the only wildcard, and host_regexp / HOST REGEXP / HOST LIKE restrictions match IPv4 clients of a dual-stack listener. #122785 (Groene AI).
• Fixed an issue where pressing Ctrl-R and then Ctrl-Z in the web terminal or the SSH interface suspended the whole server process. #125512 (Alexey Milovidov).
• Fixed TOO_LARGE_STRING_SIZE (Too large sizes of FixedString to deserialize) when reading a correctly written part whose Array(FixedString(N)) column holds more than 1 GiB of elements in one read span. Such a span is now read in bounded steps instead of being rejected, so the affected parts can be read and merged again without being rewritten. Closes #116772. #117385 (Groene AI).
• Fixed CollapsingMergeTree returning wrong rows when a row’s sign is neither 1 nor -1. #119900 (Murphy).
• Fixed a table that exceeds max_table_size_rows never being able to come back under the limit: every merge threw TABLE_SIZE_LIMIT_EXCEEDED, so the table could not be compacted and its part count grew until inserts started failing. #117229 (Alexey Milovidov).
• Fixed a shutdown stall where a background merge or mutation left in a MergeTree background executor’s queue kept holding its table lock, so any table lock taken later during shutdown waited for the full lock_acquire_timeout and then failed with DEADLOCK_AVOIDED. #113653 (Groene AI).
• Fixed count() returning the number of default values instead of the number of NULLs when optimize_trivial_count_with_sparsity_filter is used on a column widened to Nullable by a metadata-only ALTER TABLE ... MODIFY COLUMN. #120245 (Raúl Marín).
• Fixed MergeTree tables with non_replicated_deduplication_window on object storage failing to load their deduplication log after being detached and attached back without inserts in between, and fixed the deduplication log possibly dropping its active file. #121670 (Alexey Milovidov).
• Fixed ALTER TABLE ... COMMENT COLUMN and ALTER TABLE ... RENAME COLUMN making reads of an ALIAS column’s subcolumns, for example arr.size0, fail with NO_SUCH_COLUMN_IN_TABLE until the table was reloaded. Closes #121421. #121449 (Groene AI).
• Fixed CANNOT_READ_ALL_DATA (while reading prefix of column ...) during query planning when a MergeTree table has an active empty Wide part, left by a TTL drop or by a mutation that deleted all rows of a part, and the query filters on a subcolumn, such as a key of a Map(LowCardinality(String), String) column. #122756 (Groene AI).
• Fixed ATTACH PART failing with CORRUPTED_DATA for a Compact part whose default_compression_codec.txt is missing, for example a part copied or restored without that file. #119114 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR (Sizes of nested column and null map of Nullable column are not equal after deserialization) that made a SELECT of a Nested member fail with Code: 49 until the server was restarted, after an ALTER TABLE ... MODIFY COLUMN on that member had been killed and a later mutation hardlinked the column into a new wide part. #121561 (Groene AI).
• Wide parts no longer split a fixed-width value across two compressed blocks when max_compress_block_size is not a multiple of the value size. This improves the compression ratio of type-specific codecs on such tables and fixes CANNOT_COMPRESS from ALP. #122841 (Raufs Dunamalijevs).
• After ALTER TABLE ... RENAME COLUMN, the subcolumns of the renamed column, for example c0.null or a.size0, stayed resolvable under the old name. Reading them through a Buffer or Merge table, or using them in OPTIMIZE ... DEDUPLICATE BY, read uninitialized memory, and once a new column reused the old name they returned wrong values instead of throwing UNKNOWN_IDENTIFIER. #122527 (Groene AI).
• Fixed excessive memory usage when loading a data part with a corrupted columns_substreams.txt; such a part is now detached as broken. #124157 (Murphy).
• Fixed a LOGICAL_ERROR and silent data corruption when a column is named into the skp_idx_ prefix so that its on-disk stream base equals a skip index’s, making both share one marks file. Such a write is now rejected with INCORRECT_FILE_NAME naming both owners. #113093 (Groene AI).
• Fixed reads of a renamed Nested leaf column returning default values from wide parts while the RENAME COLUMN mutation was still pending. #119134 (Alexey Milovidov).
• Fix CANNOT_COMPRESS on merge or insert for a Map column with with_buckets serialization and CODEC(T64). #124695 (Raufs Dunamalijevs).
• Fixed a column added by ALTER TABLE ... ADD COLUMN ... DEFAULT <expression> being filled with the type default instead of the evaluated expression when it is read through a single-conjunct explicit PREWHERE on the expression’s dependency. #117533 (Alexey Milovidov).
• Fixed the order of rows in a data part written by a merge or a mutation with TTL ... GROUP BY ... SET when the SET clause assigns a new value to a column the sorting key is calculated from. Closes #108514. #119863 (Alexey Milovidov).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK when merging a table that has a rows TTL together with a move (TO DISK, TO VOLUME) or recompression (RECOMPRESS) TTL over a different column. #121536 (Murphy).
• Fixed INSERT into a table with a constant TTL ... DELETE WHERE whose condition is not exactly UInt8, for example c > 10 over a Nullable column: a part where no row matched the condition was marked with the delete TTL and later rewritten by a needless TTL merge, and a condition that cannot be a filter, such as an Array column, was accepted instead of being rejected as in the other TTL code paths. #122516 (Groene AI).
• Fixed INSERT into a table whose TTL expression changes its type under a session setting, for example a CAST of a Nullable column under cast_keep_nullable = 1, which failed with LOGICAL_ERROR Unexpected type ... of result TTL column. Under such a setting, ALTER TABLE ... MODIFY TTL could also store a TTL that CREATE TABLE rejects, after which the table could not be attached again. A stored TTL is now analyzed with the server settings in INSERT and ALTER, as in CREATE TABLE. #122973 (Groene AI).
• Fixed an unsorted part with a primary index that did not match the data after a TTL GROUP BY ... SET merge that rewrites the column a subcolumn sorting key (such as ORDER BY n.a) is taken from. Queries with a condition on the sorting key could return wrong results. #122373 (Alexey Milovidov).
• Fixed ALTER TABLE ... MODIFY TTL accepting a TTL whose subquery reads a table function (for example numbers or remote), a view over one, a parameterized view, a temporary table, or the table itself. After such an ALTER the table could not be loaded again, so a server with synchronous table loading did not start. Now the ALTER fails. #123561 (Groene AI).
• Fixed standalone UPDATE ... SET substituting a WITH expression alias into a nested subquery that cannot resolve it. Where that name is a table name, the statement read the alias instead of the table, or silently succeeded where the equivalent SELECT raises UNKNOWN_IDENTIFIER. #120026 (Groene AI).
• allow_non_metadata_alters = 0 now also rejects the ALTER commands that reach a MergeTree table as explicit mutations: UPDATE, DELETE WHERE, the five MATERIALIZE forms, APPLY DELETED MASK, APPLY PATCHES, and REWRITE PARTS. #115287 (Groene AI).
• Fixed Cannot convert NULL value to non-Nullable type in ALTER TABLE ... UPDATE with a scalar subquery when the subquery result is Nullable and the updated column is not. #119926 (Alexey Milovidov).
• ALTER TABLE ... DROP COLUMN and ALTER TABLE ... CLEAR COLUMN are now refused when they would leave a part with no columns at all, instead of wedging the mutation queue with a logical error or leaving an unreadable part. #121104 (Alexey Milovidov).
• Fixed a mutation such as ALTER TABLE ... MODIFY COLUMN or RENAME COLUMN committing a data part with no columns when the part held only columns the table no longer has (for example a partition detached before DROP COLUMN and re-attached after it), after which every read of the table failed with a logical error. Such a mutation now fails with NO_SUCH_COLUMN_IN_TABLE, like reads of that part, and the part and its data are kept. #122400 (Groene AI).
• Fixed ALTER TABLE ... UPDATE and MATERIALIZE COLUMN on a Memory table after MODIFY COLUMN of the updated column: later reads failed with ILLEGAL_COLUMN or returned wrong values, and BACKUP failed. #123043 (Groene AI).
• Fix Found patch part ... that intersects mutation with version ... on lightweight updates in a table created with CREATE TABLE ... CLONE AS or ATTACH PARTITION ALL FROM a table with lightweight updates. #123620 (Nikolay Degterinsky).
• A synchronous mutation (mutations_sync) on a MergeTree table no longer reports success when the table is detached before the mutation runs; it fails with UNFINISHED, and the mutation continues after the table is attached again. #124161 (Murphy).
• DELETE, UPDATE and ALTER ... DELETE/UPDATE no longer refuse a MergeTree table of a database with lazy_load_tables that has not been accessed yet, and SYSTEM DROP REPLICA ... FROM ZKPATH no longer skips such a table when checking that the replica being dropped is not a live local one. #118152 (Alexey Milovidov).
• Fixed a ReplicatedMergeTree table whose metadata carries table_readonly = 1 being unusable: its backup could not be restored, it could not be re-attached after DETACH, and ALTER TABLE ... MODIFY SETTING table_readonly = 0 was refused. Converting a table with that setting to replicated is no longer done, by either the convert_to_replicated flag or ATTACH TABLE ... AS REPLICATED. #119224 (Alexey Milovidov).
• CHECK TABLE on a ReplicatedMergeTree replica whose Keeper session was lost now fails with ABORTED instead of returning 0, which means that the check found broken data. Nothing was read in that case, so the healthy table was certified as broken. #119381 (Alexey Milovidov).
• Fixed the automatic MergeTree to ReplicatedMergeTree conversion silently not finishing for a table in a database with lazy_load_tables, which left the table read-only with no metadata in ZooKeeper and the conversion flag in place. #118135 (Alexey Milovidov).
• Fix an integer overflow in the DDL log retention check of Replicated databases that could cause premature removal of DDL log entries and replicas being incorrectly declared lost for a very large logs_to_keep. logs_to_keep is now limited to 2147483647, so replicas on older releases also stay safe during a rolling upgrade: out-of-range values are rejected on CREATE and clamped, with a warning, when replayed from existing metadata, server config, or Keeper, and existing databases that hold a larger value in Keeper are fixed automatically when a replica on the new release starts. #118103 (Kirill Shokhin). #123105 (Kirill Shokhin).
• ALTER DATABASE ... MODIFY COMMENT on a database with ENGINE = Backup no longer writes a definition the server cannot read back, which made it refuse to start; metadata already written in that form now loads too. Backward incompatibility: a backup taken by an older server of a database with ENGINE = Backup stores the locator quoted into a string literal, and such a definition is no longer accepted from a RESTORE. #118674 (Alexey Milovidov).
• Tables in a database with lazy_load_tables were missing from system.parts, system.replicas and related system tables, and refused by replica commands, after a restart. #121332 (Nikolay Degterinsky).
• Fixed the server failing to start with TABLE_ALREADY_EXISTS or Mapping for table with UUID=... already exists after it was killed during CREATE OR REPLACE TABLE, REPLACE TABLE, or CREATE TABLE ... AS SELECT in an Atomic database whose metadata is on a plain_rewritable disk (database_disk). #123902 (Alexey Milovidov). #123832 (Groene AI).
• Fixed EXISTS TABLE returning 0 for a table that is still being loaded asynchronously after server startup. #125729 (Alexey Milovidov).
• Fixed SELECT ... FINAL returning a superseded row when the partition expression distinguishes floating-point values that the FINAL merge comparator treats as equal, such as -0.0 versus 0.0 or different NaN payloads. enable_automatic_decision_for_merging_across_partitions_for_final no longer engages for such a key. #117482 (Alexey Milovidov).
• Row policies over floating-point sorting key columns are now applied after FINAL, with apply_row_policy_after_final enabled, and such columns are excluded from index analysis of deferred filters. Previously a policy condition distinguishing -0.0 from 0.0 or NaN payloads could filter out the winning row version before the merge and return stale data. #121782 (Yarik Briukhovetskyi).
• CREATE TABLE and ALTER TABLE ... ADD INDEX now reject a skip index whose expression cannot be evaluated, such as INDEX i c0 = c1 TYPE set(0) over String and Int8, and an ALTER TABLE ... MODIFY COLUMN that would leave an existing index in that state is rejected too. Such a declaration used to be accepted, after which every INSERT and merge failed with NO_COMMON_TYPE, and MATERIALIZE INDEX retried that error forever. Closes #69597. #118373 (Groene AI).
• Fixed SELECT count() returning more rows than the filter accepts when a MergeTree primary key is a Tuple holding a floating point element and the key data contains a nan. For example, with ORDER BY t over t Tuple(Float64, Int32), SELECT count() FROM t WHERE t >= (5., 3) returned 6 while SELECT * FROM t WHERE t >= (5., 3) returned 2 rows. A nan nested in a tuple is not visible in the granule bounds, so primary key analysis reported a granule of non-matching rows as wholly matching and the exact-count optimization counted its rows without evaluating the filter. Closes #117819. #118683 (Groene AI).
• Fixed WHERE f IN (nan) silently losing the NaN rows of a floating-point column: statistics-based part pruning and the minmax skip index use ranges that exclude NaN, so a part or granule holding NaN was pruned even though IN matches NaN. #117453 (Alexey Milovidov).
• Fixed wrong results when a String primary key column is compared against a FixedString constant wrapped in Variant or Dynamic. #114353 (Groene AI).
• Fixed wrong results for f = 0.0 and f != 0.0 over a float column when the primary key is an injective transform of it (for example ORDER BY toString(f)): the index treated the transformed constant as exact, although the transform distinguishes -0.0 from +0.0 while the comparison does not. #117722 (Alexey Milovidov).
• Comparing an Enum column with a string literal that is not one of its elements no longer throws UNKNOWN_ELEMENT_OF_ENUM when validate_enum_literals_in_operators is disabled (the default) and the column is part of the primary key, the partition key, or a bloom_filter index. != with such a literal now returns true for every row instead of throwing. Closes #122235, #111545. #122672 (Vladimir Cherkasov).
• Fixed an exception (such as Division by zero) thrown from index analysis when a monotonic function chain over a key column cannot be evaluated on a key range endpoint, for example SELECT count() FROM t WHERE p != 1 AND intDiv(1, p - 1) > 0 with use_skip_indexes = 0. Also fixed a possible segmentation fault caused by a poisoned entry in the index analysis function result cache. #117355 (Alexey Milovidov).
• Fix a bloom_filter skip index on mapValues or mapKeys that dropped rows for m['k'] IN (...) when the map values are an Enum with negative values and the set contains the name of the value 0. #124232 (Alexey Milovidov).
• Fix a wrong order (and wrong rows under LIMIT) from the read-in-order optimization when the ORDER BY expression maps a stored inf to NaN, as in ORDER BY x / inf. #118720 (Alexey Milovidov).
• Fixed primary key, partition key, and skip index analysis that could skip or over-count rows, and ORDER BY optimization that could return rows in the wrong order, for conditions on timezoneOffset, toTimeWithFixedDate, toYYYYMMDDhhmmss, toHour, and toMinute over a DateTime or DateTime64 column when the key range spans a change of the UTC offset, for example a daylight saving time change in America/New_York or the 30-minute change in Australia/Lord_Howe. #124474 (Groene AI). #124530 (Groene AI).
• Fixed wrong results of ORDER BY with optimize_read_in_order, GROUP BY with optimize_aggregation_in_order, and LIMIT BY with the in-order optimizations when a String sorting-key column is compared with a FixedString constant, also one held by a Variant or Dynamic constant (for example WHERE s = toFixedString('a', 2) or WHERE s = CAST(toFixedString('a', 2) AS Dynamic)), or when a Variant or Dynamic expression, such as a JSON path, is compared with a constant (for example WHERE j.a = 1 over Int64 and Float64 values). Such a comparison matches several values that sort differently, but the optimization treated them as one value and skipped sorting by them. #123677 (Ashish Kumar Choubey). #125378 (Samuel Krempaský).
• Fixed count() counting rows that the filter rejects when a whole-tuple comparison against a Tuple primary key involves a NULL or a NaN nested in the constant, or a NaN nested in the key data. #118944 (Alexey Milovidov).
• Fixed wrong result order (and wrong rows under LIMIT) from the read-in-order optimization when ORDER BY is a monotonically decreasing function of a Float sorting key containing NaN, e.g. ORDER BY negate(x). The opposite request for the same shape, ORDER BY negate(x) ASC NULLS FIRST, is now read in order instead of being sorted. #118469 (Alexey Milovidov).
• Fixed repeated reading of text index posting lists in multi-threaded queries when the global text index caches are disabled. #120844 (Anton Popov).
• The text, tokenbf_v1, ngrambf_v1, and sparse_grams skip indexes are now used for IN when transform_null_in = 1 and the set contains no NULL. Previously such a query read every granule, and one using force_data_skipping_indices failed with INDEX_NOT_USED. #121375 (Jimmy Aguilar Mena).
• Fixed wrong results and the exception Column ... already added for reading for a FINAL query with query_plan_optimize_lazy_final = 1 that reads a text index directly, when the ReplacingMergeTree sorting key is Nullable, floating-point, or mixes ASC and DESC columns. #122534 (Groene AI).
• On a column whose text index has a postprocessor, hasPhrase rejoined the index tokens with a space before matching, so it returned wrong results unless that space matched how the tokenizer splits. Allowing hasPhrase to accept an Array(String) removes joining tokens completely. #120859 (Elmi Ahmadov).
• Fixed a text index with tokenizer = 'japanese' returning fewer rows than a full scan: over a LowCardinality column every second row of a run of equal values was not indexed, and two search terms of the same byte length in one query (for example hasAnyTokens(s, 'エンジン', 'japanese') AND hasAnyTokens(s, '形態解析', 'japanese')) matched no rows. Indexes built before this fix keep the missing rows until ALTER TABLE ... MATERIALIZE INDEX is run. #123389 (Groene AI).
• Fixed a query failing with an exception like Cannot parse UInt64 from String on a table with a text index on mapKeys or JSONAllPaths, when the predicate cannot be evaluated on the value of a missing map key or JSON path. #124612 (Anton Popov).
• Fixed count() with a text index search function such as hasToken, hasAnyTokens, or hasAllTokens in WHERE and a constant GROUP BY key returning a row with 0 instead of no rows when nothing matches. #124593 (Groene AI).
• Fix hasAnyTokens with a String needle on a text index with the sparseGrams tokenizer returning fewer rows than the equivalent Array needle. Grams covered by a longer gram of the needle were dropped, which is only valid for hasAllTokens. #124613 (Jimmy Aguilar Mena).
• Fixed slow planning of queries with a text index on tables with many parts when other skip indices are stored in the packed skp_idx.packed archive. Checking for an absent optional text-index file no longer calls stat on disk for every part. #125381 (Anton Popov).
• Fix ALTER TABLE ... MATERIALIZE STATISTICS / DROP STATISTICS failing forever and wedging the table’s mutation queue when a server profile sets allow_statistics = 0 and the statement enables it per-session, the same opt-in ADD STATISTICS honours. Also close a hole where MATERIALIZE STATISTICS ignored allow_statistics = 0 altogether under validate_mutation_query = 0. #117351 (Alexey Milovidov).
• Fixed a mutation silently dropping one of two swapped columns’ statistics when the part carried a composed RENAME COLUMN pair that exchanged their names. #118977 (Alexey Milovidov).
• When the query cache is unable to store results (with enable_writes_to_query_cache = 0, in clickhouse-local, or with a zero limit in the query_cache server configuration), queries with use_query_cache = 1 no longer fail with QUERY_CACHE_USED_WITH_NONDETERMINISTIC_FUNCTIONS, QUERY_CACHE_USED_WITH_SYSTEM_TABLE or QUERY_CACHE_USED_WITH_NON_THROW_OVERFLOW_MODE: these checks protect against caching wrong results, which cannot happen then. When a query_cache limit is reconfigured to 0, the existing entries are now dropped instead of lingering. #123604 (Konstantin Bogdanov).
• LIMIT [n] AFTER start_expr UNTIL end_expr now checks end_expr only after start_expr first matches. Previously a match of end_expr anywhere before the first start_expr incorrectly made the result empty. #120514 (Nihal Z. Miaji).
• Fixed SELECT DISTINCT and LIMIT BY over a floating point column returning a different number of rows depending on whether the in-order variant of the operator was used: -0.0 and 0.0, and distinct NaN payloads, are no longer merged by the in-order variant. #118198 (Alexey Milovidov).
• Fixed a higher-order function with a non-deterministic lambda body, such as arrayMap(i -> rand64(i), ...), being evaluated once per source row instead of once per output row after ARRAY JOIN when lazy columns replication is enabled. #120796 (Yarik Briukhovetskyi).
• Fixed a performance regression where a query whose WHERE requires the same expression to equal two different constants stopped being recognized as always false when compatibility was set below 26.7, or when optimize_redundant_comparisons was disabled explicitly. Closes #121007. Closes #121012. #121065 (Groene AI).
• Fixed NOT_AN_AGGREGATE being thrown for a GROUP BY key referenced from a window function when group_by_use_nulls is enabled together with ROLLUP, CUBE, or GROUPING SETS, as in SELECT rank() OVER (PARTITION BY k) FROM t GROUP BY k WITH ROLLUP. #118173 (Groene AI).
• Fixed AMBIGUOUS_IDENTIFIER for a column qualified with the alias of a table expression when another table of the same query is named like that alias. Such a qualifier now refers to the alias, as it does in the old analyzer, both in a join that reported the ambiguity and in a two-table join that resolved it to the other side. #119082 (Alexey Milovidov).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK when a query tree pass rewrites a window function’s aggregate in place. A query whose only reference to a Nullable column was a windowed count, such as SELECT count(n) OVER () FROM t, was rejected at default settings because optimize_functions_to_subcolumns rewrote it into sum(not(n.null)) and lost the node’s window kind. optimize_uniq_to_count corrupted uniq(n) OVER () over a SELECT DISTINCT subquery the same way. Closes #119635. #119681 (Groene AI).
• Fix a wrong result from optimize_injective_functions_in_group_by on the two-argument form toString(x, timezone). GROUP BY toString(number, NULL) could be rewritten to GROUP BY number and produce one group per row instead of one. #117197 (Ilya Yatsishin).
• Fixed a server crash in CREATE TABLE ... AS <table function> when an argument of the table function is a scalar subquery with a JOIN and enable_join_runtime_filters_index_analysis is enabled in the default settings profile. #122795 (Groene AI).
• Fixed table names not being qualified with the database when a WITH name declared elsewhere in the same query happened to match them: a view whose UNION branch referenced a table with the name of a WITH RECURSIVE CTE of another branch could not be read from another database, and ALTER TABLE ... UPDATE with a common table expression referenced from a subquery failed with UNKNOWN_DATABASE or read a table of the session database. #114908 (Groene AI).
• Fixed IN against a table-backed set (ENGINE = Set), or with transform_null_in = 1, throwing CANNOT_CONVERT_TYPE for probe values outside the set key type’s range instead of evaluating them to false, which also made the query succeed or fail depending on how much the plan pruned. #117606 (Alexey Milovidov).
• Fixed a crash and wrong results with convert_query_to_cnf = 1 when WHERE or HAVING references the same SELECT alias more than once. #123414 (Groene AI).
• Fixed wrong results, Unexpected return type from tupleElement errors and NOT_FOUND_COLUMN_IN_BLOCK errors when a Nested column is read through ARRAY JOIN with a numeric tupleElement index (for example tupleElement(n, 3)) while another subcolumn is unused, and the same expression is reused by ORDER BY ALL, GROUP BY ALL or an alias. #123647 (Groene AI).
• Fix UNKNOWN_IDENTIFIER in the analyzer when a table is referenced as database.table in FROM or JOIN and another table expression in the same query uses that database name as its alias, for example FROM db.t1 AS db INNER JOIN db.t2 AS t ON .... Closes #122239. #124262 (Dmitry Novik).
• Fix a wrong result, a leaked totals row, and a LOGICAL_ERROR exception with rewrite_in_to_join when the subquery of IN uses WITH TOTALS. #124096 (Alexey Milovidov).
• Fix Unexpected return type from __topKFilter logical error raised by ORDER BY ... LIMIT over a MergeTree table with more than one part when the sort column contains a Nullable element, for example Tuple(UInt64, Nullable(UInt64)). use_top_k_dynamic_filtering is enabled by default, so no opt-in setting was required to hit this. #113406 (Groene AI).
• Reject a duplicate column alias name in a CTE or table expression, which previously could return an incorrect result instead of failing. Closes #114427. #114505 (Yarik Briukhovetskyi).
• Fixed APPLY lambda(tuple(x), ...) followed by an operator, as in SELECT * APPLY lambda(tuple(x), toString(x)) > '0', which failed with UNKNOWN_IDENTIFIER. A view using APPLY (lambda(...)) followed by an operator also stopped working after a server restart. #124142 (Groene AI).
• Fix a logical error Equal values are not contiguous within the range assumed to be sorted (and potential wrong DISTINCT results) when optimize_distinct_in_order is enabled and an ORDER BY over an arrayJoin-derived column is followed by another ARRAY JOIN that reproduces a column of the same name. #110778 (Groene AI).
• Fixed wrong results and a LOGICAL_ERROR (Cannot capture column ... because it has incompatible type) when a matcher such as b.*, *, COLUMNS('...'), or b.* EXCEPT c appears inside a lambda body. A matcher expanded inside a lambda kept its pre-JOIN column types, so join_use_nulls = 1 did not turn columns from the NULL side of a LEFT/FULL/RIGHT JOIN into Nullable (unmatched rows produced a value computed over defaults instead of NULL), and group_by_use_nulls = 1 with WITH ROLLUP / WITH CUBE / GROUP BY GROUPING SETS failed with NOT_AN_AGGREGATE. A matcher expanded inside a lambda now carries the same types it would have if the columns were written out explicitly, and a lambda inside PREWHERE is no longer promoted to Nullable by join_use_nulls. Also fixed wrong results for a matcher inside a lambda when analyzer_compatibility_join_using_top_level_identifier = 1 makes a JOIN ... USING key resolve to a SELECT projection alias: the matcher selected the merged USING key instead of the right table’s column. #111993 (Groene AI).
• Fixed wrong results and NOT_FOUND_COLUMN_IN_BLOCK for three-argument multiIf and if-chain GROUP BY keys with ROLLUP, CUBE, or GROUPING SETS and group_by_use_nulls = 1. Also fixed constant-condition keys reused through aliases with enable_identifier_resolve_cache, lost grouping sets when a key collapses to another key, and a Bad cast exception for a correlated subquery reading that column. Closes #122404. Part of #103419. #123065 (Groene AI). #122512 (Groene AI).
• Fixed wrong results of ORDER BY, LIMIT BY, and GROUP BY with optimize_aggregation_in_order over a function of the sorting key with a NULL constant argument, such as toString(dt, NULL) or toString(dt, {tz:Nullable(String)}) with a NULL parameter. Such a function returns NULL for every row, but reading in order treated it as strictly monotonic and kept the next sorting key column in table order. #125042 (Groene AI).
• Fixed a server crash, wrong results, or a rejected query in ORDER BY ... WITH FILL ... INTERPOLATE when one INTERPOLATE expression has to be converted to its target type and another item reads the same column or literal, for example INTERPOLATE (a AS b, b AS a) with targets of different types. #125129 (Groene AI).
• A WITH list whose elements refer to the preceding ones, or a chain of common table expressions that read the previous one more than once, now fails with TOO_BIG_AST past max_expanded_ast_elements instead of exhausting the server’s memory in EXPLAIN AST optimize = 1, CREATE VIEW, a full ATTACH VIEW ... AS SELECT and ALTER TABLE ... MODIFY QUERY. #122362 (Alexey Milovidov).
• Fix excessive memory usage for queries with ARRAY JOIN and filters on both rows and array elements: the row filter is now applied before expanding the array. #125577 (Konstantin Vedernikov).
• Fixed ORDER BY ... LIMIT over a column of an Interval type (or Nullable, LowCardinality, Tuple or Array of one) failing with TYPE_MISMATCH when use_top_k_dynamic_filtering is enabled, which is the default. Closes #125197. #125221 (Groene AI).
• Fixed a heap-use-after-free when a correlated subquery reading a Merge table was planned inside EXPLAIN, a subquery in FROM, a view, or an IN subquery. #122475 (Groene AI).
• Fix a correlated scalar subquery used as the whole WHERE, HAVING, or QUALIFY condition, e.g. WHERE (SELECT count() > 0 FROM t AS i WHERE i.k = o.k), being rejected with ILLEGAL_AGGREGATION or NOT_AN_AGGREGATE, or failing with NOT_IMPLEMENTED (NOT_FOUND_COLUMN_IN_BLOCK on older versions) when the outer column it uses is not read elsewhere in the query. Closes #124499. #124594 (Groene AI).
• Fixed NOT_IMPLEMENTED (None of the algorithms enabled by the 'join_algorithm' setting ... can execute this SEMI/ANTI JOIN) for a correlated EXISTS/NOT EXISTS subquery, for an IN (subquery) under rewrite_in_to_join, or for an ANY LEFT JOIN of an ENGINE = Join table under an always-false WHERE, when join_algorithm is set to full_sorting_merge or partial_merge. #118315 (Alexey Milovidov).
• Comparing two deeply nested Array, Tuple, Map, or Object values doubled the work for every nesting level, because the ordering asked each equal element both whether it was less and whether it was greater. A single-value aggregate such as min, max, argMin, groupArraySorted, or an -ArgMin combinator over a 30-level nested argument therefore ran for minutes on a few hundred rows, and could not be interrupted by max_execution_time. Such a comparison now visits each element once. #116455 (Groene AI).
• Fixed a server double free when finalizing an aggregate function with the -State combinator nested under a combinator that transfers several substates in one call, such as -Resample, -ForEach, -Map, or -Tuple, if an exception is thrown part-way through the transfer. #111287 (Groene AI).
• Fixed wrong results from quantileExactWeightedInterpolated, quantileInterpolatedWeighted, and their quantiles* forms over integer-backed types. #114571 (Groene AI).
• Fixed topK, topKWeighted, approx_top_k, approx_top_count, approx_top_sum, uniqUpTo, uniqCombined, uniqCombined64, exponentialMovingAverage, stochasticLinearRegression, stochasticLogisticRegression, and groupNumericIndexedVector state columns with a Decimal or wide-integer parameter. The parameter was persisted as a quoted string that reparsed as String, so after a restart the table was unreadable, uninsertable, and undroppable. #121491 (Groene AI).
• Fixed a server abort when the flameGraph aggregate function was constructed on a thread without a query context, for example while attaching a stored table whose column type names AggregateFunction(flameGraph, ...). #121894 (Groene AI).
• Fixed runningAccumulate returning an accumulation computed over only part of the rows when it reads a state column under if, multiIf, and, or or at the default short_circuit_function_evaluation = 'enable'. Inside a boolean predicate this also changed which rows passed the filter. An argument computed inline by another lazily evaluated function, whether it is the state or the grouping key, is still filtered. Closes #117121. #117126 (Groene AI).
• Fixed arrayIntersect returning a NULL that is present only in its first argument, which made the result depend on the order of the arguments. NULL is now in the result only when every argument array contains one. #118433 (Alexey Milovidov).
• arrayUniq, arrayEnumerateUniq, arrayEnumerateDense, and arrayJaccardIndex now distinguish NULL from the nested type’s default value for nullable arrays that use generic hashing. #120676 (xiaohuanlin).
• Fixed an out-of-bounds heap write in stem. The output buffer now grows when the stemmed word is longer than the estimate from the input. #121402 (Jimmy Aguilar Mena).
• Fix array and map search functions returning a match for a NULL String needle when an empty string is present. #118741 (Minh Vu).
• Fixed least and greatest returning wrapped-around values for a LowCardinality numeric argument of mixed signedness, intDiv/%/positiveModulo throwing Division by zero on NULL rows of a LowCardinality(Nullable(...)) denominator, and fromModifiedJulianDay, runningConcurrency, hasPhrase, formatRow, now, and CAST rejecting valid LowCardinality arguments. #117450 (Alexey Milovidov). #122418 (Groene AI).
• Fix has, notHas, indexOf, countEqual, and mapContains silently finding nothing in a constant array or map with Enum keys searched for the name of an enum value, e.g. mapContains(map('a', 1)::Map(Enum8('a' = 1), UInt8), 'a') returned 0. They now compare by the name of the enum value, consistently with equals and with the same functions over a non-constant argument. Also fix indexing a Map with Enum keys by a numeric value that is not in the Enum throwing UNKNOWN_ELEMENT_OF_ENUM during query analysis instead of returning the default value, and an exception in the analysis of a bloom_filter index over mapKeys of a Map with Enum keys. Also fix a bloom_filter index over an Array(String) or Array(FixedString) column wrongly pruning the granules that match hasAny or hasAll with a constant Array(Enum) argument. #116093 (Alexey Milovidov).
• Fixed comparison of tuples with an untyped NULL element, such as (1, 2) != (3, NULL) or (1, 2) != toNullable((3, NULL)): the result was NULL (or 0 for a Nullable tuple) even when another element decides it. It is now the same as with a typed NULL: 0 or 1 when another element decides the comparison, NULL otherwise. A filter and its negation no longer both drop a row, e.g. with WHERE NOT (_partition_value > toNullable((..., NULL))). #123896 (Groene AI).
• Fixed pointInPolygon with a constant polygon returning wrong results for some points of ordinary polygons, for example 0 for the point (0.76, 0.01) inside the polygon [(1, 0), (0, 1), (-1, 0), (0, -1)], or 1 for a point inside a small hole. The results now match those for a non-constant polygon. #123767 (Groene AI).
• Arithmetic between a Decimal and an integer that does not fit the native width of the decimal result now raises DECIMAL_OVERFLOW instead of silently wrapping the integer operand. #118721 (Alexey Milovidov).
• Fixed wrong results of pointInPolygon for polygons with huge coordinates. With a constant array polygon, a filter on a table whose primary key covers the point columns could also return fewer rows than a full scan. pointInPolygon now throws BAD_ARGUMENTS where it would compute with a polygon coordinate larger than 1e150 in absolute value. #123740 (Groene AI).
• Fixed date_time_overflow_behavior being ignored while parsing Date and DateTime. Closes #113151. #116099 (Yarik Briukhovetskyi).
• toStartOfInterval no longer widens Date and DateTime inputs to Date32 and DateTime64 when enable_extended_results_for_datetime_functions = 1; only Date32 and DateTime64 inputs keep the extended result types, matching the toStartOfX functions. This also removes the saturated result (2106-02-07 06:28:15) that toStartOfInterval(<Date>, INTERVAL ... DAY) returned under that setting for dates after 2106-02-07. #119806 (Groene AI). #121809 (Kirill Kopnev).
• Fixed age('week', ...) sometimes being off by one week when the dates fall on different weekdays. #122633 (Samuel Krempaský).
• Fixed wrong results when an Enum8 or Enum16 value is converted to DateTime or Date. A negative member wrapped instead of saturating like the Int8 or Int16 it is stored as, so toDateTime of an enum returned a far-future timestamp and primary-key analysis discarded granules that did contain matching rows. #118039 (Groene AI).
• Fixed toDateTime64, toDateTime with a scale and toTime64 throwing CANNOT_PARSE_DATETIME for a Nullable(String) or Nullable(FixedString) argument that contains NULL; such rows now return NULL. For toDateTime64 this was a regression in 26.5, where cast_string_to_date_time_mode became best_effort by default. #123311 (Groene AI).
• Fixed dateDiff and age with unit day returning a result off by 65536 (for example -65535 instead of 1) when an argument is a DateTime on the local day 1969-12-31 in a time zone west of UTC, such as toDateTime(0) in America/New_York. If a table stores such a value in a sorting key, partition key, skip index, projection or MATERIALIZED/DEFAULT column, rebuild it. #124555 (Groene AI).
• Setting date_time_overflow_behavior is now honored by CAST and the conversion functions for numeric conversions to Date, Date32, DateTime, Time, DateTime64, and Time64. For DateTime64 and Time64, the same now applies to the INSERT ... VALUES expression fallback and the values table function, and for DateTime64 sort keys to numeric WITH FILL bounds. Previously an out-of-range value was silently clamped even in the throw mode. accurateCast, accurateCastOrNull, and accurateCastOrDefault no longer clamp an unrepresentable value into DateTime64 or Time64, and saturation now reaches the extreme representable tick of the target rather than the last whole second. #117667 (Alexey Milovidov).
• Fixed a constant of Variant type reached through an Array, Tuple, or Map being mis-serialized to a secondary server, so a query that returned a result on a single node failed with CANNOT_CONVERT_TYPE through parallel replicas, remote, or Distributed. A String member of a multi-member Variant constant, scalar or nested, arrived there as whichever member its text parsed as, and an empty Map whose key or value is a Variant was rejected outright. #117831 (Groene AI).
• Fixed LOGICAL_ERROR when reading a sizeN subcolumn of a shared-data JSON path, for example a size1 subcolumn over data of type Array(Array(Nullable(Int64))). It affected parts written with the advanced shared data serialization version, which is the default for merged parts. #120039 (Pavel Kruglov).
• Fixed wrong results when reading the null subcolumn of a Nullable field of a Variant, Dynamic, or JSON element, for example value.`Tuple(a Nullable(UInt32), b String)`.a.null. On rows holding a different alternative the subcolumn returned 0 although the field itself was NULL, so optimize_functions_to_subcolumns (enabled by default) rewriting IS NULL to that subcolumn silently dropped matching rows. Closes #119935. #120099 (Groene AI).
• Fix JSONHas and JSONExtractRaw returning an absent result for a typed JSON path whose literal value is NULL but which contains a non-empty sub-object when type_json_skip_null_typed_paths is enabled. #120321 (JasonLi-cn).
• Fixed max_memory_usage and max_server_memory_usage not being respected when reading a JSON column with the FLATTENED serialization version from the Native format. #117730 (Pavel Kruglov).
• Fixed reading flattened Nested members next to a JSON member in Wide parts. #122519 (Mohammed Sohail).
• Allow the JSON type and the JSON* functions to parse documents whose String values hold bytes that are not valid UTF-8, which previously failed with INCORRECT_DATA. As a side effect, isValidJSON and the JSONExtract* family now accept such documents instead of reporting them invalid. #119122 (Groene AI).
• Fix stack exhaustion when reading a deeply nested Dynamic or JSON value as a Field or printing it in a text format. Both paths now raise TOO_DEEP_RECURSION instead of terminating the server. #121441 (Groene AI).
• Fix reading bucketed runtime Map values from JSON columns using advanced shared-data serialization. #118577 (Rory Shanks).
• Fixed a Dynamic constant losing the type of its value when a query is sent to other servers (Distributed, remote, parallel replicas): a constant holding Enum8('7' = 3) arrived as the number 3, so IN and NOT IN matched the wrong rows and = failed with NO_COMMON_TYPE, and a UInt64, IPv4, or Map value arrived as a different type. Closes #119745. #118988 (Groene AI).
• Fixed a crash on a query with a deeply nested Dynamic or JSON constant. #120885 (Pavel Kruglov).
• Fix inconsistent results and missing type errors when optimize_functions_to_subcolumns rewrites a JSON path accessed with tupleElement or arrayElement using an unsupported key type. Such expressions now behave the same with and without the optimization. #125471 (Groene AI).
• Fixed TYPE_MISMATCH when a LowCardinality join key is also selected from the NULL-extended side of a LEFT, FULL, or ASOF LEFT join with join_use_nulls = 1 and join_algorithm = 'full_sorting_merge' or 'parallel_full_sorting_merge'. Closes #120424. #120469 (Groene AI).
• Fixed ILLEGAL_PREWHERE on sharded GLOBAL RIGHT JOIN queries that used PREWHERE on the left table. The filter is applied to the broadcast side before the join instead of being sent to a Memory temporary table that does not support PREWHERE. PREWHERE on a remote table in a join is also preserved when that table is read through a subquery wrap. #119780 (Anton Ivashkin).
• Fixed a Pipeline stuck exception when a query combined a merge join using join_algorithm = 'parallel_full_sorting_merge' with a step that glues independent streams together, such as reading from a Merge table, or a UNION ALL narrowed by max_streams_for_union_step. Such a query now runs instead of failing. #121227 (Groene AI).
• Fixed a query failing with an error such as CANNOT_READ_ARRAY_FROM_TEXT when a subquery’s SELECT list computes an expression that the same subquery’s WHERE guards against, and that subquery is the probe side of a JOIN for which a runtime join filter is built (enable_join_runtime_filters = 1, enabled by default). The guarded expression was evaluated on rows that the WHERE had already removed. #122143 (Groene AI).
• Fixed lost matches on a Dynamic join key when outer joins are reordered with a join order conflict detector enabled (query_plan_optimize_join_order_conflict_detector) and join_use_nulls = 1. #123113 (Fisnik Kastrati).
• Fix NOT_IMPLEMENTED error for a LEFT JOIN over a dictionary or other key-value storage under join_algorithm = 'direct'. #123491 (Hechem Selmi).
• Fix exponential planning time in the transitive outer-to-inner JOIN conversion when the expression DAG has shared nodes, such as a column used twice (x + x) across nested subqueries. The analysis now visits each node once instead of every path. #123509 (Hechem Selmi).
• Fix a performance regression where the join runtime filter was skipped for joins that combine an = key with a null-safe <=> (IS NOT DISTINCT FROM) key once hash table statistics were collected. #122618 (Manuel).
• Fixed indexHint over the non-preserved side of an OUTER join silently returning extra rows, and join_algorithm = 'direct' failing with NOT_IMPLEMENTED when a column-free conjunct such as materialize(1) is added to the WHERE clause. #117569 (Alexey Milovidov).
• Fix NOT_IMPLEMENTED for an outer JOIN with join_algorithm = 'full_sorting_merge' when optimize_and_compare_chain derived a condition on the preserved side of the join from the ON clause. #124233 (Yarik Briukhovetskyi).
• Fix UNKNOWN_IDENTIFIER errors in INSERT ... SELECT from a parameterized view when parallel_distributed_insert_select is enabled. The view is now analyzed with the analyzer instead of the legacy interpreter. #119495 (Yarik Briukhovetskyi).
• Fixed a query over a Distributed table (or cluster/remote/clusterAllReplicas with an inner view) spanning more than one shard that references no column of the table anywhere in the query, such as SELECT count(*) OVER () FROM distributed_table. Such a query silently returned no rows, and returned LOGICAL_ERROR Invalid number of columns in chunk pushed to OutputPort when the table produced a constant. The fix covers the analyzer; enable_analyzer = 0 is fixed separately by #113955. #113469 (Groene AI).
• Fixed a server crash when a scalar or IN subquery inside a standalone expression, such as a CHECK constraint, a TTL expression, or a row policy, reads a Distributed table or the remote table function. #121418 (Groene AI). #121639 (Alexey Milovidov).
• Fix NOT_FOUND_COLUMN_IN_BLOCK for grouping over a Distributed table with optimize_skip_unused_shards = 1: a constant that a query tree pass builds is now named on a shard exactly as the initiator names it. #119273 (Alexey Milovidov).
• Refuse a query with ORDER BY over a Distributed table whose declared column type differs from the shard table’s type in a way that does not preserve the order, for example Int8 declared over a String column. The shards sorted by their own type and the initiator merged the streams as if they were sorted by the declared type, which returned rows in the wrong order, cut the wrong rows with LIMIT, and duplicated values under DISTINCT. The check runs when a shard is the local server, where the shard table’s types are known, and only for the columns the query sorts by. Exact widenings (Float32 to Float64, Date to Date32, DateTime to DateTime64, a Decimal with no smaller scale and no fewer integer digits, FixedString to String) are now recognized as order-preserving by both this check and the one of the Merge engine. #118445 (Alexey Milovidov).
• Fixed queries over a Distributed table when optimize_skip_unused_shards skips every shard and the query uses distributed_group_by_no_merge or reads the table through another Distributed table. Such queries failed with NUMBER_OF_COLUMNS_DOESNT_MATCH, CANNOT_CONVERT_TYPE, or std::out_of_range (with ALIAS columns) instead of returning an empty result or the aggregate of an empty set. #122982 (Groene AI).
• Fixed queries that go through another server (remote, Distributed, parallel replicas, GLOBAL IN, and QueryRunner jobs sent to a cluster) failing with SYNTAX_ERROR, UNKNOWN_CODEC, or CORRUPTED_DATA when output_format_native_encode_types_in_binary_format or input_format_native_decode_types_in_binary_format is enabled. The blocks of these queries now carry type names on the wire, and a shard with this change also works with an initiator without it. clickhouse-client --query_kind secondary_query follows the same rule. #123649 (Groene AI). #124716 (Groene AI).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK (or THERE_IS_NO_COLUMN) when a query sent to other servers (a Distributed table, remote, parallel replicas) has an IN over a parameterized view called without its database name inside an expression the servers return, for example countIf(x IN (SELECT y FROM pv(p = 1))), a GROUP BY key, or a SELECT item. Also fixed NOT_AN_AGGREGATE when one query spells the same parameterized view call both with and without its database name. #123662 (Groene AI).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK when reading a subcolumn (Array .size0, a Tuple element, Map .keys/.values, Nullable .null, a JSON path, or a Dynamic typed subcolumn) through a VIEW over a Distributed table. #119293 (Alexey Milovidov).
• Fixed wrong results of aggregations, LIMIT, DISTINCT, LIMIT BY and window functions over a Buffer table whose destination is a Distributed table that a single shard finishes (one shard, distributed_group_by_no_merge, or a GROUP BY by the sharding key): the rows still in the buffer and the rows of the destination were processed separately and both results were returned. #124981 (Groene AI).
• Fix wrong results when a condition outside a distributed subquery with QUALIFY was pushed into it and evaluated before the window functions QUALIFY filters by. #125491 (Nikita Taranov).
• Fixed THERE_IS_NO_COLUMN and NOT_FOUND_COLUMN_IN_BLOCK errors in queries over remote or with parallel replicas that compare tuples where one pair of elements is constant and false, such as (n, 1) = (n, 2), or where NULL is compared with a column, such as (n, s) = (NULL, s). #125195 (Groene AI).
• Fixed wrong results with parallel_replicas_plan_based = 1 when the plan fragment sent to the replicas has no output columns, for example SELECT count() FROM (SELECT 1 FROM t LIMIT 500). Such a query returned 0, or an empty result set, instead of the correct value. #115871 (Groene AI).
• Fixed wrong results and a LOGICAL_ERROR when a query reads with parallel replicas inside a distributed sub-query whose cluster_for_parallel_replicas numbers its shards differently from the cluster the query is distributed over: a VIEW carrying its own SETTINGS cluster_for_parallel_replicas, a read through clusterAllReplicas of that same cluster, a cluster built by cluster discovery or by a Replicated database when the two servers see a different set of shards under one name, or a plain remote_servers cluster whose per-server copies number the shards differently while a configuration change rolls out. The shard number shipped by the initiator was applied to the other numbering and selected the wrong shard. Parallel replicas are now declined for such a read. #113589 (Groene AI). #124724 (Igor Nikonov).
• Fixed SELECT count() returning a value multiplied by the number of replicas when executed inside a transaction with parallel replicas enabled. Trivial count optimization now is not rejected inside a transaction, so it is executed solely on the initiator. Followers no longer apply the trivial count optimization during query planning, which previously caused each of them to independently answer with the total row count read from the table’s metadata whenever the query reached them. #116365 (Kirill Shokhin).
• Fixed max_bytes_to_read, max_bytes_to_read_leaf, and read-speed and timeout limits being ignored for reads executed from a query plan shipped to another node, which happens with parallel_replicas_plan_based and make_distributed_plan. #119958 (Igor Nikonov).
• Fixed LOGICAL_ERROR Cannot find a matching set in the map of sets from single-replica plan, and silently lost parallel replicas, when automatic parallel replicas were enabled and an IN subquery carried its own SETTINGS clause naming force_primary_key or automatic_parallel_replicas_mode. #120799 (Groene AI).
• Fix UNKNOWN_FUNCTION on a shard or a replica for a higher-order function whose lambda body is an AND of comparisons when serialize_query_plan = 1 or parallel_replicas_plan_based = 1: a JIT-compiled lambda body was serialized into the plan. #119332 (Alexey Milovidov).
• Fix missing total rows in progress (and therefore a non-working progress bar) for queries with parallel replicas when parallel_replicas_local_plan is enabled (the default). #123928 (Alexey Milovidov).
• Fixed duplicate filter push-down in plan-based parallel replicas. #117731 (Nikita Taranov).
• Fixed wrong results and UNKNOWN_IDENTIFIER errors when a query reads a MergeTree table after reading one of its projections with the mergeTreeProjection table function, for example in a JOIN or UNION ALL of the two. With parallel replicas such a query could also fail with Got read request from replica N for unknown stream. Closes #102182. #123915 (Groene AI).
• FINAL on table functions now disables parallel replicas the same way as FINAL on tables, so a TimeSeries table function such as timeSeriesTags no longer returns duplicate rows under parallel replicas. #122492 (Groene AI).
• Fixed parallel replicas not being used when reading a Distributed table whose underlying table is another Distributed table and every shard of the outer cluster has a single replica. #123566 (Igor Nikonov).
• Fixed TOO_MANY_ROWS being raised for a query that stays within its max_rows_to_read budget when parallel_replicas_min_number_of_rows_per_replica is set. The estimate sizing the replica set enforced the limit against whole granules at planning time. #121269 (Groene AI).
• Fix a NOT_IMPLEMENTED exception for JOIN queries with IN subqueries when using plan-based parallel replicas. #121508 (Manuel).
• Parallel replicas now honor a SETTINGS clause that changes table eligibility when it is written inside a subquery or a view body. Such a clause was ignored, so a query over a view whose body sets parallel_replicas_for_non_replicated_merge_tree = 0, or sets parallel_replicas_allow_view_over_mergetree = 0 and reads a further view, returned max_parallel_replicas times too many rows, and a UNION over a view overriding parallel_replicas_allow_view_over_mergetree per subquery could throw Can't determine table for parallel replicas. #110020 (Groene AI).
• A query that reads a MATERIALIZED CTE more than once, for example inside an IN subquery, now works with parallel replicas when the initiator also reads locally (parallel_replicas_local_plan = 1). The other replicas no longer fail with UNKNOWN_TABLE on the CTE’s temporary table. Closes #112642. #123121 (Groene AI).
• Fixed a LOGICAL_ERROR “Block structure mismatch in QueryPipeline stream” when selecting a computed column through the merge table function or the Merge engine. If constant folding folded that column for one matched table and not another, the query failed. #115971 (Groene AI).
• Fix the exception No set is registered for key for a SELECT that reads a Distributed table through a Merge table or the merge table function and whose GROUP BY, DISTINCT, or LIMIT BY key contains an IN-family function. #117817 (Groene AI).
• Fix wrong results when reading a subcolumn of a Merge table column whose type differs between the source tables. Such a subcolumn exists only in the common type Merge derives, so it was filled with the default value for every row instead of being read. Most visibly, variantElement over a Merge table returned NULL for every row unless the call sat in the top-level projection next to another use of the whole column. #121462 (Alexey Milovidov).
• Fixed a server crash when a SELECT with serialize_query_plan = 1 read a Merge table through a table that forwards its reads, such as Alias, Buffer with a destination, or a MATERIALIZED VIEW with a TO target. Closes #119899. #119387 (Groene AI).
• Fixed wrong results when a Merge table (or the merge table function) reads a subcolumn of a column that one of its child tables does not have: for that child, x.null of a missing Nullable column x was read as 0, so x IS NULL returned no rows from it. Element subcolumns of Tuple, Variant and JSON columns and the sizes of a missing flattened Nested member were affected the same way. Closes #122523. #123633 (Groene AI).
• Fixed a Merge table returning the value of another column, or failing with a CAST error, for a column that one of its child tables does not have, when that child has a row policy or an ALIAS column. Such a column is now filled with its default value. #124674 (Groene AI).
• Fixed merge and Merge tables with a Distributed table among them returning the same GROUP BY group several times, each copy with a part of its aggregates, when the aggregation is two-level and a table lacks a grouping column or has it under another type than the common one. #124720 (Groene AI).
• Fixed a use-after-free when a query reads through an Alias table whose target is dropped concurrently. #120098 (Groene AI).
• Fix CREATE MATERIALIZED VIEW ... POPULATE AS SELECT count() FROM <MergeTree table> counting rows inserted during the population twice when materialized_views_populate_atomically is enabled. #123182 (Murphy).
• Fixed a materialized view over a system log table such as system.query_log silently stopping to receive rows after the server renamed the table to <name>_N because its definition changed (a server upgrade, or a change of <engine> or <ttl> in the config). Closes #124177. #124217 (Groene AI).
• Fixes wrong results when a materialized view reads an ordinary view over a table with a row policy that contains a subquery (x IN (SELECT ...) or a scalar subquery). The policy’s subquery now reads the real source table instead of the block being inserted into the materialized view. #124875 (Aditya Kumar).
• SYSTEM WAIT VIEW on a refreshable materialized view in a Replicated or Shared database no longer returns Ok. while a refresh requested on another replica has not run yet. Requested refreshes are now published in Keeper. #120623 (Samuel Krempaský).
• Fixed a segmentation fault when a refreshable materialized view is started up and shut down at the same time, for example when a failed ATTACH DATABASE rolls the database back, or when the server shuts down while tables are still loading asynchronously. #121037 (Groene AI).
• Fixed a refreshable materialized view with REFRESH DEPENDS ON (or REFRESH AFTER ... DEPENDS ON) and no REFRESH EVERY re-running its refresh without limit after a failure used up refresh_retries. It now waits for its dependencies’ next refresh. #123369 (Ashish Kumar Choubey).
• Fixed plain_rewritable disks leaving object storage and the in-memory filesystem in disagreement when a failed metadata transaction could not be reversed completely. Transient failures during reversal are now retried until object storage accepts the operation, and rollback runs for all exception types. Missing blobs are reported as CORRUPTED_DATA, and replacing a file with a missing source no longer deletes the existing target. The disagreement previously surfaced only after a restart, when a directory could appear under a name the server never committed. #119235 (Sema Checherinda). #122696 (Sema Checherinda).
• Fixed a query reading from file, s3, url, azure, or hdfs with session_timezone set getting the DateTime schema inferred by an earlier session with a different session_timezone, which formatted and computed the values in that other session’s time zone. #117811 (Alexey Milovidov).
• Fixed unnecessary object-storage reads, and <Error>-level exception messages such as Code: 499 ... The specified key does not exist ... (S3_ERROR), while destroying a data part on an object-storage disk when its skip indices are packed into skp_idx.packed. Evicting the part’s index marks from the mark cache no longer inspects the part’s on-disk layout, so it performs no I/O. #113454 (Groene AI).
• Fixed hive partition columns of glob-path object storage tables being permanently disabled when the first query over the table ran while the path matched no files. #120308 (Zach Rippas).
• Fixed unbounded memory and CPU usage when a path passed to file, s3, azureBlobStorage, or hdfs contains many {a,b,c} glob groups or long {N..M} ranges. Such a path is now rejected with an error. #118467 (Alexey Milovidov).
• Fix copyObjectToAnotherObjectStorage copying an object twice when the source and destination are the same object storage, which doubled the data transferred by a copy between two disks backed by that storage. #120953 (Alexey Milovidov).
• Fixed UNKNOWN_IDENTIFIER for a DEFAULT column that depends on a column used by a row policy, in tables with the File, URL, S3, and other object-storage engines. #122714 (Nikolai Kochetov).
• Fix object storage tables with partition_strategy = 'hive' being silently empty for formats whose name is not the file extension real files carry. The read glob was built from the lowercased format name (**.jsoneachrow, **.csvwithnames), so pointing a JSONEachRow table at a lake of .json / .jsonl files — or a CSVWithNames table at .csv files — read zero rows with no error or warning. The glob is now built from every file extension registered for the format (and for its WithNames/WithNamesAndTypes base format), plus additionally registered read extensions: JSONEachRow matches .json/.jsonl/.ndjson files and TabSeparated matches .tsv. Files written by ClickHouse itself keep matching. #120319 (Zach Rippas).
• Fixed reads of Hive-partitioned files (.../key=9/...) that need no column of the data file, such as count() with optimize_count_from_files = 0 or SELECT key, _row_number, returning no rows (Native threw an exception) when a partition column was smaller than the data columns. The wrong 0 was also cached and returned by later count() queries. The same reads of a structure that lists only partition columns returned no rows or threw an exception for most formats. Affects file, url, s3 and other object storages. #123381 (Groene AI).
• A query over an object storage table (e.g. S3) no longer waits for another query that is resolving the table’s hive partitioning on its first use, so KILL QUERY and max_execution_time apply to it while the endpoint is slow or unreachable. #123532 (Groene AI).
• Fixed BAD_ARGUMENTS (Expected positional arguments to go before key-value arguments) for a valid S3 argument list that combines headers(...) with a key-value argument, in either order, and for the engines and table functions sharing that parser (Iceberg, DeltaLake, Hudi, Paimon, GCS, S3Queue). With file_like_engine_default_partition_strategy = 'wildcard' the server persisted such a list itself, so a created table could no longer be attached and the server refused to start; affected tables load again. #119070 (Groene AI).
• Fix the logical error Attempt to read beyond right offset when the right bound of a read from Azure Blob Storage is lowered below bytes the reader has already buffered. #120957 (Alexey Milovidov).
• Fixed a WHERE condition or a row policy on a Hive partition column (use_hive_partitioning = 1) returning too few rows when the Parquet data file also has a column with that name and other values: Parquet row-group, page, and dictionary pruning used the file’s values instead of the path’s. The same applied to a virtual column such as _file when the data file has a physical column with that name. Affects file, url, s3, and other object storages. #124991 (Alexey Milovidov).
• Fixed an exception when a WHERE condition on the _file, _path, or _part virtual column guards a branch with if, and, or, or multiIf, for example WHERE _file LIKE 'events_%' AND toDate(substring(_file, 8, 10)) >= '2024-01-01'. Pruning files and parts by virtual columns now respects short_circuit_function_evaluation instead of evaluating the excluded branch for every file or part name. The same path now also rebuilds copied indexHint expressions and reduced and expressions with the right arguments, avoiding argument-count exceptions during lazy evaluation. #125033 (Groene AI). #125150 (Groene AI). #125126 (Groene AI).
• Fixed loading, short ATTACH, and UNDROP TABLE for existing SQL-defined S3 disks whose endpoints are absent from remote_url_allow_hosts; new full-definition CREATE and ATTACH still reject disallowed endpoints. #124930 (Alexey Bakharew).
• Fixed filesystem cache accounting when the directory for a new cache key cannot be created, for example because the cache disk is out of space or inodes. The failed reservation stayed counted in the cache until the query released the file segment. #122211 (Sema Checherinda).
• Fixed Parquet writing size statistics twice for a column chunk that falls back from dictionary encoding, which made unencoded_byte_array_data_bytes and the level histograms too large. The ParquetMetadata format now also reports size_statistics for each column chunk. #120027 (Erez Rokah).
• Fix two JSON object-size-limit bugs: bytes before the opening bracket are now counted toward the limit, and valid data is no longer rejected just because earlier rows in the same block already used the shared buffer. #115795 (Pavel Kruglov).
• Schema inference for the Freeform format no longer exhausts memory on a wide row of strings: the search for the row structure is bounded by the setting input_format_freeform_max_search_steps (default 4096, 0 means unlimited) and now also stops at max_memory_usage, max_execution_time, and KILL QUERY. #122447 (Mohammed Sohail). #123238 (Groene AI).
• Fix a crash when reading an ORC file whose root type is not a struct (for example, a file written with a bigint or array<bigint> root type). Such files are now rejected with an INCORRECT_DATA error. Closes #123305. #123337 (Groene AI).
• Fixed a possible out of bounds access when reading an ORC file. #123464 (Pedro Ferreira).
• Fix reading Avro files with a union whose named branches (records, enums, fixed) have the same structure. Before the fix, such files failed with The number of (non-null) union types in Avro record (N) does not match the number of types in destination Variant type (M), even with the inferred schema. #123236 (Dmitry Kovalev).
• Avro schema parsing no longer overflows the stack on a deep chain of named types. #124053 (Groene AI).
• Fixed a crash (stack overflow) on an Avro schema with a long chain of named type references, in schema inference (DESC, or SELECT without an explicit structure, for Avro and AvroConfluent), in reads into a Variant column and in Iceberg manifest reading. Such a schema now fails with an error instead. #124644 (Groene AI).
• Fixes a crash when reading an ORC file whose metadata has fewer stripe statistics than the file has stripes (for example, after a corrupt metadata_length) with input_format_orc_filter_push_down enabled. The ORC input format now reads such files without stripe-level filtering. #124315 (Groene AI).
• Fixed a crash (stack overflow) when reading an Avro file with an explicit structure that does not read a field nested through a long chain of named types, or a field of a recursive type holding deeply nested data such as a long linked list. Such a file now fails with TOO_DEEP_RECURSION instead. Closes #124925. #124964 (Groene AI).
• Fixed a stack overflow crash on a deeply nested CapnProto schema that the nesting depth check missed: a comment or a byte order mark between import and its path, an escape sequence in the import path, or a backtick text line. A backtick text line with many brackets is no longer rejected as nested too deeply. #125487 (Groene AI).
• Fixed a crash (stack overflow) on a CapnProto schema with a long chain of type expression suffixes (Int32.a.a..., Int32()()...) or of declarations that refer to each other (constants, aliases, struct fields, annotations, imports). Such a schema is now rejected with The CapnProto schema is nested too deeply. Closes #125782. #125900 (Groene AI).
• Fixed a data race on the file list of a File engine table. Two concurrent INSERTs with engine_file_allow_create_multiple_files = 1 could pick the same file name, which made SELECT count() over-report, rejected a valid INSERT with CANNOT_APPEND_TO_FILE, and could corrupt the heap. Closes #121517. #121598 (Groene AI).
• Fixed count() over file('*.csv', 'CSVWithNames') with optimize_count_from_files = 1 over-counting a CSV file by one row in rare cases. #120552 (Knut-Aage-Hofseth).
• Fixed concurrent INSERTs and SELECTs on a table with the File engine stalling for lock_acquire_timeout (120 seconds by default) and then failing with TIMEOUT_EXCEEDED. A single SELECT can open several readers on one File table, and an INSERT arriving between two of them made the query wait for itself. Closes #115931. #121779 (Groene AI).
• Fixed TinyLog returning fewer rows than the table contains when one of the data files being read ends before the others. An Array, Map, or Nested column contributes an elements file that holds one entry per element rather than per row, so it runs out early whenever the trailing rows hold empty containers, and a column holding only empty arrays contributes an elements file of zero bytes. The read stopped at the first data file by name to reach its end, cutting the result at a block boundary. #120307 (Groene AI).
• Fixed BACKUP and RESTORE queries with SETTINGS <name> = DEFAULT: such a query either failed to parse or was formatted into text that could not be parsed back, and the other settings of the same clause were discarded. Repeated async items now use the last value, async = '1' no longer fails after the operation has started, and compression is rejected in the SETTINGS clause as in other queries. #112978 (Groene AI).
• Fixed BACKUP writing a .backup manifest that is not well-formed XML when a value in it contained &, <, or >, for example BACKUP ... SETTINGS id = 'a&b', or an incremental backup whose base_backup path contained one of those characters. The backup reported BACKUP_CREATED, but every later RESTORE from it failed with SAXParseException: Invalid token, leaving the data unreachable. Such values are now escaped. Backups already written with an unescaped value remain unreadable and have to be taken again. #118898 (Julia Kartseva).
• Fixed RESTORE of system.functions failing, and aborting the whole restore, when the backup holds a SQL user-defined function whose definition the version that created it accepted but CREATE FUNCTION now rejects, for example CREATE FUNCTION f AS lambda(identity(x), x) created on 25.5 or earlier. #122912 (Groene AI).
• Fixes BACKUP ... ON CLUSTER creating a backup that cannot be restored when a host names its data files differently from the initiator: with data_file_name_generator = 'checksum' and a data_file_name_prefix_length other than the hosts’ backups.data_file_name_prefix_length (3 by default), or when a host’s config sets backups.data_file_name_generator = checksum and the initiator uses the default generator. Backups already created this way still cannot be restored. #122624 (Groene AI).
• Fixed RESTORE of a lightweight snapshot from Azure Blob Storage when the source disk endpoint includes a key prefix below the container name. #122697 (Garrett Thomas).
• Fixed column comments of a ReplicatedMergeTree table being lost in a backup when they were changed after the last metadata ALTER. #121309 (Nikolay Degterinsky).
• Fixed BACKUP / RESTORE ... ON CLUSTER hanging forever, unaffected by KILL QUERY and blocking server shutdown, when compatibility is set to 24.10 or older and a host fails on a Keeper error before it starts. Such a compatibility value no longer sets backup_restore_finish_timeout_after_error_sec, backup_restore_failure_after_host_disconnected_for_seconds, backup_restore_keeper_max_retries_while_initializing, and backup_restore_keeper_max_retries_while_handling_error to 0, a combination no release ever used. #123752 (Groene AI).
• Fixed BACKUP failing with A concurrent backup writing to the same destination ... detected when nothing else was writing there. The lock file is claimed with a conditional write, and a retry of a request that had already been applied made the backup mistake its own lock for someone else’s. #116376 (Alexey Milovidov).
• BACKUP and RESTORE to S3 now retry requests that fail with ExpiredToken after refreshing credentials. The retry time limit is 120 seconds by default and can be changed with backup_s3_expired_token_retry_timeout_ms. #123195 (Nikolai Ovchinnikov).
• BACKUP ... TO File(...) / Disk(...) now fsyncs the backup data files, the .backup manifest and the containing directories before reporting BACKUP_CREATED, so an acknowledged backup to local storage survives power loss. The guarantee covers backups.allowed_path and everything the backup writes below it. On a Disk(...) over object storage that keeps its metadata in local files, those metadata files are fsynced instead, since the backup is reached only through them. Controlled by the new backup setting fsync_backup_files (default true). S3(...)/AzureBlobStorage(...) destinations are unaffected. #111394 (Groene AI).
• Iceberg tables whose manifest file headers carry a schema that conflicts with the schema registered for the same schema-id in metadata.json are now readable: the metadata.json schema is preferred and the divergent manifest header copy is ignored with a warning, no matter which of the two copies is read first, matching the behavior of other query engines such as Spark, Trino, PyIceberg, and DuckDB. Controlled by the new setting iceberg_tolerate_conflicting_manifest_schemas (default true); compatibility below 26.9 restores the previous strict behavior, which failed such queries with ICEBERG_SPECIFICATION_VIOLATION. #119743 (Melvyn Peignon).
• Fixed Iceberg manifest statistics recording the wrong file count for added files. #120858 (Konstantin Vedernikov).
• Fixed Iceberg INSERT writing no lower_bounds or upper_bounds at all into the manifest entry when the table has a column whose type cannot be serialized into a statistics bound, or a column that is entirely NULL. Such a field is now omitted individually and the bounds of every other column are still written, so min/max file pruning keeps working. Closes #112706. #112934 (Groene AI).
• Fixed Iceberg date partition transforms being computed in the server time zone instead of UTC. #118583 (Konstantin Vedernikov).
• Fixed SHOW TABLES and system.tables failing on a database with the DataLakeCatalog engine over an Iceberg REST catalog when an unrelated namespace was dropped by another client while the listing was in flight. #113740 (Groene AI).
• Fixed validation of Iceberg manifest-list entries during writes. #122829 (Arsen Muk).
• Fixed metadata of empty Iceberg tables created by ClickHouse, so that Spark and other Iceberg Java-based tools can open them before the first insert and evolve their partition spec. #122077 (Asya Shneerson).
• Prevent loss of newer committed Iceberg data by rejecting OPTIMIZE when iceberg_metadata_file_path, iceberg_snapshot_id, or iceberg_timestamp_ms is explicitly set. #122683 (Arsen Muk).
• Fixed reading an Iceberg table partitioned by DateTime64 when a data file does not store the partition column: the query failed with DECIMAL_OVERFLOW. Decimal partition values stored in a manifest in their raw form are decoded as well, and system.iceberg_files.partition now shows partition values in a readable form instead of raw bytes. #117090 (Alexey Milovidov).
• Fix CREATE TABLE and DROP TABLE in Iceberg REST catalog databases: the catalog now writes the initial metadata file, ORDER BY and table properties are sent to the catalog, multi-level namespaces are created correctly, and DROP TABLE uses the correct endpoint. #123631 (Asya Shneerson).
• Fix SELECT count() over Iceberg tables returning too few rows after compaction leaves dangling position-delete files. Closes #123494. #123851 (Konstantin Vedernikov).
• Fixed reading Iceberg tables through the DataLakeCatalog database engine when the catalog reports a column type in a spelling the Iceberg type parser did not know: varbyte (Glue’s name for the Iceberg binary type), list<...> (the Iceberg spelling of array<...>, also written by Glue), and timestamp with local time zone (Hive 3’s name for the Iceberg timestamptz type). Such tables were rejected with Unknown Iceberg type. Closes #121850. Closes #87716. #121974 (Groene AI).
• The iceberg_metadata_staleness_ms setting is now respected when parsing Iceberg metadata, so Iceberg table functions and ON CLUSTER workers no longer list the whole metadata/ directory on every query. #124113 (Konstantin Vedernikov).
• Fix a server abort when reading an Iceberg table whose sort order names a transform ClickHouse cannot parse, e.g. bucket[xyz]; ClickHouse now ignores that sort-order hint and keeps the table readable. #121687 (Ilias).
• Fixed OPTIMIZE TABLE for OSS Iceberg tables so compaction no longer removes version-hint.text and breaks hint-based reads. #122833 (Arsen Muk).
• Fixed a crash (SIGSEGV) when reading a malformed deletion vector with the Puffin input format, for example one whose roaring run container has zero runs. Structurally invalid roaring bitmaps are now rejected with an error instead of crashing or returning wrong deleted-row positions. Closes #124688. #124735 (Groene AI).
• Fixed an INSERT into a Parquet Iceberg table failing with Logical error: 'Iceberg data file has 1 columns while its schema has 2 fields' when ALTER TABLE ... ADD COLUMN committed while the INSERT was running. #121821 (Groene AI).
• Fixed Iceberg v3 virtual columns _row_id and _last_updated_sequence_number: filters on them no longer return no rows for a data file rewritten by copy-on-write when the manifest has statistics for the table columns but not for the row-lineage columns, and rows of files written before the table was upgraded to format version 3 no longer return NULL for these columns. Closes #124271. #124591 (Groene AI). #124598 (Groene AI).
• Fixed OPTIMIZE TABLE ... MANIFEST failing with BAD_ARGUMENTS on an Iceberg table whose current snapshot summary omits an optional total-records, total-files-size, or total-data-files metric, which made manifest compaction permanently unavailable on tables written by engines that omit those metrics. Closes #117000. #117012 (Groene AI).
• Fixed Iceberg min/max pruning skipping a data file that holds matching rows when the file’s uuid bound was written by an external engine such as Spark, Trino, or PyIceberg. Such a bound now prunes only when its lower and upper bounds are equal. This fixes the Iceberg manifest bounds only: the uuid statistics of the Parquet data file itself are a separate defect, so a query at default settings can still miss rows. Closes #118366. #118380 (Groene AI).
• Fixed Iceberg writes storing ClickHouse unsigned integers in a narrower Iceberg integer type than the source. #118596 (Konstantin Vedernikov).
• ALTER TABLE ... EXECUTE remove_orphan_files on Iceberg tables now rejects an older_than that is less than iceberg_orphan_files_min_retention_seconds (default 1 day) before now, so it no longer deletes the files of a write that started within that window and has not committed yet. Setting iceberg_orphan_files_min_retention_seconds to 0 restores the previous behavior. Closes #125331. #125376 (Groene AI).
• Fixed DROP TABLE in a DataLakeCatalog database with an Iceberg REST catalog ignoring iceberg_delete_data_on_drop from the server’s default profile: ClickHouse always sent purgeRequested=False, so the catalog never deleted the table’s data, and Microsoft OneLake rejected the drop with HTTP 405. ClickHouse now sends purgeRequested=true or purgeRequested=false (lowercase, as the Iceberg REST spec defines), chosen by that setting; a session-level SET does not reach this drop. Closes #125410. #125469 (Groene AI).
• Fixed add.modificationTime in Delta Lake commit logs being written in seconds instead of the milliseconds the Delta protocol requires, so a data file written by ClickHouse appeared to Spark, delta-rs, and delta-kernel-rs as last modified in 1970. Closes #120523. #120564 (Groene AI).
• Fixed wrong values and a LOGICAL_ERROR when reading a partitioned DeltaLake table with allow_delta_kernel_rs = 0 and column types that differ from the ones in _delta_log. #121544 (Raúl Marín).
• Fix reading Delta Lake tables with column mapping where a field name with a dot, like a.b, equals the path of a nested field (the field b of a struct a). Such a field returned the other field’s values or NULL, SELECT * failed with COLUMN_QUERIED_MORE_THAN_ONCE, and a partitioned table failed with a LOGICAL_ERROR. #123102 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK when a query reads a struct column of a DeltaLake table with column mapping together with one of its fields, such as SELECT s, s.x, SELECT *, s.x, or SELECT s ... WHERE s.x = 1. Closes #123811. #123852 (Groene AI).
• Fixed INSERT into DeltaLake tables through a Unity DataLakeCatalog: the catalog now requests write credentials instead of reusing read-only ones. #123543 (Konstantin Vedernikov).
• Fix DeltaLake statistics pruning for filters that compare a Date32 column through CAST, toDate, or toDate32, including a wrong-results case where 1 >= d was treated as d >= 1. #119184 (Arsen Muk).
• Fixed INSERT into a DeltaLake table with a timestamp_ntz column writing the Parquet data file with isAdjustedToUTC=true, which made the committed table unreadable by Spark and other Delta engines. Such a column is now annotated TIMESTAMP(isAdjustedToUTC=false), as the Delta protocol requires. Closes #120653. #120863 (Groene AI).
• Fixed SHOW TABLES, system.tables scans, and table resolution hanging indefinitely against a paimon_rest catalog whose listings paginate. #115534 (JIaQi Tang).
• Fixed data lake tables created on top of a server disk (SETTINGS disk = '...'): they no longer overwrite the disk’s object storage settings with the table’s or the query’s settings, and they follow the disk’s configuration on SYSTEM RELOAD CONFIG instead. #117398 (alesapin).
• Reading an Apache Paimon primary-key table now throws NOT_IMPLEMENTED instead of silently returning wrong results. Merge-on-read is not implemented, so the reader returned the raw union of the snapshot’s data files, including the row versions superseded by later upserts. Closes #113360. #113460 (Groene AI).
• Fixed SHOW CREATE DATABASE and BACKUP of a DataLakeCatalog database omitting its COMMENT, and ALTER DATABASE ... MODIFY SETTING erasing that comment from the stored metadata, so that it was lost after a restart. #122942 (Groene AI).
• Reject Paimon manifest and data paths that escape the table directory. A crafted table could previously read files outside its own directory through an absolute path or .. components in metadata paths. #123099 (Konstantin Vedernikov).
• Fixed a data race between concurrent queries on one Hudi table: the list of the table’s data files was filled on first use without synchronization. #119137 (Alexey Milovidov).
• Fixed the Paimon incremental read silently duplicating and dropping data. Advancing the consumed-snapshot watermark in Keeper is now fenced by the processing_lock and monotonic, so a read that lost the lock can no longer commit over another consumer’s progress. A cursor left ahead of the warehouse, and a snapshot that fails to load for any reason other than Paimon expiration, now raise an error instead of being silently skipped. #117630 (JIaQi Tang).
• Fixed two ways a table with the Redis engine mishandled an MGET reply. Reading the whole table stopped at the first key without a string value and silently dropped every key scanned after it, so one key of another Redis type was enough to truncate the result. A reply carrying a different number of elements than keys requested could terminate the server; such a reply is now rejected with INTERNAL_REDIS_ERROR. #120230 (Groene AI).
• Fixed a broker-side queue leak when dropping a RabbitMQ table. DROP TABLE asked the broker to delete the table’s queues without waiting for its own consumer channels to close, so the broker could still see a registered consumer and refuse the deletion with PRECONDITION_FAILED. #116050 (Groene AI).
• Fixed spurious NATS connection failures when the vendored nats-io client gets EINTR while waiting for a socket, and picked up the upstream JetStream pull-subscription deadlock fix in the vendored client. #120202 (Alexey Milovidov).
• Fixed arrowFlight and the ArrowFlight table engine hanging forever when the remote Flight server accepts a request and never answers it. Requests are now bounded by the new arrow_flight_request_timeout_sec setting (300 seconds by default, 0 disables the bound), and a stalled read honors KILL QUERY and max_execution_time. #120231 (Groene AI).
• Fixed SELECT queries on an EmbeddedRocksDB table with a composite PRIMARY KEY returning fewer rows than match the predicate when the predicate combines IN sets with OR and one set’s element type cannot be converted to the key type. Such a set no longer narrows the candidate keys. #122261 (Groene AI).
• Fixed a FileLog table that stopped reading all of its files forever after a single read error (such as a line that failed to parse or a partially written last line): every later poll reported Another select query is running on this table and every direct SELECT failed with CANNOT_SELECT until the server was restarted. Related to #123368. #123395 (Groene AI).
• Fixed a NATS engine table reading from a JetStream stream (nats_stream) silently stopping to consume messages after the NATS server restarted or the connection was re-established. #115343 (Alexey Milovidov). #119301 (Alexey Milovidov).
• Fix a rare server crash (SIGSEGV in mongoc_log) when a MongoDB dictionary source or MongoDB table was created while another thread created the first MongoDB connection in the process. Messages logged by the MongoDB driver now go to the server log. #124235 (Vladimir Cherkasov).
• RENAME DATABASE is now refused when the database holds a loaded S3Queue or AzureQueue table whose keeper_path expanded the {database} macro when the table was loaded. Previously the rename succeeded silently, the table bound to a different Keeper path at the next server start and reprocessed every file it had already consumed. Not checked: detached tables, tables not yet materialized under lazy_load_tables = 1, and a keeper_path whose macros a config reload redefined after the table was loaded. #120103 (Groene AI).
• SELECT count() and system.tables.total_rows of a Join table that keeps every row of a key, ALL, or ANY / SEMI / ANTI with RIGHT, returned the number of distinct keys instead of the number of rows. #124099 (Groene AI).
• Fixed reading Enum8 and Enum16 columns of PostgreSQL and MySQL tables, Nullable(Enum8) and Nullable(Enum16) columns of PostgreSQL tables, and Enum range columns of a PostgreSQL dictionary source. Values are now returned as the declared Enum type, parsed by element name or by number, instead of the raw remote text typed as String or an Unexpected value N in enum error. A remote value that is not an element of the enum is now rejected with an error. #125198 (Groene AI).
• Fixes RabbitMQ tables with rabbitmq_num_consumers > 1 that insert into materialized views with parallel_view_processing = 0 (the default) falling behind: the consumers were read one after another, all but the first inserted a single message per flush interval, and their messages waited many intervals (up to about 28 s at 100 msg/s with a 1 s interval). They are now read in parallel. #125106 (Sven Eliasson).
• Fixed a server crash (segmentation fault) in a direct SELECT from a NATS table while the table replaces a connection that the NATS client library closed for good, for example after the broker rejected rotated credentials. Closes #120199. #125494 (Groene AI).
• Fixed a heap-use-after-free when an EmbeddedRocksDB table with a composite PRIMARY KEY was filtered by a tuple equality whose tuple has a LowCardinality element, e.g. WHERE (k, id) = (SELECT (name, id) FROM t LIMIT 1). Closes #120527. #125509 (Groene AI).
• Fixed the Arrow Flight interface returning garbage values for a MergeTree column stored in sparse serialization, which is the default for a column that holds mostly default values. Affected DoGet with a SQL ticket, GetFlightInfo and PollFlightInfo. #125553 (Groene AI).
• Fixed a server crash when a RabbitMQ table with the headers exchange type had a routing key without =, or when rabbitmq_routing_key_list was empty. Routing keys without = for the headers exchange type are now rejected with an exception. The problem was found and reported by @dstesk. #125661 (Alexey Milovidov).
• StorageKafka2 now counts only live replicas when distributing partition locks, so a replica that lost its Keeper session no longer leaves some partitions unconsumed. #120111 (johnjing).
• Fixed memory growth in Kafka tables with offsets stored in Keeper when partitions move between consumers: prefetched messages for released partitions are now freed rather than retained until restart. Closes #125077. #125076 (Aparajita Pandey).
• Fixed spurious TLS/SSL error: Interrupted system call failures in mysql, the MySQL table engine, the MySQL database engine, and MySQL dictionaries by updating mariadb-connector-c to v3.4.10, with a patch that clears the OpenSSL error queue before every TLS operation. #120526 (Groene AI).
• Fixed a durability bug in the MySQL and PostgreSQL database engines. The marker file recording whether a table is detached or dropped was created and removed without fsyncing the metadata directory, so a power loss could silently revert an acknowledged DETACH TABLE ... PERMANENTLY, DROP TABLE, or ATTACH TABLE and change the table’s visibility on restart. It is now fsynced when fsync_metadata is enabled (the default). A failed ATTACH TABLE no longer leaves the table half-attached, which made a retry report TABLE_ALREADY_EXISTS. #112170 (Groene AI).
• Using the PostgreSQL engine, inserts now quote table and column names so PostgreSQL does not fold them to lowercase. The destination table can therefore use any identifier casing. #120435 (Thomas Sileghem).
• Fixed KILL QUERY and max_execution_time having no effect on a query reading from PostgreSQL when the remote host stops sending data, and DROP of a refreshable materialized view whose refresh reads from PostgreSQL hanging for the same reason. #118995 (Nikolay Degterinsky).
• Fixed PostgreSQL connections losing send_timeout after the handshake. #123094 (Sema Checherinda).
• HTTP dictionary sources (SOURCE(HTTP(...))) now respect the max_http_get_redirects setting, so dictionaries can be created from URLs that redirect. Previously such dictionaries always failed with TOO_MANY_REDIRECTS. #101768 (Yash ).
• Fix optimize_inverse_dictionary_lookup rewriting dictGet comparisons incorrectly: complex-key dictionaries with a single key column called with a tuple-wrapped, possibly Nullable, key expression failed with ILLEGAL_TYPE_OF_ARGUMENT; key expressions that rely on dictGet’s implicit key type conversion, e.g. a String expression for a UUID key, failed with NO_COMMON_TYPE, and lossy ones silently returned results where dictGet throws; a match on a dictionary key containing NULL produced NULL instead of false; a single matching floating-point NaN or zero key produced wrong results; and mismatched key shapes were folded to a result instead of the dictGet error. Key expressions that would need a conversion are no longer rewritten and keep the dictGet lookup, except a signed integer probe of a simple-key dictionary, which is converted to the UInt64 lookup type the way dictGet converts it. #114865 (Artur Beglaryan).
• Fixed a dictionary with a local CLICKHOUSE source that stopped updating with ACCESS_DENIED and kept serving stale data after a new role was granted to its source user and the old role was dropped. Every query of such a source now uses the current default roles of the user, like a new session of that user. Closes #122663. #124051 (Groene AI).
• Fixed ClickHouse Keeper terminating when a client sent a multi request with no subrequests. Such a request was written to the Keeper changelog before it failed, so the termination repeated on every restart. #119847 (Alexey Milovidov).
• Fixed ClickHouse Keeper discarding a world:anyone ACL entry with all permissions when the request contained other entries. Such a node was stored more restrictively than requested and could end up repairable only with a super credential; with <path_acls> configured and no <identity>, a ReplicatedMergeTree under the configured path could not be created at all. These nodes now get the requested permissions, which grants every session full access to them, as ZooKeeper does. Closes #120177. #120265 (Groene AI).
• Fixed Keeper persistent recursive watch events to report the path of the changed node, matching ZooKeeper behavior. Previously, they reported the path where the watch was registered. #121333 (Kirill).
• Fix a segmentation fault in ClickHouse Keeper when preallocating space for a changelog file fails because the disk is full. #122337 (Alexey Milovidov).
• ClickHouse Keeper no longer sends the same watch event to one session several times when more than one of the session’s watches matches a change, for example persistent recursive watches on /a and /a/b for a create /a/b/c, or a single persistent watch on a removed node. Like ZooKeeper, Keeper now sends each event to a session once. Closes #122291. #122706 (Groene AI).
• Fixed a Keeper durability gap: with force_sync enabled, the directory entry of a new changelog segment, or of a segment continued after a restart, was never fsynced, so acknowledged writes in that segment could be lost after a power loss on filesystems where a file fsync does not persist the file’s name. Keeper now fsyncs the log directory before the first entries of such a segment are reported durable, and stops if that fails. #122755 (Groene AI).
• Fixed Keeper silently giving a new session to a ZooKeeper client, for example the Java client or kazoo, that reconnects and asks to continue its session. Such a client now gets SessionExpired, ZooKeeper’s answer for a session that cannot be continued, instead of losing its ephemeral nodes later without notice. Keeper still does not restore sessions. Closes #116690. #122767 (Groene AI).
• Fixed the Keeper HTTP storage API (/api/v1/storage) storing a truncated value and reporting success when a POST or PUT body was larger than max_request_size. Such requests are now rejected with 413, and the node is left unchanged. Closes #122286. #122928 (Groene AI).
• Fix Keeper recovery with remove_orphaned_nodes_on_startup leaving an unrepaired snapshot on disk. Persist the repaired snapshot so followers can install it successfully and the recovered node can restart with orphan recovery disabled. #123206 (Seva Potapov).
• Fixed a Keeper node that restarted after a crash sometimes giving a new client the session ID of a client that had connected to it before the restart, after which the new client received that client’s responses. #124307 (Groene AI).
• ClickHouse now rejects relative paths in checksums.txt. #120356 (alesapin).
• Encryption codecs (AES_128_GCM_SIV, AES_256_GCM_SIV) now also cover the structural substreams of a column. #120636 (Raufs Dunamalijevs).
• Rejected a compressed block whose codec stores data verbatim, such as NONE or Quantized, when its declared uncompressed size does not match the body length. Previously that passthrough read path accepted such a block while every other decompression path rejected it. #114344 (Groene AI).
• Reject invalid HTTP header names supplied to SQL-facing HTTP clients (the url table function and engine, HTTP dictionaries, S3/object storage, and DataLake/REST catalogs). A header name must be a valid RFC 9110 token (letters, digits and !#$%&'*+-.^_`|~); any other byte, :, whitespace, control characters, or separators such as / and =, is rejected. Such names could previously bypass the <http_forbid_headers> restriction. #118830 (Sergei Trifonov).
• Fixed ALTER QUOTA silently not taking effect when a quota is changed to KEYED BY client_key: a user whose quota context was already cached kept the previous, more permissive limits, and the failed refresh also stopped every other cached context from picking up unrelated quota changes. #120240 (Groene AI).
• Kafka producer and consumer trace logs now show secret librdkafka properties, such as sasl.password, as [HIDDEN] instead of their cleartext values. #119244 (János Benjamin Antal).
• EXPLAIN AST no longer prints secret function arguments, such as the key of encrypt, in cleartext, including with optimize = 1 for the body of a view the user may only SELECT from. #120313 (Harikrishnan Prabakaran).
• ClickHouse now masks credentials in system.query_log even for invalid URLs of external table engines. #122057 (alesapin).
• The value of role_session_name is now hidden in query_log. #122423 (Alexey Bakharew).
• BACKUP ... TO S3(...) and RESTORE ... FROM S3(...) now enforce remote_url_allow_hosts on the initial S3 URL, instead of checking only redirect targets. Endpoints that the s3 table function would reject now fail with UNACCEPTABLE_URL here as well. #123114 (Konstantin Vedernikov).
• A password or presigned signature in a URL is no longer printed in clear by the ConfluentSchemaRegistry trace log, by the URL ... is not allowed in configuration file, see <remote_url_allow_hosts> error, or by the Unsupported scheme in URI error. These now show scheme://[HIDDEN]@host and X-Amz-Signature=[HIDDEN], as system.query_log already does for format_avro_schema_registry_url. #123226 (Groene AI).
• Fixed SYSTEM SYNC TRANSACTION LOG (local, non-ON CLUSTER form) not checking the SYSTEM_SYNC_TRANSACTION_LOG privilege. The ON CLUSTER form already required it, so a user without the grant could run the local form when experimental transactions were enabled. #112441 (Azhar Izzannada E).
• The values of MongoDB connection options that carry a secret (tlsCertificateKeyFilePassword, sslClientCertificateKeyPassword, AWS_SESSION_TOKEN in authMechanismProperties) are now hidden in SHOW CREATE, system.query_log and system.dictionaries.source for the MongoDB table engine, the mongodb table function and the MongoDB dictionary source, system.dictionaries.source no longer shows the MongoDB password, and a named-collection uri override after another override is no longer shown as written. #123240 (Groene AI).
• A SELECT row policy on the mergeTreeProjection table function (_table_function.mergeTreeProjection) is now also applied to a read shipped to another server as a serialized query plan (serialize_query_plan = 1, with process_query_plan_packet enabled there), as it already is for a local read or with serialize_query_plan = 0. #123609 (Groene AI).
• SYSTEM DROP DATABASE REPLICA without a database name is now denied for a user without the SYSTEM DROP REPLICA privilege, instead of silently doing nothing and reporting success. This makes it consistent with SYSTEM DROP REPLICA. #117964 (Alexey Milovidov).
• ALTER TABLE ... MODIFY SETTING disk = disk(...) could register a disk defined in SQL whose local path is outside custom_local_disks_base_directory, and create its directory, bypassing the check that CREATE TABLE applies. Now it is rejected the same way, and a statement rejected this way leaves no disk and no directory behind. #119850 (Alexey Milovidov).
• Fixed ALTER TABLE ... ATTACH PARTITION ... FROM ... and CREATE TABLE ... CLONE AS ... requiring the ALTER DELETE privilege on the destination table. These statements only add data, so INSERT is now enough. REPLACE PARTITION, which does drop the destination data, still requires ALTER DELETE. A session temporary table used as the source of ATTACH or REPLACE PARTITION ... FROM also no longer wrongly requires SELECT on a same-named table of the current database. #117936 (Alexey Milovidov).
• Fixed the server terminating when an argument of a disk(...) setting value is a key without a value, such as disk(equals(name)). A malformed disk(...) argument no longer causes the whole query, including its other secrets, to be logged unmasked. #123847 (Groene AI).
• Mask secrets in the queries that a cluster-mode QueryRunner table writes to system.query_log as well as in queries executed as startup scripts. #118978 (Miсhael Stetsyuk).
• A RESTORE TABLE ... AS <name> of a backup whose engine arguments were stripped now validates the Keeper path it derives from default_replica_path. Previously the {database} and {table} substitutions were expanded against the restored name without being checked, so a name containing / resolved into another table’s Keeper subtree and left an inactive replica there. #121546 (Groene AI).
• Hide secrets in the system.mutations table. #122717 (alesapin).
• Creating a FileLog table now requires the READ and WRITE privileges, as for a File table. #125701 (Robert Schulze).
• Fixed an out-of-bounds read when reading a Variant column from a truncated Native stream using compact discriminators serialization. #120252 (Raufs Dunamalijevs).
• Reject malformed PostgreSQL startup messages whose parameter data exceeds the declared payload size. #112780 (Minh Vu).
• Reading settings from the Native protocol no longer sizes a string from the length a peer declares before the payload arrives, so a Query packet or a server hello can no longer turn a few bytes on the wire into a large memory allocation. This covers settings and query parameters, on the server and in the client. #118612 (Groene AI). #120931 (Alexey Milovidov).
• Fixed a crash when a client sends a block over the native TCP protocol that selects the internal Detached serialization kind for a column. #117699 (Pavel Kruglov).
• COPY options are no longer silently ignored on the PostgreSQL wire protocol: WITH (FORMAT csv) and the legacy WITH CSV are now applied, unsupported options are reported instead of skipped, and TSV header auto-detection no longer drops the first data row of a COPY ... FROM STDIN. #119376 (Alexey Milovidov).
• Bound the header column count in RowBinary formats to avoid excessive memory allocation from a crafted header. #120282 (George Larionov).
• Reject malformed Native blocks or corrupted parts that could make ClickHouse read past the end of a JSON or Dynamic column, for example because of duplicate types or paths, a truncated stream, or a part-only serialization version sent over the wire. #117765 (Pavel Kruglov).
• Fixed handshake_timeout_milliseconds not bounding a connection that stops sending during the handshake, which held a thread for receive_timeout (300 seconds by default) instead. The setting now also covers the MySQL and PostgreSQL handshake phases and the TLS negotiation; previously a peer that dribbled bytes during a TLS handshake, or stayed silent on the PostgreSQL port, could hold a thread indefinitely. HTTP header reading keeps its own http_headers_read_timeout budget, now enforced by the same mechanism. On tcp_port_secure, ClickHouse no longer writes an exception to a client whose TLS handshake has already timed out, so such a peer no longer holds a handler thread for twice the configured budget. #121547 (Sema Checherinda). #123101 (Sema Checherinda).
• Reject malformed binary input with out-of-range VarInt, integer, or enum values, including settings, client info, query plans, distributed task descriptions, and text indexes, instead of silently truncating them. #113944 (sunliqiang). #123242 (Sergei Trifonov).
• The collections of the Native protocol, the settings and query parameters of a query, the client information, the granted roles, and the table-status lists, are now bounded, so a small packet can no longer make the server allocate memory for data that has not arrived. #117015 (Alexey Milovidov).
• Fixed Code: 102. UNEXPECTED_PACKET_FROM_SERVER (... got Progress) aborting an INSERT that requests a table structure from the server, when a subquery is evaluated while the server analyses the query. The Progress and ProfileEvents packets sent in that window are now drained instead of rejected, both by the client and by the initiator of an INSERT into a Distributed or remote table, where rejecting them also hid the shard’s own error. Affects INSERT ... SELECT, input inserts, and INSERT ... VALUES with async_insert = 0. #118946 (Groene AI).
• Fixed the PostgreSQL wire protocol rejecting a simple query that consists only of comments, such as the -- ping that the Go driver pgx v5 sends from Ping, with Empty query. Such a query is now answered with EmptyQueryResponse, as in PostgreSQL. Closes #84085. #125161 (Groene AI).
• Fixed http_max_multipart_form_data_size silently truncating oversized multipart data instead of failing the query, and made it span a form’s external-table parts instead of resetting per part. #120036 (Sema Checherinda).
• KILL QUERY, query timeouts, and client disconnects now stop a query that reads over HTTP, for example over the url table function, while it is retrying a request, instead of waiting until all retry attempts are exhausted. #112930 (Alexey Milovidov).
• Fixed gRPC queries with transport compression (gzip, deflate) crashing the server, as they could in official builds since 26.4. Closes #115121. #122668 (Groene AI).
• Prevented asynchronous inserts that only need to buffer data from being blocked by another insert waiting for a flush worker while holding their shared queue lock. SYSTEM FLUSH ASYNC INSERT QUEUE now also waits for batches removed from the queue that are still waiting for a worker. #118352 (RJ).
• Fixed asynchronous INSERT under a quota KEYED BY client_key: the flush ran without the client’s quota key, so it failed with QUOTA_REQUIRES_CLIENT_KEY (or charged the wrong bucket). The quota key is now carried through the async batch and restored on flush, and background quota recomputes no longer log a spurious QUOTA_REQUIRES_CLIENT_KEY error. #122296 (Mikhail f. Shiryaev).
• Fixed max_memory_usage_for_user being exceeded by memory the user no longer holds. Memory that a query allocates but does not free itself, such as the data of in-memory table engines, cache entries and data waiting in the asynchronous insert queue, is now accounted to whoever owns it, and whatever a query leaves behind is settled when it ends. max_memory_usage_for_user is also applied when a query hands over memory it allocated before it had a user. #113647 (Raúl Marín).
• Queries that send data over pooled connections, such as INSERT into a Distributed table, are no longer charged for connection buffers that stay in the pool after they finish, which could count against max_memory_usage_for_user while another query of the same user was running. #125573 (Raúl Marín).
• Fix a possible hang of a query under workload CPU scheduling with cpu_slot_preemption enabled: when all its threads were preempted, the query could keep CPU slots it had already used and wait forever for one of them. #123640 (Sergei Trifonov).
• Fixed vector search returning too few rows for a query with arrayJoin below ORDER BY <distance> LIMIT. A row policy now counts as an additional filter for vector search, so vector_search_filter_strategy = 'prefilter' is honored and the index fetch multiplier is applied for a query filtered by a row policy. Fixed NOT_FOUND_COLUMN_IN_BLOCK for a non-rescoring vector search query when a WHERE filter or a row policy reads the vector column. #114074 (Alexey Milovidov).
• Fix a vector search returning fewer rows than its LIMIT, and missing true nearest neighbors, after a lightweight DELETE (or with a delete mutation applied on the fly): the vector index is no longer used for a part while it still holds deleted rows. #119371 (Alexey Milovidov).
• CREATE TABLE now rejects a SETTINGS name that is neither a setting of the table engine nor a query setting for the File, URL, S3 (with COSN, OSS, GCS), AzureBlobStorage, HDFS, Log, TinyLog, and StripeLog engines, with UNKNOWN_SETTING, as MergeTree already did. The name = DEFAULT and param_name = value spellings are rejected too, except where the whole clause is dropped before any engine sees it. Such a name used to be accepted, stored in the table metadata, and never read, so a typo like input_format_csv_delimiter for format_csv_delimiter silently did nothing. A table whose stored definition already names a non-setting still loads. Closes #120705. #120898 (Groene AI).
• arrayJoin in a projection or in a Distributed sharding key is now rejected at CREATE/ALTER time. It changes the number of rows, which made every insert into a table with such a projection fail with an exception, and misrouted or failed inserts into such a Distributed table. #117791 (Alexey Milovidov).
• Fixed ALTER TABLE ... MODIFY ORDER BY accepting a key expression over a column that the same ALTER retypes through a subcolumn, for example MODIFY COLUMN x Tuple(a UInt32, b UInt32), MODIFY ORDER BY (k, x.a), or renames, for example RENAME COLUMN x TO y, MODIFY ORDER BY (k, y). Existing parts are not sorted by such a key, so queries that rely on the sort order could return wrong results. Such an ALTER is now refused with BAD_ARGUMENTS. Closes #122685. #122765 (Groene AI).
• Fixed ALTER TABLE ... MODIFY QUERY deleting the column comments of a materialized view with a TO table. A column that the new query still produces now keeps its comment. #122775 (Groene AI).
• Fixed ALTER TABLE <mv> COMMENT COLUMN <col> '...' and comment-only ALTER TABLE <mv> MODIFY COLUMN <col> COMMENT '...' returning success and storing nothing on a materialized view with an inner table, which left no supported way to change a column comment on such a view after CREATE. #119469 (Groene AI). #123033 (Pedro Ferreira).
• Fix DROP NAMED COLLECTION succeeding while a MongoDB table uses the collection, after which the server could not attach the table on restart. The drop is now refused with NAMED_COLLECTION_IS_USED, as for other table engines. #123636 (Groene AI).
• A sharding key of a Distributed table, a Remote table or the remote/cluster table functions that contains IN with a subquery or with a non-Set table is now rejected when the table is defined. The set of such an IN was never built, so every INSERT failed with a Not-ready Set logical error and optimize_skip_unused_shards could prune to the wrong shard and return wrong results. A table defined earlier with such a key still loads, but an INSERT that needs the key fails with BAD_ARGUMENTS and shard pruning does not use it. Pruning by a key with IN over a Set table, which can change after rows are placed, now needs allow_nondeterministic_optimize_skip_unused_shards, as for joinGet. #124646 (Groene AI).
• Reject a RENAME DATABASE that would make the table dependency graph cyclic, for example one that leaves an Alias table referring to itself. Such a rename put a cycle into the server-wide table dependency graph, after which every unrelated CREATE that adds a dependency edge failed with INFINITE_LOOP and the renamed database stopped being listable after a restart. #117253 (Alexey Milovidov).
• Fixed formatting of a function named any with a single subquery argument: SELECT 1 = `any`((SELECT 1)) lost its backticks, and because bare ANY before a subquery is the SQL quantifier, formatting the query changed its meaning to IN. This affected round-trips such as stored view definitions and SHOW CREATE. #120175 (Groene AI).
• Fixed parsing of a column named from inside parentheses, such as WHERE (from IN ('a')), which since 26.8 was read as a subquery starting with the FROM clause and failed with Unknown table function IN, Unknown table expression identifier, or a syntax error. #120950 (Alexey Milovidov).
• Preserve quoting for identifiers named cube, rollup, and recursive so formatted queries can be parsed back correctly. #121896 (Jaysurya Ray).
• Fix SET <setting> = DEFAULT ignoring an active compatibility setting: the setting now gets the value of that version, unless the settings constraints forbid it, and SET compatibility = DEFAULT reverts the settings it changed. Closes #116883. #116933 (Groene AI).
• Fixed direct config substitutions via from_env failing when XML-1.0-legal values contain XML special characters like &, <, >, or the ]]> sequence; existing <include from_env="..."/> XML-subtree substitutions are unchanged. An explicit <include from_zk="..." yaml="true"/> substitution can read a YAML subtree wherever the generic <include> form is used, including inside a leaf element, so it must not be used for a literal secret or setting. Without the explicit yaml="true" attribute, a non-< value is literal text, but direct non-< leaf from_zk substitutions are recommended for exact-byte secrets or settings, because surrounding whitespace of an <include> becomes part of a leaf value. Values beginning with < continue to be interpreted as XML fragments for backward compatibility. Upgrade note: a direct from_env value or a non-< leaf from_zk value that was previously XML-entity-encoded, for example a&b, now resolves to its exact literal bytes and must be stored raw as a&b instead. #101773 (Yash ).
• Fixed the server refusing to start with UNKNOWN_ELEMENT_IN_CONFIG when the definition files of executable user-defined functions or of their drivers are placed in config.d (for example with <user_defined_executable_functions_config>config.d/*_function.xml</user_defined_executable_functions_config>), which also merges their <function> or <driver> elements into the server config. Closes #123308. #123586 (Groene AI).
• Queries over system.parts, system.parts_columns, system.projection_parts, system.projection_parts_columns, and system.dropped_tables_parts now respect cancellation and max_execution_time while the result is being built. #113909 (Alexey Milovidov).
• Fixed a data race on the exception message for a cancelled query: threads no longer append context concurrently to one shared exception object. #119077 (Groene AI).
• Fixed unbounded memory growth from time zone names that were not names of the time zone database. A time zone name now has to be one that system.time_zones lists, or a fixed UTC offset spelled Fixed/UTC±HH:MM:SS. #120288 (Alexey Milovidov).
• Fixed a server abort during transaction rollback: internal exception-reporting helpers could call std::terminate while no exception was being handled. #122891 (Murphy).
• Fixed a crash when synchronous logging (logger.async = 0) had no log sink that accepted a message. #123336 (Groene AI).
• Fixed a permanent hang of the server on macOS, where threads that format a time (for example while signing S3 requests, reporting an exceeded quota or reloading a dictionary) could wait forever on the C library’s internal time zone lock while the query profiler was enabled. #125369 (Groene AI).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK on a table that has a physical column named _part shadowing the virtual column, for queries such as SELECT count() FROM t WHERE isNull(_part > 0) and SELECT count() FROM t WHERE _partition_id = 'all'. #117539 (Alexey Milovidov).
• The statistics cache is now refreshed for read-only MergeTree tables. #119559 (Rory Shanks).
• A MergeTree table with table_disk = 1 mounts the parts already present at the disk root. If it declares a sorting key those parts were not written with, reading it raised LOGICAL_ERROR and aborted the server in builds with assertions enabled. It now reports INCORRECT_DATA and names the offending part. #119295 (Groene AI).
• Fix a refresh of a table whose disks are all read-only (SYSTEM RESTART DISK, or the background refresh_parts_interval task) failing with a LOGICAL_ERROR and hiding committed parts when a rolled-back part covers them. Follow-up to #100992, which fixed the same case on the startup part-loading path. #111167 (Groene AI).
• SYSTEM SYNC MERGES no longer returns while the part covering the scheduled merges is still uncommitted: it now waits until a new transaction can see that part. #123396 (Murphy).
• Reject empty part names in CHECK TABLE. #123625 (Minh Vu).
• Partition commands on a DateTime64 key accept a literal with . as the date separator again, such as '2024.02.19 00:00:00.500'. #122995 (DeviousCardi).
• A CHECK TABLE running while a part writes txn_version.txt no longer fails with Unexpected file txn_version.txt.tmp in data part. #124158 (Murphy).
• Fixed ALTER TABLE ... MATERIALIZE TTL failing with UNKNOWN_IDENTIFIER and being retried forever on a table with ttl_only_drop_parts = 1 when a part still holds a column that the table dropped while that part’s partition was detached. Mutating such a partition after cloning it into another table with ATTACH PARTITION FROM no longer fails with a logical error either. #115962 (Groene AI).
• Fixed a member of a Nested added by ALTER TABLE ... ADD COLUMN to a Memory table being read as empty arrays for the rows inserted before the ALTER, unless another member of the same Nested was read at the same time. It now gets the array sizes of the other members, like in MergeTree. This fixes wrong results of SELECT, WHERE, PREWHERE, ARRAY JOIN, and row policies over such a column, and SIZES_OF_ARRAYS_DONT_MATCH errors. #122991 (Groene AI).
• Fixed SELECT from system.part_moves_between_shards and KILL PART_MOVE_TO_SHARD failing with Code: 999 ... No node, path .../part_moves_shard when any replicated table on the server is attached without its Keeper tree. #122132 (Groene AI).
• Fixed an exception (BAD_GET or NOT_IMPLEMENTED) when a filter on a table with a _part_offset projection had a Float or Nullable(Nothing) condition over a column the projection does not store. #124568 (Groene AI).
• Fixed a skip index recording the wrong expression type when ALTER TABLE ran in a session whose settings change how that expression resolves. toString(v) over a Dynamic column is Nullable(String) under cast_keep_nullable and String at the server baseline, so such an ALTER stored Nullable(String) while the writers resolved String, and the next index rebuild failed an internal type assertion. #119799 (Groene AI).
• Fixed hasPhrase rejecting FixedString input for some tokenizers, such as icu and splitByRegexp('\W+'). Closes #123426. #123585 (Elmi Ahmadov).
• Fixed statistics about the number of selected rows in trivial count queries answered from a text index. #125380 (Anton Popov).
• ALTER TABLE ... RESET SETTING auto_statistics_types now rebuilds the automatic column statistics from the default types immediately. Previously the table kept the statistics of the old setting value until the server was restarted, so system.columns and the statistics written into new parts disagreed with the table definition. #123599 (Rion Bora).
• Fix basic statistics of parts written by 26.6 or 26.7: after ALTER TABLE ... UPDATE or a lightweight DELETE, non-Nullable columns reported a default-value count of 0, which made selectivity estimates wrong. Query results were not affected. To repair parts that are already affected, run ALTER TABLE ... MATERIALIZE STATISTICS ALL. #123536 (Samuel Krempaský).
• Fixed indexHint with a Nothing-typed argument, such as indexHint(assumeNotNull(materialize(NULL))), which threw Cannot create non-empty column with type Nothing instead of only narrowing the read. #117635 (Alexey Milovidov).
• Fixed arrayReduce over a non-deterministic aggregate such as an unseeded groupArraySample being treated as deterministic, which pushed a HAVING predicate below the aggregation and collapsed per-row draws into a single one. #118971 (Alexey Milovidov).
• Fixed an internal error when two * matchers whose column transformers differ in kind at the same position are compared, for example two lambdas sharing an alias with bodies * EXCEPT (a) and * REPLACE (1 AS a). Such a query now reports MULTIPLE_EXPRESSIONS_FOR_ALIAS. #120234 (Groene AI).
• Fixed a LOGICAL_ERROR (Can't set alias of ... of QualifiedAsterisk) when a column transformer that assigns a name, such as APPLY with a name prefix or REPLACE, is applied to an expression that expands to an asterisk or a COLUMNS matcher. Such queries now return a clear BAD_ARGUMENTS exception. #109223 (Groene AI).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK for LIMIT BY on an injective function of a GROUP BY key, such as GROUP BY toString(c0) LIMIT 1 BY toString(c0). #120070 (Alexey Milovidov).
• Fixed LOGICAL_ERROR: Incorrect table expression when EXPLAIN AST optimize = 1 or EXPLAIN SYNTAX reads from the experimental table function eval. #121272 (Groene AI).
• A non-finite FROM, TO, STEP, or STALENESS value in ORDER BY ... WITH FILL (nan, inf, or -inf) is now rejected with INVALID_WITH_FILL_EXPRESSION. Previously such a value either made the query generate fill rows endlessly (TO nan, TO inf ascending, TO -inf descending, and the same values for STALENESS) or produced a wrong result: a non-finite FROM emitted the bound itself as a row and dropped the fill rows that TO asked for, while a non-finite STEP, or a TO bound the fill runs away from, returned the rows unfilled. Pass a finite value instead, or omit it. #118434 (Alexey Milovidov).
• Fixed exception Trying to compare constants of incompatible types for a query with an IN list over a column of a table that has a CONSTRAINT ... ASSUME, when both convert_query_to_cnf and optimize_using_constraints are enabled. #118278 (Alexey Milovidov).
• EXPLAIN AST no longer throws when a Backup locator name looks like a COLUMNS matcher. #122870 (Groene AI).
• ORDER BY ... WITH FILL ... INTERPOLATE of the fill column itself is now rejected over a Distributed table as well, with the same exception a local table already reported, instead of returning rows whose fill column was overwritten by the interpolated expression. #118304 (Alexey Milovidov).
• Fix NOT_FOUND_COLUMN_IN_BLOCK when a query over the filesystem table function combines indexHint with an ordinary condition on the same column. #118671 (Alexey Milovidov).
• Fixed wrong results from optimize_using_constraints with a CONSTRAINT ... ASSUME NOT (...): a negated comparison could be added to the constraints graph as if it were positive, making predicates such as a < b appear always true and a >= b always false. #121255 (Alexey Milovidov).
• A HAVING clause referencing a SELECT alias whose quoted name contains a dot, as in SELECT sum(t1.c1) AS `t1.c1` FROM t1 AS t1 HAVING `t1.c1` > 0, is no longer rejected with NOT_AN_AGGREGATE. The identifier resolution cache conflated that quoted one-part name with the qualified t1.c1, which could also return a wrong result instead of an error when both spellings met in one query. Default settings were affected from 26.6, where enable_identifier_resolve_cache became enabled by default. #122017 (Groene AI).
• The analyzer now rejects arrayJoin in INTERPOLATE and in HAVING with WITH TOTALS. Before, INTERPOLATE used only the first array element, and HAVING failed only during execution. #123538 (Yarik Briukhovetskyi).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK when a comparison of single-element tuples whose elements are tuples, such as tuple((a, b)) = tuple((c, d)), is part of an alias used in GROUP BY and also in ORDER BY, HAVING, or WHERE. Such comparisons are now rewritten completely, so WHERE tuple(tuple(x)) = tuple(tuple(y)) over a JOIN no longer returns rows where x and y are both NULL, the same as WHERE tuple(x) = tuple(y). #124546 (Groene AI).
• Fixed the exception Not-ready Set is passed as the second argument for function 'in' when the additional_result_filter setting contains an IN subquery. Such a query now applies the filter instead of failing. #117719 (Groene AI).
• A nested correlated subquery that cannot be decorrelated now reports NOT_IMPLEMENTED instead of the internal error Cannot find column ... in actions DAG input. #118317 (Alexey Milovidov).
• Fixed NOT_IMPLEMENTED for a correlated subquery in the SELECT list of a query that has a WHERE, when the correlated column is not also selected, and NOT_FOUND_COLUMN_IN_BLOCK for a correlated subquery in HAVING of a query with WITH TOTALS. #118402 (Alexey Milovidov).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK for a correlated subquery that reads a tuple element of an outer column by position (o.p.1), and for the same shape on the left of IN (subquery) under rewrite_in_to_join. #118287 (Alexey Milovidov).
• Fixed Cannot evaluate correlated scalar subquery for a query that uses one correlated scalar subquery twice in the same expression, for example in both branches of if(1 = 1, ..., ...). #118321 (Alexey Milovidov).
• Fixed an out-of-bounds read when finalizing a crafted analysisOfVariance aggregate state whose internal arrays have mismatched lengths. #107860 (uwezkhan).
• Fixed an out-of-bounds read of the parsed pattern in sequenceMatch, sequenceCount, and sequenceMatchEvents. When the trailing part of a pattern consisted only of actions that a zero-length match satisfies (.* and (?t...) conditions), the matcher read one element past the end of its action list, so sequenceMatch and sequenceCount could return 0 for a pattern that does match. #119871 (Groene AI).
• Fix an out-of-bounds read when deserializing a uniqTheta state shorter than 24 bytes, for instance from CAST(<string> AS AggregateFunction(uniqTheta, ...)). #119598 (Alexey Milovidov).
• Fixed regr_r2 returning 1 instead of nan when the data contains a NaN or an infinity. regr_sxx and regr_syy returned 0 in the same case. #123249 (mosya415).
• Reject malformed serialized groupBitmap states with INCORRECT_DATA instead of failing later during bitmap operations. #124049 (Elmi Ahmadov).
• Fixed wrong results in the largestTriangleThreeBuckets aggregate function for DateTime64 arguments close to the type’s upper bound: the maximum DateTime64(9) value was returned as 1677-09-21 00:12:43.145224192. Such values are now saturated instead of wrapping, and a deserialized state whose Date, Date32, DateTime or DateTime64 coordinates are out of range or fractional is rejected with an error. #120284 (Groene AI).
• Fix a wrong result, or an exception in debug builds, when optimize_aggregators_of_group_by_keys eliminated any, min or max of a GROUPING SETS key with group_by_use_nulls. #125575 (Alexey Milovidov).
• Fixed several URL-parsing bugs in domainRFC and portRFC for bracketed IPv6 hosts: an embedded IPv4 tail (for example [::ffff:192.168.0.1]) was rejected, non-IPv6 bracket contents and garbage following the closing bracket were silently accepted, and userinfo state could leak into host validation. Also fixed domainRFC and portRFC resolving the wrong host when a URL’s userinfo contains more than one @, rejected a host starting with a literal dot in domain and domainRFC, and fixed firstSignificantSubdomainRFC and cutToFirstSignificantSubdomainRFC returning a bogus subdomain for a bracketed IPv6 host instead of an empty result. #118979 (Luís Lizardo).
• Fixed IPv4StringToNum, IPv6StringToNum, isIPAddressInRange, parseDateTime, parseDateTime64, parseDateTimeInJodaSyntax, parseTimeDelta, parseReadableSize, toModifiedJulianDay, JSONMergePatch, stringToH3, and the H3 index functions throwing an exception on entirely valid LowCardinality data. #117359 (Alexey Milovidov).
• Fixed the exception Cannot convert out of range floating point value to integer type when a logical function (and, or, xor) receives a constant floating point argument outside {0, 1}, such as SELECT 1 OR 2.0. #117965 (Alexey Milovidov).
• timeSeriesRange and timeSeriesFromGrid now throw ARGUMENT_OUT_OF_BOUND when the requested number of points would exceed function_range_max_elements_in_block, matching the behavior of range. Previously an unbounded request would run for tens of seconds and allocate hundreds of gigabytes before hitting the memory limit. #121056 (Ilya Yatsishin).
• bitShiftLeft, bitShiftRight, bitRotateLeft, and bitRotateRight no longer answer, once the expression is compiled, for the arguments their interpreted implementation rejects: a big-integer or negative shift amount. #118716 (Alexey Milovidov).
• Fix mapPopulateSeries keeping the wrong value when duplicate keys are present. #119832 (Minh Vu).
• Fix format handling of escaped braces around replacement fields. #120704 (Minh Vu).
• jaroSimilarity and jaroWinklerSimilarity now return 0 when one input is empty. #119710 (Minh Vu).
• Fixed plus and minus over an array of dates and an array of tuples holding NULL, for example toString([toDateTime64(1, 3)] + [(2, NULL)]), and the exception ColumnTuple cannot have ColumnConst as its element when adding a constant interval to a non-constant interval of another unit, for example toIntervalDay(number) + INTERVAL 1 HOUR. #122451 (Groene AI).
• Fixed timeSeriesCopyTags and timeSeriesCopyTag when both group arguments are constant. They returned a single row regardless of the number of input rows, so such a call over several rows failed with Invalid number of rows in Chunk (LOGICAL_ERROR), and a single-row query printed an extra row in clickhouse-client. #122990 (Groene AI).
• Fixed if and CASE over tuples failing with ColumnTuple cannot have ColumnConst as its element when an element is NULL in both branches, for example if(number = 0, (1, NULL), (number, NULL)). #123268 (Groene AI).
• Fixed geometryIntersectCartesian and geometryIntersectSpherical returning an arbitrary result (usually 1) for an empty LineString, or a MultiLineString with an empty element, against a Ring, Polygon, or MultiPolygon. For such a MultiLineString, polygonsDistanceCartesian and polygonsDistanceSpherical could return 0. #124101 (Groene AI).
• Fixed an infinite, uncancellable loop in functions hop, hopStart and hopEnd when subtracting the hop interval does not move the window end, for example hopEnd(toDateTime(0, 'UTC'), toIntervalQuarter(413844), toIntervalQuarter(413844)), where the date saturates at the calendar limit. Such arguments now throw the Time overflow error. #124570 (Groene AI).
• Fixed the result type of randConstant, rand, rand32, rand64, randCanonical, generateUUIDv4, generateUUIDv7 and generateULID when their optional ignored argument has type Nothing: they declared Nothing rather than their documented type. randConstant with more than one argument also raised LOGICAL_ERROR during query analysis instead of reporting the wrong argument count. #116072 (Groene AI).
• KILL QUERY and max_execution_time (including timeout_overflow_mode = 'break') now stop AI functions in the middle of a block, instead of only after all requests for the block have finished. #125698 (George Larionov).
• Fixed the exception Provided time zone must be non-empty when the time zone argument of formatDateTime, formatDateTimeInJodaSyntax, fromUnixTimestamp, or fromUnixTimestampInJodaSyntax is a LowCardinality column. #117006 (Alexey Milovidov).
• Fixed toWeek and toYearWeek with mode 8 returning week 135 for Date32 and DateTime64 values in the year 1900 and earlier. The wrong values also made the primary key analysis drop matching rows for predicates on toYearWeek(key, 8). #117694 (Alexey Milovidov).
• An out-of-range integer constant in an IN set for a DateTime64 or Time64 column no longer wraps to a negative timestamp and matches an unrelated row. #118656 (Alexey Milovidov).
• Fixed makeDate and makeDate32 with two arguments to accept day 366 in leap years. #119363 (Minh Vu).
• Fixed CAST of an empty container to a container of Variant or Dynamic, which was rejected with Cannot convert type Nothing to Variant(...) or Variant type should have at least one nested type. This also fixes a query returning an empty Array(Variant(...)) failing once part of it runs on a secondary server, where the folded constant is re-rendered as a _CAST of an empty container. #118038 (Groene AI).
• Fixed TYPE_MISMATCH (CAST AS Array can only be performed between same-dimensional array types) when casting an Array(Map(...)) to an array whose element type contains the Map inside a Variant, or to Array(Dynamic), Array(JSON), or Array(String). #120948 (Alexey Milovidov).
• Fixed arraySlice and arrayElement on a QBit column replicated by ARRAY JOIN or a JOIN: debug builds raised LOGICAL_ERROR, while release builds could return wrong results. #120224 (Groene AI).
• Fixed a function over a Dynamic column reporting LOGICAL_ERROR when an unrelated exception, such as MEMORY_LIMIT_EXCEEDED, was raised while converting the nested result to the declared result type. Such exceptions now keep their own error code, message, and stack instead of being reported as an internal error. #121280 (Groene AI).
• The JSON type and the JSON* functions now also accept String values holding bytes that are not valid UTF-8 on builds where simdjson uses its scalar fallback backend (s390x, riscv64 and loongarch64 without RVV / LSX, amd64compat on CPUs without SSE4.2 or PCLMULQDQ), which previously failed with INCORRECT_DATA. #123945 (Groene AI).
• Fixed an exception when a function applied to a Dynamic column returns a result of the Nothing type, for example arrayElement on an empty array literal. Such queries now return NULL for every row instead of failing with std::out_of_range or Variant type should have at least one nested type. #115982 (Groene AI).
• Fixed CAST AS QBit can only be performed from String, Array or another QBit (and Unsupported types for IN) errors when a query with a QBit constant, such as WHERE q = CAST([1, 2], 'QBit(Float32, 2)') or q IN (...), runs over remote, a Distributed table, or parallel replicas. #125038 (Groene AI).
• Fix LOGICAL_ERROR: Column identifier ... is already registered when analyzer_compatibility_join_using_top_level_identifier = 1 and a JOIN ... USING key resolves to an ALIAS column of a table function such as merge or remote. NATURAL JOIN now also matches ALIAS columns of a table function, as it already did for a table, and no longer takes a non-selectable EPHEMERAL column as a join key on either side; with no other shared name it behaves as a CROSS JOIN rather than failing with Code: 16. #118734 (Groene AI).
• With analyzer_compatibility_join_using_top_level_identifier = 1, a JOIN USING identifier of the outermost JOIN is now also resolved from an alias defined in the WITH clause or in another clause of the query, not only from the SELECT list, and duplicated aliases with identical definitions no longer prevent the resolution; an inner JOIN of a query with several joins resolves it from its left table, as the old analyzer did. This fixes UNKNOWN_IDENTIFIER for such queries and a USING clause expected identifier error when an inner USING join is placed below an aliased outermost USING key. Closes #121401. #121975 (Vladimir Cherkasov). #123103 (Vladimir Cherkasov).
• Fixed a null dereference while formatting the LOGICAL_ERROR for a USING column with no expression; the query now reports the column name instead. Closes #123292. #123296 (Raúl Marín).
• Fix the AST predicate push-down into distributed subqueries treating non-deterministic EXECUTABLE and WASM user-defined functions as deterministic. It pushed a predicate over such a function’s result to the shard, which evaluated the function again. #119865 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR (Initiator received more initial requests than there are replicas, or Duplicate announcement received for replica number N) when a mergeTreeAnalyzeIndexes predicate contained a correlated subquery and parallel replicas were enabled. #119860 (Groene AI).
• Parallel replicas are kept enabled for cluster engines when automatic_parallel_replicas_mode is set. Also, from now on, with enable_parallel_replicas = 2 parallel replicas are forced and the automatic mode is not applied. #122306 (Nikita Taranov).
• Fixed NO_SUCH_DATA_PART in a make_distributed_plan bucketed read when a worker’s own index analysis prunes a part that the coordinator selected, for example with use_index_for_in_with_subqueries = 0 and an IN (subquery) predicate. Such parts are now skipped instead of producing a false “part … selected by the coordinator is not available on this replica” error. Added DistributedPlanWorkerPartsReceived, DistributedPlanWorkerPartsScanned, and DistributedPlanWorkerPartsPruned profile events. #115514 (Groene AI).
• A serialized query plan carrying max_block_size = 0 is now rejected when it is read, instead of constructing steps with a zero block size that could make the receiving server hang. #120606 (Nihal Z. Miaji).
• Fixed a case where make_distributed_plan fell back to local execution for the outer query but still built inner IN subquery sets as distributed plans inside views, scalar subqueries, GLOBAL IN subqueries, Distributed shard plans, materialized CTEs, and TimeSeries reads. #119986 (Antonio Filipovic).
• make_distributed_plan queries that resolve dictionary functions, joinGet, or MergeTree column defaults only on the initiator now fall back to local execution, avoiding worker-side Dictionary (...) not found failures and similar unresolved-reference errors. #121198 (Antonio Filipovic). #121567 (Antonio Filipovic).
• A cancelled or failed distributed query with make_distributed_plan = 1 no longer waits for the handshake timeout of a streaming exchange whose producer does not answer. Closes #122395. #122422 (Groene AI).
• Fixed distributed query plans losing a LIMIT early-stop signal on an in-memory exchange when a port closed concurrently. Such queries could keep reading whole inputs or wait for exchange streams after the LIMIT was already satisfied. #122453 (Groene AI).
• Fixed the LOGICAL_ERROR Reading from materialized CTE '...' before its materialization completed - DelayedPortsProcessor gate is missing in the query plan on a table read with more than one stream, raised by a lightweight UPDATE or DELETE whose predicate passes an IN subquery that defines a MATERIALIZED CTE to a function such as ignore, and by a lightweight UPDATE/DELETE or a SELECT when an IN or other set subquery defines a MATERIALIZED CTE that two other MATERIALIZED CTEs of that subquery read. #119715 (Alexey Milovidov). #121010 (Groene AI).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK when a predicate on a Merge table’s _table or _database virtual column is wrapped in indexHint. #117612 (Alexey Milovidov).
• Fix Column `__array_join_exp_1` not found in table for ARRAY JOIN with a filter on the array-joined column over a Merge table whose child is a Distributed table. #119307 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR Unexpected return type when reading a Merge table whose underlying table is another Merge table with different column types, with optimize_prewhere_after_pushdown = 1 and move_all_conditions_to_prewhere = 0. #124037 (Groene AI).
• Fixed TRUNCATE TABLE failing with DEADLOCK_AVOIDED, leaving the data in place, while the table is being read concurrently, for a MergeTree table in a database with lazy_load_tables = 1 and for an Alias table whose target is a MergeTree table. #119105 (Groene AI). #121703 (Groene AI). #113634 (Groene AI).
• TRUNCATE ALL TABLES FROM and TRUNCATE TABLES FROM ... LIKE ... now skip Alias tables, which hold no data of their own. Previously such a statement could fail with LOGICAL_ERROR, truncate a target in another database, or fail when the alias target was missing. #121134 (Groene AI).
• Fix SHOW CREATE TABLE, system.tables.create_table_query, and metadata files of Alias tables including a column list, which is invalid for ENGINE = Alias and goes stale when the target table changes. #115141 (Enric Calabuig).
• Fixed a logical error when a non-APPEND refreshable materialized view was created in a non-Atomic/Replicated database via a full-definition ATTACH, or via a RESTORE of a view that had a UUID into such a database. CREATE already rejected this combination; ATTACH and RESTORE bypassed the check and produced a broken view with no UUID. Backups of views that never had a UUID stay restorable, and refresh now fails cleanly instead of aborting in debug and sanitizer builds. #111928 (Groene AI).
• Fixed SELECT count() never terminating, or returning an invented row count, on an Avro file with a corrupted block header. The bundled ClickHouse/avro now rejects a block row count larger than its payload can hold wherever it can derive a lower bound for the schema, so SELECT count() over snappy-compressed Avro files still uses the block header row count instead of decoding every row. #116156 (Groene AI). #121211 (Groene AI).
• Fix rows_read and rows_before_limit_at_least in the statistics of the JSON and XML output formats being reported as 0 or too small when a query with LIMIT reads from remote servers (parallel replicas or the remote table function), because the output format could be finalized before the remaining progress was drained from the connections. Also fix a query broken off by timeout_overflow_mode = 'break' answering with a truncated response - or with no response body at all - instead of a complete document with the statistics, and, over the native protocol, dropping the rows that had already been produced when the time limit was checked. #118000 (Alexey Milovidov).
• Fix CANNOT_SEEK_THROUGH_FILE when reading a subset of columns from an ArrowStream through a pipe, for example with clickhouse local reading stdin. #125689 (Mishalski).
• Fixed FileLog on macOS re-reading a file from offset 0 when it is renamed during a directory rescan. #122448 (Groene AI).
• Fixed FileLog silently skipping the first rows of a file moved into the watched directory from elsewhere (for example a staging file renamed into place) when that file got the inode of a file just deleted from the directory. Such a file is now read from the beginning. #123593 (Groene AI).
• A FileLog file that cannot be opened (removed, a symlink whose target was removed, not readable by the server) no longer stops the table: the other files keep being consumed, and it is skipped, logged once and retried until it can be opened while the table is loaded, unless the directory reports that it was removed, renamed or replaced. With the default handle_error_mode, a record that cannot be parsed is now logged and skipped while the table streams into materialized views, instead of failing the whole insert together with the rows read from the other files; a direct SELECT still throws. #123595 (Groene AI).
• Fixed the warning about a setting duplicated in the definition of an S3Queue or AzureQueue table, which said that the first value is used while the table uses the last one. #124916 (Vadim Lopatin).
• Fix the KafkaConsumersWithAssignment metric drifting negative: it was decremented twice for every consumer-group rebalance. #113919 (kalavt).
• Fixed the histogram metric keeper_read_wait_for_write_time_milliseconds and the OpenTelemetry span keeper.read.wait_for_write always reading zero and never being emitted when use_new_dispatcher is enabled, which is the default. #115252 (Groene AI).
• TestKeeper no longer treats nodes that only share a name prefix as descendants: removeRecursive of /q keeps /q2, and listRecursive and listWithOptions of /q return its children when a sibling such as /q-x exists. #122769 (Groene AI).
• Fixed Keeper ProfileEvents KeeperPacketsSent, KeeperPacketsReceived, KeeperRequestTotal, KeeperRequestTotalWithSubrequests, and KeeperLatency being counted twice. After upgrading, these counters grow at half the previous rate. #122811 (myeongjun).
• Fix unbounded recursion when decoding a deeply nested value inside a binary-encoded type, for example an AggregateFunction parameter in a Dynamic value read from RowBinary or Native. #122667 (Pavel Kruglov).
• Fixed SHOW COLUMNS failing with Cannot parse string 'Decimal' as Int8 (CANNOT_PARSE_TEXT) for Nullable(Decimal(P, S)) columns when the connection is made through the MySQL wire protocol. #75694 (ultrabig).
• Fix a logical error Cannot write to canceled buffer in the PostgreSQL wire protocol handler when a client disconnects while the server is still sending the result of a query. #118369 (Alexey Milovidov).
• Fixed a signed overflow in the workload throttler that made system.scheduler.throttling_us report a negative value for a workload whose max_bytes_per_second, max_queries_per_second, or max_cpus is small enough to postpone a request for longer than a year. Such a postponement is now saturated at one year. #113697 (Groene AI).
• Fixed the throughput column of system.scheduler reporting a small fraction of the real value (about 0.05x initially and 0.3x after 4 seconds) for about 20 seconds after a scheduling node is created, for example right after server start or CREATE WORKLOAD. The client’s CPU usage indicator, which uses the same EventRateMeter, was affected the same way on its first samples. Closes #121962. #122006 (Groene AI).
• Fix a logical error Cannot add step Limit to QueryPlan because it has incompatible header with root step Sorting when a vector search query with vector_search_with_rescoring = 0 was executed with the experimental make_distributed_plan setting. #114457 (Alexey Milovidov).
• Fix assignCentroid(vec, 'dict') with an unqualified dictionary name over a Distributed table, cluster/remote, or parallel replicas. Remote shards and replicas resolved the name in their own default database, so the query failed with Dictionary (dict) not found or silently used a dictionary with the same name from default. The name is now resolved in the current database of the initiator, as for dictGet. #123711 (Groene AI).
• Fixed UNDROP TABLE failing with LOGICAL_ERROR (Failed to get metadata path from database) instead of UNKNOWN_TABLE when the target database engine stores no table metadata on disk, for example a Memory database that reused the name of a dropped Atomic one, or the default database of clickhouse-local. #120945 (Groene AI).
• Fixed ALTER TABLE ... MODIFY SETTING name = DEFAULT skipping the checks an engine makes on a setting reset. #119475 (Raúl Marín).
• Fixed TRUNCATE ALL TABLES FROM <db> failing with NOT_IMPLEMENTED for the whole database when it held a table whose engine stores no rows of its own, such as Merge, Null, GenerateRandom, Executable, ExecutablePool, FuzzQuery, FuzzJSON, or URL. Such tables are now skipped, as views and dictionaries already were. #121248 (Groene AI).
• A table created without a column list over a source that carries a per-column TTL no longer inherits that TTL when its own engine does not support one, such as Merge, Buffer, Distributed, or Remote. #121335 (Groene AI).
• Fixed the logical error Cannot find column <name> in ColumnsDescription when a single ALTER statement on a materialized view applies COMMENT COLUMN or MODIFY COLUMN ... COMMENT after a MODIFY QUERY whose new query no longer produces that column. Such a statement now reports the missing column instead of aborting in debug and sanitizer builds. #121469 (Groene AI).
• Fix LOGICAL_ERROR Cannot capture column N because it has incompatible type when a materialized view with a lambda over an Array(Tuple(...)) column receives a block while the source table’s column type is being changed by ALTER TABLE ... MODIFY COLUMN. #122345 (Jimmy Aguilar Mena). #123514 (Jimmy Aguilar Mena). #124740 (Groene AI).
• In a Memory database, the name of a detached table, view or dictionary is now reserved, as in Atomic databases. A statement that would create a new object under it (CREATE, CREATE OR REPLACE [MATERIALIZED] VIEW, or ATTACH with a full definition) fails with TABLE_ALREADY_EXISTS, IF NOT EXISTS makes it a no-op, and a short ATTACH restores the detached object. Previously these statements were accepted, but SHOW CREATE TABLE and the table dependencies kept describing the detached object, and a later ALTER of such a view failed with a LOGICAL_ERROR exception. #122670 (Groene AI).
• Fixed a window definition that places a frame before the parent window name, as in WINDOW x AS (ROWS UNBOUNDED PRECEDING w). That order is now a syntax error, while a parent window named after a frame keyword, as in WINDOW x AS (rows), now parses. #120156 (Groene AI).
• Formatting EXPLAIN AST (<query>) FORMAT ... or ... SETTINGS ... now keeps the parentheses for INSERT, DELETE, CREATE, a nested non-SELECT EXPLAIN, other non-SELECT queries, and a SELECT or EXPLAIN with output options of its own, and EXPLAIN SYNTAX SYSTEM ... FORMAT ... is no longer formatted with parentheses that do not parse, so the formatted query parses back with the same output options. #123189 (Groene AI).
• Fixed formatting of PostgreSQL COPY queries so table names, columns, direction, format, and header options are preserved. #123601 (Minh Vu).
• Keep empty LIKE and ILIKE patterns in SHOW statements, including SHOW FUNCTIONS. #123624 (Minh Vu). #123626 (Minh Vu).
• A one-argument tuple in the ORDER BY of a projection, such as ORDER BY tuple(c0), is now kept when the definition is formatted, so the formatted query parses back to the same definition. #123162 (Alexey Milovidov).
• Fixed formatting of the tuple element access operator when its index carries an alias. SELECT ((1,2) AS t).(1 AS a) was formatted as SELECT ((1, 2) AS t).(1) AS a, which re-parses with the alias attached to the whole tupleElement instead of to the index. This changed the column names of a query passed through formatQuery or sent to a remote shard, and raised Inconsistent AST formatting on builds with assertions enabled. #120217 (Groene AI).
• Fixed formatQuery and formatQueryFromJSON dropping FULL from SHOW FULL DATABASES and the other supported SHOW FULL forms. SHOW FULL DICTIONARIES no longer selects a non-existent engine column, and builds with assertions enabled no longer fail with Inconsistent AST formatting. Closes #123967. #124036 (Groene AI).
• Fixed a crash and a LOGICAL_ERROR when a lambda function reached query analysis without the lambda(tuple(...), body) shape the SQL parser always produces. An AST deserialized from JSON (dialect = clickhouse_json) can no longer mark an arbitrary function as a lambda, and CREATE FUNCTION f AS g(tuple(x), x) is now rejected with BAD_ARGUMENTS instead of failing when f is called. Closes #120380. #120483 (Groene AI).
• Fixed server termination on CREATE TABLE/ALTER TABLE when a constraint expression restored through the clickhouse_json dialect held a function without an arguments list. Such a payload now fails with BAD_ARGUMENTS while the AST is deserialized, so shapes that used to be accepted and fail later at INSERT (for example notEquals with no argument list) are rejected at parse time inside that dialect. #120507 (Groene AI).
• Fixed formatQueryFromJSON accepting enumerator names that name no value (an enumeration bound, or an unset default), such as {"type":"SystemQuery","query_type":"UNKNOWN"}, which raised LOGICAL_ERROR: Unknown SYSTEM command instead of reporting bad input. Such names are now rejected with BAD_ARGUMENTS. Closes #120741. #120886 (Groene AI).
• Fixed two segfaults in formatQueryFromJSON when a node omits a child that formatting dereferences unconditionally. Closes #120739. #120876 (Groene AI).
• Fixed a server crash when a clickhouse_json payload restores an expression holding a function with no arguments list, a shape the SQL parser cannot produce. Such a payload is now rejected with BAD_ARGUMENTS, as is an empty TTL list, which was persisted as metadata that no longer loaded. #120882 (Groene AI).
• Fixed formatQueryFromJSON raising LOGICAL_ERROR: Unknown element in AST: Foreign Key for a payload of type ForeignKeyDeclaration. That node type has no formatter and no query can produce one, so it is now rejected with BAD_ARGUMENTS. Closes #120738. #120915 (Groene AI).
• Fixed stack exhaustion when deserializing an AST from JSON through formatQueryFromJSON or the clickhouse_json dialect. Such payloads now fail with TOO_DEEP_RECURSION. #121693 (Groene AI).
• Fixed a silently corrupted query when a node restored from the AST JSON format (formatQueryFromJSON, dialect = 'clickhouse_json') carried a shape the parser cannot produce, such as a bare select query argument or an expression list where an argument was expected. Such a node is now rejected with BAD_ARGUMENTS instead of formatting into SQL that parses back as a different query. #120725 (Groene AI).
• Fixed stack exhaustion in parseQueryToJSON when serializing a query nested deeper than the serializer could follow, reachable by raising max_parser_depth above its default. Such a query now fails with TOO_DEEP_RECURSION instead of terminating the server. #121856 (Groene AI).
• A CREATE DICTIONARY sent in the AST JSON format (dialect = 'clickhouse_json') is now rejected with BAD_ARGUMENTS when the brackets of a key-value argument do not match its value, a bracketed list holds something other than key-value pairs, the SOURCE has no brackets, the primary key is empty or not a list of identifiers, a DEFAULT is not a literal, or the attributes or the definition are missing or empty, and so is any other query that carries them. Such queries could crash the server or prevent it from starting. #122981 (Groene AI).
• Fix a server crash (SIGSEGV) when a CREATE TABLE ... PROJECTION / ALTER TABLE ... ADD PROJECTION statement is restored from a clickhouse_json AST whose projection query subtree contains a function node without an arguments list. The malformed payload is now rejected with BAD_ARGUMENTS during AST JSON deserialization. #121113 (Ilias).
• A clickhouse_json payload that restores a function without an arguments list into an expression slot read by the pre-analyzer stack is now rejected with BAD_ARGUMENTS while it is being deserialized. Previously such a payload was accepted and could reach a visitor that dereferences the missing list, terminating the server. #121119 (Groene AI).
• Fixed parseSize and parseTime logic to correctly distinguish between size thresholds and time-based rotation schedules such as 12:00. Fixes #103108 and #100910. #103127 (Jianmei Zhang).
• A configuration file can be written in XML or in YAML, but only config.xml was looked up as the default configuration file. Now clickhouse-server, clickhouse-keeper, clickhouse-local, clickhouse-keeper-client and clickhouse-disks also find config.yaml and config.yml (keeper_config.yaml and keeper_config.yml for Keeper). Previously, clickhouse-server silently started with the configuration embedded into the binary when only a config.yaml was present. #118422 (Alexey Milovidov).
• clickhouse install (and therefore apt/dpkg package upgrades) no longer prompts interactively to set up a password for the default user when that user has been explicitly removed from configuration via <default remove="1" /> in users.d. Previously the installer only checked whether the default user already had a password configured, missing the case where the user itself does not exist in the merged configuration at all, so it kept asking for a password on every upgrade even though there was no user to set it for. #110438 (Eduardo Gómez).
• Fixed the per-query MemoryAllocatedWithoutCheck and MemoryAllocatedWithoutCheckBytes values in system.query_log. A thread’s untracked memory balance is flushed when the thread is attached to a query, and that flush was reported as the query’s, so memory allocated before the query existed was charged to it. #116505 (Groene AI).
• Fixed AsynchronousInsertQueueSize and AsynchronousInsertQueueBytes in system.metrics growing permanently after SYSTEM FLUSH ASYNC INSERT QUEUE. #120215 (Sema Checherinda).
• Fixed a trailing NUL byte in the notation column of system.unicode, which made length return a wrong value and comparisons such as notation = 'U+0041' never match. #122119 (David Meng).
• Fixed CGroup* asynchronous metrics being silently missing on cgroup v2 systems that run ClickHouse in a nested cgroup, such as a typical systemd service. Closes #103968. #122918 (Raúl Marín).
• Fixed NO_COMMON_TYPE for use_variant_as_common_type = 0 in the jemalloc fragmentation dashboard. #121621 (Azat Khuzhin).
• Fixed TYPE_MISMATCH being counted in system.errors on every execution of a SELECT comparing a DateTime64 column against a DateTime, Date or Date32 constant inside AND. A date or time constant with a part the column type cannot hold is no longer pushed down to MongoDB as a truncated bound. #121078 (Navneet Kumar).
• Fixed reading system.parts, system.parts_columns and system.projection_parts for a table whose PARTITION BY contains a modulo with an unsigned left and a signed right operand: the query aborted with Bad get: has Int128, requested UInt128 (BAD_GET) when the partition value was wider than 8 bytes, and otherwise showed a reinterpreted value that ALTER TABLE ... DROP PARTITION rejects. The same value is rendered while a merge or mutation is scheduled, while a system.part_log row is written, and while a partition is pruned, so on such a table merges failed to start, part_log rows were lost, and a filtered read could abort during planning. #119333 (Groene AI).
• Fix system.server_settings reporting the async logger queue size under the misspelled key logger.async_queye_max_size instead of logger.async_queue_max_size. #123578 (Raúl Marín).
• Fix SELECT from system.stack_trace failing with Cannot read from file /proc/<tid>/status ... No such process when a thread exits while the table is being scanned. #123805 (Groene AI).
• Fix system.jemalloc_sampled_allocations and the collapsed output of system.jemalloc_profile_text on sampled allocations whose backtrace jemalloc could not unwind: the allocation is now reported with trace = [] or under [unknown] instead of failing with CANNOT_PARSE_TEXT or being omitted. #123818 (Groene AI).
• Fix per-protocol asynchronous metrics such as TCPThreads and HTTPThreads not being reported for listeners configured through the <protocols> section. Such a listener is now reported under the same metric as the equivalent built-in port. Metrics are also now reported for the tcp_with_proxy and tcp_ssh listeners, which previously had none. #112977 (Chester).
• Fix a deadlock when an obsolete union system log table is replaced with synchronous drops. #124610 (Azat Khuzhin).
• Fixed the max_local_read_bandwidth_for_server, max_local_write_bandwidth_for_server, max_remote_read_network_bandwidth_for_server and max_remote_write_network_bandwidth_for_server rows of system.server_settings reporting the reading session’s own bandwidth limit instead of the server-wide one. getServerSetting on those names is fixed the same way. #119556 (Groene AI).
• Fix tables of the clickhouse-local default database becoming invisible (UNKNOWN_TABLE) when the same --path is later used with another default_database. #123124 (Nikolay Degterinsky).
• Fixed missing crash reports of clickhouse-local when neither stdin nor stderr is a terminal: a fatal signal printed nothing, and --client_logs_file was not written. #115102 (Groene AI).
• Fixed clickhouse-local hanging forever on DROP TABLE|VIEW|DATABASE ... SYNC when --path is set and no metadata is loaded from it, for example a new directory on released versions or --only-system-tables. #123862 (Groene AI).
• The wasm64 clickhouse local now starts with 2 GiB of linear memory, so it no longer fails at startup with memory access out of bounds when Node.js races a memory.grow across worker threads. Closes #124067. #123944 (Groene AI).
• Fixed SET profile having no effect on subsequent queries in clickhouse-local. Profile settings such as readonly and join_use_nulls now apply to subsequent queries again. Closes #123916. #125171 (Alexey Milovidov).
• Fixed promql_evaluation_time silently rounding the requested instant, which could make an instant query return a different sample. The setting also now accepts an integer Unix timestamp over the native protocol, which it previously refused. #120942 (Valery Petrov).
• PromQL queries with regex label matchers (=~, !~) again work under --dialect promql. #115073 (Valery Petrov).
• Allow PromQL # and #! line comments to end at EOF without a trailing newline. #113201 (Minh Vu).
• Fix PromQL duplicate-series detection to check one step at a time, so series that share tags but have values at different steps are now merged instead of failing with Multiple series have the same tags. Binary operators check for duplicates after matching series, so unmatched series no longer cause this exception. #124086 (Vitaly Baranov). #124879 (Vitaly Baranov).
• PromQL: fixed expressions over an instant vector that is known to be empty, such as (0 < clamp(up, 1, -1)) or vector(7), which failed in range queries with Number of values doesn't match number of steps or returned samples at 1970-01-01. #123879 (Valery Petrov).
• Fixed SELECT count() returning the whole-table row count for a streaming read (FROM ... STREAM), ignoring the CURSOR bound. The implicit min-max/count projection answered the query from part metadata instead of the streaming reader. #117188 (Groene AI).
• Fixed a server crash (std::terminate with write failed) in the embedded SSH client when the client’s terminal was closed before the line reader was destroyed. #120764 (Alexey Milovidov).
• Fixed Code: 1001 STD_EXCEPTION (filesystem error: ... File name too long) when a query names a table or dictionary longer than the filesystem’s NAME_MAX in clickhouse local, or in a server holding a Filesystem or URL database. Such a name is now reported as missing rather than as a filesystem error. Closes #121304. #121471 (Groene AI).
• A malformed argument list of an optimization of the mergeTreeAnalyzeIndexes table function is now reported as BAD_ARGUMENTS instead of a logical error or a raw std::exception. #118644 (Alexey Milovidov).
• Fixed max_execution_time not being enforced in clickhouse-local for a query blocked inside a non-interruptible read. clickhouse-local did not start the watchdog that turns a time limit expiry into a query cancellation, so a query reading from an unresponsive s3 endpoint ran to the end of the S3 retry budget (s3_retry_attempts defaults to 500, which takes minutes) instead of failing with TIMEOUT_EXCEEDED at its limit. #120933 (Groene AI).
• Fixed distributed queries hanging at 100% CPU on macOS when max_threads = 1. #122905 (Raúl Marín).
• Fixed a server abort on macOS when an HTTP client resets the connection during header reading. #123074 (Groene AI).
• Fixed a LOGICAL_ERROR (Unexpected return type from materialize, a server abort in debug and sanitizer builds) when reading a Buffer table whose destination declares a column with a different but convertible type and a forwarded filter references that column: with a row policy and the implicit WHERE to PREWHERE move, and under query_plan_optimize_lazy_final. The same cause made such a read fail with NOT_FOUND_COLUMN_IN_BLOCK when a text index on that column was read directly and the predicate appeared in both PREWHERE and WHERE, fail with TYPE_MISMATCH over a URL(Parquet) destination, abort with Bad cast from ColumnMap to ColumnArray when a row policy and an explicit PREWHERE both referenced it, or fail with AMBIGUOUS_COLUMN_NAME when several PREWHERE conditions referenced the converted column. #113900 (Groene AI). #124632 (Groene AI).
• Fixed SYSTEM DROP DNS CACHE leaving entries in the DNS cache when it ran at the same time as the periodic DNS cache update. #123699 (Groene AI).
• Fix UNSUPPORTED_METHOD error for queries using IN/NOT IN/GLOBAL IN with a subquery inside AND/OR when enable_function_early_short_circuit is enabled. #116037 (fhw12345).
• Fixed a crash caused by a stack overflow when a deeply nested JSON document was parsed into a JSON column or passed to the JSONExtract family of functions. Such documents are now rejected with the TOO_DEEP_RECURSION error. #120887 (Pavel Kruglov).
• Fixed the LOGICAL_ERROR Not-ready Set is passed as the second argument for function 'in' for a query with ORDER BY ... WITH FILL ... INTERPOLATE whose INTERPOLATE expression contains IN with a subquery or a table that is not used anywhere else in the query. #124518 (Groene AI).
• Fixed wrong results of a WHERE equality between Tuple values with Nullable elements, or between Variant values, over a comma, CROSS or INNER join: the pair whose equality is NULL, such as tuple(NULL) = tuple(NULL), was returned. With allow_dynamic_type_in_join_keys = 0, a comma or CROSS join with a WHERE equality on a Dynamic value no longer returns no rows or fails with ILLEGAL_COLUMN. #124854 (Groene AI).
• Fixed secret arguments being exposed in system.query_log and the server log when a table function or function was called with the wrong letter case, such as S3 instead of s3, MySQL instead of mysql, or ENCRYPT instead of encrypt. Such calls are still rejected. #124862 (Groene AI).
• Fixed tables of a MySQL database ignoring the database’s connection_auto_close setting: with connection_auto_close = 0, every read still closed its connection and the next read reconnected. A table attached again with ATTACH TABLE also lost the database’s settings. INSERT into any MySQL table ignored the setting and always kept its connection; with the default connection_auto_close = 1 it now closes it after the query, as reads do. Closes #124749. #124884 (Groene AI).
• Hide the password of a URL in SHOW CREATE TABLE, system.tables and system.query_log only within the URL’s authority: a @ in the query string no longer rewrites the shown host, and a password containing @ is hidden whole. #124885 (Raúl Marín).
• Fixed merge and Merge tables over a Distributed or remote table rejecting valid queries on a column that some of the tables lack: aggregates over Decimal and Map columns failed with errors such as THERE_IS_NO_COLUMN, and an integer column of this kind in GROUP BY, ORDER BY or LIMIT BY failed with Positional argument number 0 is out of bounds. #124976 (Groene AI).
• Fixed trace samples in system.trace_log losing their query_id while a thread attaches to or detaches from a query, so memory profiling samples remain attributable to the query. #125176 (Groene AI).
• Fixed wrong window function results in the totals row of a WITH TOTALS query when an arrayJoin expands the totals into several rows: those rows were not sorted by the window’s PARTITION BY / ORDER BY keys. #125324 (Groene AI).
• Fixed primary key and skip index analysis that could skip or over-count rows for conditions on date and time functions such as toHour, toDayOfMonth, timezoneOffset or toTimeWithFixedDate over a Nullable(DateTime) or LowCardinality(Nullable(DateTime)) key whose time zone differs from the session time zone. #125327 (Groene AI).
• Fixed wrong results of ORDER BY, DISTINCT, LIMIT BY and aggregation in order over a Merge table or the merge table function when it reads an Iceberg table with a sort order: the rows came back unsorted or duplicated. #125383 (Groene AI).
• Fixed count() with a WHERE condition on the sorting key returning too many rows on a MergeTree table with a normal projection: rows counted from the primary key by the exact count optimization were counted again when the projection was read. Closes #125183. #125412 (Groene AI).
• Fixed a possible segmentation fault when reading an Avro map into a Map column (for example Map(DateTime, String)) with the AvroConfluent format, including Kafka tables. Closes #125457. #125505 (Groene AI).
• Fix a server crash on CREATE DICTIONARY without a SOURCE clause when restore_replace_external_dictionary_source_to_null is enabled. Such a query now fails with INCORRECT_DICTIONARY_DEFINITION, as it does with the setting disabled. With the same setting, a local CLICKHOUSE source that uses a constant function such as DB currentDatabase() is no longer rejected. #125659 (Groene AI).
• Fixed the Arrow and ArrowStream input formats rejecting valid dictionary-encoded columns whose dictionary contains repeated values. Closes #125139. #125677 (Jay Pokale).
• Fixed the Arrow Flight interface writing the text of each query, including secrets such as IDENTIFIED BY passwords and storage keys, to the server log and system.text_log. The gRPC interface did the same for the first 64 characters of each query. Now queries from these interfaces are logged only with their secrets hidden. #125716 (Groene AI).
• Fix a server crash on macOS where an idle thread pool worker aborted with condition_variable wait failed: Invalid argument. The EINVAL that Darwin’s pthread_cond_wait returns when the kernel fails the wait is now treated as a spurious wakeup. Closes #84761. #125721 (Groene AI).

​

Build/Testing/Packaging Improvement

• Harden official builds by compiling with -fstack-protector-strong and -fstack-clash-protection, and by linking with -Wl,-z,separate-code. #119039 (Raúl Marín). #123073 (Raúl Marín).
• The musl build of ClickHouse is now a static position-independent executable, so it gets address space layout randomization like the glibc build. #125669 (Raúl Marín).
• Added an optional packaging mode SYMBOL_SECTIONS=gnu_debugdata that replaces .symtab and .strtab in release binaries with compressed .gnu_debugdata (MiniDebugInfo), reducing symbol data from about 148 MiB to about 10 MiB in the official build. Stack traces and symbol access from gdb, lldb, elfutils, and perf remain available. #115322 (Andrew Grigorev).
• Update cctz to 2026d. Time zone names with the file: prefix are now rejected. #121325 (Actuele AI). Update cctz to 2026e: America/Winnipeg, America/Rainy_River, and Canada/Central now stay at UTC−5 from 2026-11-01, reflecting Manitoba’s removal of the DST fall-back. #124249 (Actuele AI).
• Update OpenSSL to 3.5.9. #124278 (Actuele AI).
• Update aws-sdk-cpp to 1.11.907, the AWS CRT libraries to 0.43.8, and aws-c-common to 1.0.2. #125080 (Groene AI).
• The delta-kernel-rs library no longer carries its own copy of OpenSSL, so the binary always links the OpenSSL built from contrib/openssl with the current build flags instead of an older copy kept by the Rust build cache. #123889 (Groene AI).
• Rust crates are now built in parallel with each other and with the C++ code, which reduces the build time. #124645 (Antonio Andelic).
• Builds of the experimental GPU engine are faster: nvcc compilations are now cached with ccache, and unused PTX plus CUDA line info are no longer generated. #120421 (Konstantin Vedernikov).
• Fixed building with -DNO_ARMV81_OR_HIGHER=1 on AArch64: the compatibility profile no longer disables NEON for Rust, so ring-based components (chdig, delta-kernel-rs) compile and stay enabled. This unblocks building for ARMv8.2-A CPUs that lack rcpc, such as HiSilicon Kunpeng 920. #107440 (Tiaonmmn).
• Compile code linked into the position-independent clickhouse executable with -fPIE instead of -fPIC, avoiding unnecessary symbol indirection. Apply the flag to the C++ runtime libraries from contrib too, so builds with compilers that do not default to PIE link successfully, including Fedora’s clang on aarch64 and x86_64. #125094 (oleg-milanov). #125488 (Groene AI).
• Allow building with XRay instrumentation under musl. #118914 (Konstantin Bogdanov).
• Fixed a missing CMake link dependency that could make src/Functions/array/distanceTransposed.cpp fail with simsimd/simsimd.h file not found on a from-scratch build. #120504 (Craig Ringer).
• Reduce linker warnings in Rust/wasm builds: remove --icf=safe warnings about SHT_LLVM_ADDRSIG in the stripped Rust libraries, and apply the __jit_debug_register_code workaround only to the clickhouse binary and only when wasmtime is linked. #122904 (Raúl Marín). #123287 (Azat Khuzhin).
• Update google-protobuf to v36.1. google-cloud-cpp protobuf sources are now regenerated when protoc changes, so incremental builds do not fail after such an update. #118926 (Actuele AI). #123603 (Alexey Milovidov).
• Update rocksdb to v11.8.1. #118929 (Actuele AI).
• Update abseil-cpp to 20260817.0. #118897 (Actuele AI).
• Update PostgreSQL to 18.6. #124281 (Actuele AI).
• Update thrift to v0.25.0. #124280 (Actuele AI).
• Update cppkafka to v0.4.1. #119468 (Actuele AI).
• Update h3 to v4.5.0. #119328 (Actuele AI).
• Update flatbuffers to v25.12.19. #119100 (Actuele AI).
• Update croaring to v5.2.2. #119464 (Actuele AI).
• Update simdutf to v9.1.1. #119324 (Actuele AI).
• Update double-conversion to v3.4.0. #119095 (Actuele AI).
• Update libcpuid to v0.8.2. #121382 (Actuele AI).
• Update liburing to 2.15. #118895 (Actuele AI).
• Update libuv to v1.52.1. #119099 (Actuele AI).
• Update libcotp to v4.2.2. #122000 (Actuele AI).
• Update vectorscan to 5.4.13. #125774 (Actuele AI).
• Update snappy to 1.2.2. #119325 (Actuele AI).
• Update brotli to v1.2.0. #119096 (Actuele AI).
• Update libdeflate to v1.26. #119349 (Actuele AI).
• Update xxHash to v0.8.4. #124231 (Actuele AI).
• Update google-benchmark to v1.9.5. #118893 (Actuele AI).
• MemoryWorker now corrects MemoryTracker using allocated memory from ASan, TSan, and MSan instead of resident memory in sanitizer builds when cgroups are available. #119956 (Konstantin Morozov).
• Add support for cross-compiling ClickHouse for illumos. Closes #23777. #125073 (Alexey Milovidov).

​

ClickHouse release 26.9, 2026-09-21. Presentation, Video

​

Backward Incompatible Change

• The analyzer can no longer be disabled: the enable_analyzer setting (and its old name allow_experimental_analyzer) is obsolete, an attempt to set it to 0 is rejected, and the compatibility setting no longer reverts it. The analyzer has been the default since 24.3, and the query analysis used before it is no longer supported. To compare the behaviour or the performance of a query with the old query analysis, use a ClickHouse version older than 26.9. #118629 (Alexey Milovidov).
• The default compression method is changed from LZ4 to ZSTD(3). Client/server and server/server network communication switches to ZSTD(3) uniformly. For MergeTree column data the built-in default is now size-aware: LZ4 for parts below 100 MB and ZSTD(3) at or above it, judged by the part size known at write time (the size of the source parts for a merge or a mutation). The StripeLog data stream, the persistent files of the Set and Join engines, and MergeTree auxiliary files such as checksums.txt switch to ZSTD(3) uniformly. HTTP compress=1 now uses the network_compression_method setting too; the default is still ZSTD(3), and compatibility set to 26.8 or earlier restores LZ4 over HTTP as well. This improves compression ratios and reduces storage and network usage out of the box, at a modest increase in CPU usage. To restore the previous behavior: for network compression, use the compatibility setting or set network_compression_method = 'LZ4'; for MergeTree column data, set CODEC(LZ4) on the column or table, or set the server <compression> default to LZ4; for the Log and TinyLog engines, set a column CODEC(LZ4). The remaining streams always use the built-in default and have no runtime rollback; reading stays correct because every compressed frame is self-describing. #108786 (Alexey Milovidov). #119825 (Alexey Milovidov). This change is backward compatible, it is listed here for your convenience.
• Removed the CatBoost integration. The catboostEvaluate function, the system.models table and the SYSTEM RELOAD MODEL(S) queries are no longer available. Queries, views and grants referring to them have to be migrated before the upgrade: in particular, revoke SYSTEM RELOAD MODEL from all users and roles beforehand, because an access entity whose grants still mention the removed privilege cannot be parsed on startup. Also let any in-flight SYSTEM RELOAD MODEL(S) ON CLUSTER entries drain from the distributed DDL queue before the upgrade: upgraded hosts can no longer parse such entries and will finish them with an error status instead of executing them. The same applies to persisted table metadata: a table whose column DEFAULT/MATERIALIZED/ALIAS expressions, secondary indices, constraints, projections, partition/order/sample keys, or TTL expressions still reference catboostEvaluate cannot be loaded after the upgrade (UNKNOWN_FUNCTION), so drop or rewrite such definitions beforehand. The revoke also has to cover every other carrier of serialized grants: a GRANT SYSTEM RELOAD MODEL line in the grants section of users.xml must be removed as well, because it can no longer be parsed and the server then refuses to load the users configuration at startup; on clusters with replicated (Keeper-backed) access storage, revoke the privilege before upgrading any replica, because with the default access_control_improvements.throw_on_invalid_replicated_access_entities = 0 an upgraded replica that reads a user or role still carrying SYSTEM RELOAD MODEL from ZooKeeper cannot parse it and silently drops that entity from its in-memory copy (set throw_on_invalid_replicated_access_entities = 1 to fail loudly instead); and access backups (BACKUP ... ACCESS) taken while the grant was still present cannot be restored with RESTORE ... ACCESS on the new version, so re-create such backups after the revoke. #109710 (Robert Schulze). #116586 (Alexey Milovidov).
• Removed the deprecated read_resource and write_resource disk options from storage_configuration. Attach a scheduler resource to a disk with CREATE RESOURCE name (READ DISK disk, WRITE DISK disk) instead. #115286 (Sergei Trifonov).
• Removed the WasmEdge engine for WebAssembly UDFs. wasmtime, which has always been the default, is now the only supported value of the webassembly_udf_engine server setting. A server configured with webassembly_udf_engine = 'wasmedge' will not start; remove that setting or set it to wasmtime. On binaries where wasmtime is unavailable, including MemorySanitizer, arm_v80compat, and riscv64, WebAssembly UDFs are not available at all: the server still starts, but system.webassembly_modules is absent and any attempt to create or use a WebAssembly UDF fails with SUPPORT_IS_DISABLED. #115316 (Alexey Milovidov).
• The experimental WINDOW VIEW feature is removed, together with the WATCH statement and the windowID function. It has been non-functional since version 24.12 (inserts into the source table were silently not delivered to window views) and was never supported by the analyzer, which is enabled by default since 24.3. The settings allow_experimental_window_view, window_view_clean_interval, window_view_heartbeat_interval, and wait_for_window_view_fire_signal_timeout are kept as obsolete no-ops. The time-window functions tumble, tumbleStart, tumbleEnd, hop, hopStart, and hopEnd are kept. If you still have window views defined, drop them before upgrading: a server that finds a window view in its metadata fails to start (Cannot parse definition from metadata file ... (SYNTAX_ERROR)), the same way the removal of the Lazy database engine behaves. Similar functionality will be reimplemented on top of streaming queries (issue #114738). #114747 (Alexey Milovidov).
• Removed the sharded GROUP BY aggregation (enable_sharding_aggregator, disabled by default) in favor of the adaptive aggregation (enable_adaptive_aggregator, enabled by default), which targets the same high-cardinality workloads without the weaknesses on skewed or low-cardinality keys and with support for external aggregation. The enable_sharding_aggregator setting is kept as obsolete: setting it is still accepted and has no effect. #116532 (Nihal Z. Miaji).
• Fixed wrong results from runningConcurrency under LEFT JOIN and ANY JOIN, and under lazy evaluation in if, multiIf, and, and or, by treating it as non-deterministic. Queries using it may be slower. A sorting key, partition key, or text index preprocessor/postprocessor containing runningConcurrency is now rejected, and a table created earlier with one will not load after upgrading: remove or replace it before the upgrade. #114403 (Groene AI).
• BACKUP ... TO Disk(...), RESTORE ... FROM Disk(...), and CREATE DATABASE ... ENGINE = Backup(..., Disk(...)) now require SOURCES grants for the target Disk(...) location: WRITE ON DISK to write and READ ON DISK to read. Previously a user with BACKUP but without SOURCES could still use any disk listed in backups.allowed_disk; such users now need the corresponding DISK grant, while users already granted SOURCES are unaffected. #117406 (Groene AI).
• Compression codecs are now gated by dedicated per-codec settings whose declared tier defines the codec maturity. The ALP codec is promoted to beta through enable_alp_codec, and allow_experimental_codecs is now obsolete: it no longer enables any codec, so use the per-codec enable_<codec_name>_codec settings instead. #116337 (Raufs Dunamalijevs). #117545 (Raufs Dunamalijevs).
• Reading a File source with rename_files_after_processing set now requires the WRITE ON FILE grant on top of READ ON FILE (on the initiator for fileCluster): the setting renames the files a SELECT has read, which is a write to the source, and it used to happen for anyone who could read them. A query relying on it now needs GRANT WRITE ON FILE. A per-query renaming rule is also no longer cached by a Filesystem database, where it could make an unrelated later query rename files, or be dropped silently. #117895 (Groene AI).
• The interface and http_method columns of system.query_log, system.query_thread_log and system.processes are now Enum8 instead of UInt8, over the value set system.session_log.interface already uses, plus Unknown for an unrecognized interface. Comparing to a number keeps working (WHERE interface = 1) and comparing to a name now works too (WHERE interface = 'TCP'), but a plain SELECT renders the name, and arithmetic on the column (interface + 0) is no longer accepted. The default-enabled system.user_query_log and the opt-in all_ union tables copy these types. As with any system log schema change, the existing log tables are renamed with the first free suffix (query_log_0, query_log_1, …) on upgrade; their rows read as names after ALTER TABLE system.query_log_0 MODIFY COLUMN interface Enum8(...), and the same for http_method, preceded by MODIFY SETTING table_readonly = 0 if that table was frozen. #118136 (Groene AI).
• CREATE TABLE ... AS mergeTreeIndex(...) is no longer accepted, nor the same form over mergeTreeProjection, mergeTreeAnalyzeIndexes, mergeTreeAnalyzeIndexesUUID, mergeTreeTextIndex and mergeTreeCodecBlockCounts. The storage these functions return holds its source table’s storage object, so once such a persisted table had been read, DROP TABLE <source> SYNC never returned, the source stayed in system.dropped_tables forever and its data was never reclaimed from disk. Reading these functions directly, as in SELECT ... FROM mergeTreeIndex(...), is unaffected. An existing table of this shape keeps loading at server startup, but ATTACH TABLE, RESTORE and Replicated database replica recovery now refuse it: list them with SELECT database, name FROM system.tables WHERE engine = 'Proxy', drop the ones over these six functions, and read the functions directly instead. #118969 (Groene AI).
• CREATE TABLE ... AS timeSeriesSamples(...) is no longer accepted, nor the same form over timeSeriesData, timeSeriesTags, or timeSeriesMetrics. Reading those functions directly is unaffected. Existing tables of this shape still load at server startup, but ATTACH TABLE, RESTORE, and Replicated replica recovery now refuse them, because once such a persisted proxy table had been read, DROP TABLE <ts> dropped its live target and destroyed its rows, and later reads could fail with QueryPlan was not initialized. Find candidates with SELECT database, name, create_table_query FROM system.tables WHERE engine = 'Proxy' and keep only the ones whose outer call is one of these four functions; a view(...) wrapper is still allowed. Drop such a table on a freshly started server, before anything reads it, so the drop removes only the proxy and leaves the TimeSeries data intact. Do not use DETACH TABLE instead, because a detached definition can no longer be ATTACHed. If metadata loading of such a table already fails after a restart, move the metadata file out of store/ and restart: until then, unfiltered reads of system.tables can fail with ASYNC_LOAD_WAIT_FAILED. #119786 (Groene AI).
• Removed the s3_disable_checksum setting. It existed to skip a second read of the source data for the checksum of an S3 upload; the checksum is now computed during the single read, so there is nothing left to disable. #119496 (Azat Khuzhin).
• Nullable(Tuple(...)) is now generally available and enabled by default. Nullable(Tuple(...)) columns can be created without any setting, schema inference of nullable structs in the Parquet, Arrow, ORC, Avro and JSON formats returns Nullable(Tuple(...)), and a Tuple subcolumn extracted from a Variant, Dynamic, JSON or Nullable(Tuple) column is Nullable(Tuple(...)) and NULL where the value is missing, instead of a tuple of default values. The previous behavior is available with enable_nullable_tuple_type = 0 and allow_nullable_tuple_in_extracted_subcolumns = 0; the subcolumn setting is read at server startup, so it has to be set in the default profile. Upgrading existing tables whose partition keys, sorting keys, TTL expressions, or skip indexes depend on a tuple subcolumn extracted from a Dynamic, Variant or Nullable(Tuple) column or from a JSON path not declared as Tuple(...) is unsafe with the new default: tables may fail to attach, queries may fail, or data parts may be detached during startup. Before the first startup after upgrading, set allow_nullable_tuple_in_extracted_subcolumns = 0 in the default profile, or preserve the previous behavior through compatibility in that profile. Changing the setting after parts have been detached does not restore them automatically; those parts must also be reattached. #119222 (Nihal Z. Miaji).
• New setting validate_group_by_all_key_types (default true) gates the key-type validation that GROUP BY ALL applies to the grouping keys it expands the SELECT expressions into. Since 26.7 it rejects a Variant or Dynamic grouping key, for example an untyped JSON subpath, that earlier versions accepted with the analyzer enabled (the default), with no way to keep such a query running across an upgrade. Setting compatibility to a version before 26.7, or validate_group_by_all_key_types = 0, restores the earlier behavior. An explicit GROUP BY is unaffected and still rejects such a key, as it did before 26.7. Closes #119852. #119868 (Groene AI).

​

New Feature

• DISTINCT can now spill data to disk, like external aggregation and external sort: the new settings max_bytes_before_external_distinct and max_bytes_ratio_before_external_distinct set the threshold; by default, spilling starts once query memory usage exceeds half the available memory under the applicable server or user memory limits, or as soon as the next hash table allocation would exceed it. #116569 (Nihal Z. Miaji).
• Refreshable materialized views can now refresh incrementally with REFRESH ... APPEND INCREMENTAL, appending only the rows committed to the source table since the previous refresh instead of re-reading the whole source. #114152 (Smita Kulkarni). An incremental refresh into an Iceberg target is now exactly-once: the refresh cursor is committed atomically inside the Iceberg snapshot summary. #114823 (Smita Kulkarni).
• New LIMIT forms that select rows by boundary conditions instead of by position, following the stream order (use ORDER BY to define it): LIMIT n AFTER cond returns up to n rows starting at the first row where cond is true (SELECT number FROM numbers(10) ORDER BY number LIMIT 3 AFTER number >= 5 returns 5, 6, 7); LIMIT UNTIL cond returns the rows before the first row where cond is true (LIMIT UNTIL number >= 3 returns 0, 1, 2); LIMIT AFTER cond1 UNTIL cond2 returns the rows between the two (LIMIT AFTER number >= 2 UNTIL number >= 6 returns 2, 3, 4, 5); LIMIT n AFTER cond ALL opens a range of n rows at every matching row and returns their union without duplicates (LIMIT 2 AFTER number IN (2, 6) ALL returns 2, 3, 6, 7). #99508 (Zakhar Kravchuk).
• Added the CREATE TOKEN statement, which generates a random secret, adds it to the current user as an additional authentication method and returns it together with its expiration deadline: CREATE TOKEN VALID FOR INTERVAL 30 DAY GRANTS (SELECT ON db.table). It is a convenient way to create tokens for applications: a token is tied to the user, can be limited in time and in privileges, and stops working when the user is deleted. Without an explicit VALID UNTIL or VALID FOR clause a token lives for create_token_default_ttl_seconds, 30 minutes by default. The statement requires the new CREATE TOKEN privilege, which also authorizes the equivalent ALTER USER <current user> ADD IDENTIFIED ... statement. #116957 (Alexey Milovidov).
• Authentication methods in CREATE USER and ALTER USER ... ADD IDENTIFIED support a GRANTS (SELECT ON db.table, ...) clause which limits the access rights of sessions authenticated with that method to the intersection with the listed grants. This allows using additional credentials as tokens for applications: ALTER USER vasya ADD IDENTIFIED WITH password BY '...' VALID UNTIL '2026-12-31' GRANTS (SELECT ON db.table). Closes #109117. #110144 (Alexey Milovidov).
• Added the SQL standard linear regression aggregate functions regr_slope, regr_intercept, regr_r2, regr_count, regr_avgx, regr_avgy, regr_sxx, regr_syy and regr_sxy. #119347 (mosya415).
• Added MergeTree settings max_table_size_rows, max_table_size_bytes_compressed, and max_table_size_bytes_uncompressed, which limit the total size of a table. The limits are checked at the start of INSERT and when new parts are committed, including the results of background merges and mutations, but not on replicated fetches. Closes #115063. #115069 (Alexey Milovidov).
• Added the max_tables database setting to limit the number of tables in Atomic and Ordinary databases. Every table-like object counts toward the limit, including tables, views, and dictionaries. #115202 (Alexey Milovidov).
• Added bracket syntax for JSON subcolumns, such as json['key1']['key2']; the parser translates it to nested arrayElement calls. #107039 (Pavel Kruglov).
• Added the system.statements system table, which exposes the documentation of SQL statements such as their name, syntax, description, and examples. #115341 (Robert Schulze).
• Added virtual columns for the Iceberg v3 manifest fields first_row_id and last_seq_num. #115603 (Konstantin Vedernikov). The same fields are now also written when inserting into an Iceberg v3 table. #116171 (Konstantin Vedernikov).
• Added the opt-in MergeTree setting skip_empty_columns_on_insert, which avoids storing columns that contain only their type-default value. It requires serialization_info_version = 'with_missing_columns'; during rolling upgrades with servers that cannot read the new part metadata, keep the older serialization version. #98472 (Amos Bird).
• Added support for + and - between DateTime and Time data types. #82980 (Yarik Briukhovetskyi).
• Added the parseISO8601Duration function, which parses ISO 8601 duration strings such as PT1H30M into seconds. #116993 (Navneet Kumar).
• Added the type_json_skip_null_typed_paths setting for JSON. When it is enabled, typed paths in JSON columns that have NULL values are treated as absent, matching the behavior of dynamic paths. Closes #99191. #106959 (Pavel Kruglov).
• Added the arrayFlattenedLength function, which returns the total number of elements in a multidimensional array as if it was flattened with arrayFlatten. Unlike cardinality and length, which count only the elements of the outermost array, it counts elements at every nesting level, matching PostgreSQL’s cardinality. #113696 (David Meng).
• clickhouse-client now accepts --ssh-key-file without a value: the key is looked up the same way as ssh does it, from ~/.ssh/config, the default identity files such as ~/.ssh/id_ed25519, and the keys held by ssh-agent. Authentication with a key held by ssh-agent is now supported without passing its path or passphrase explicitly. #115072 (Alexey Milovidov).
• The splitByRegexp tokenizer, used in text indexes and the tokens function, now accepts an optional extract argument. With splitByRegexp(re, 1), re no longer acts as a separator: each match contributes at most one token, the first capture group of the match or the whole match if re has no capture group. #116127 (Jimmy Aguilar Mena).
• Added the system.session_query_ids system table, which contains the query ids of the queries executed in the current session, in execution order. It makes it easy to find “the queries I just ran” in system.query_log without assigning query_id client-side or tagging queries with log_comment, which is especially useful for testing and benchmarking. The history size is bounded per session by the new session_query_ids_history_size setting (default 1000, 0 disables recording), and TRUNCATE TABLE system.session_query_ids clears the current session’s history. #110536 (Vladimir Cherkasov).
• Added a zero-argument form of year that returns the current year, mirroring now and today: year() is equivalent to toYear(today()). year is now a function in its own right rather than a case-insensitive alias of toYear, and is reported non-deterministic, so year(<date>) is excluded from features that require a deterministic function: the query result cache, mutations of Replicated* tables, on-the-fly mutation application (apply_mutations_on_fly), mutation partition pruning, and the trivial-view pushdown to Distributed tables; use toYear(<date>) where determinism is required. Results, result types, index analysis and projection selection for year(<date>) are unchanged. #109097 (Groene AI).
• Added a memory fragmentation profiler for jemalloc. The new system.jemalloc_sampled_allocations table lists the sampled allocations that are currently alive, with their backtrace, age and size class; together with system.jemalloc_arena_bins it shows which code holds the allocations that keep fragmented slabs alive, and the same data is shown in a new “Fragmentation” tab of the jemalloc page of the server’s web UI. #115456 (Azat Khuzhin).
• Support BACKUP and RESTORE for WORKLOAD and RESOURCE entities. SQL-defined CREATE WORKLOAD and CREATE RESOURCE definitions are now included when backing up system.workloads and system.resources and are recreated on restore, including ON CLUSTER. #108431 (Sergei Trifonov).
• Added the workload_admission_timeout_ms query setting that bounds how long a query waits to acquire its workload query slot and memory reservation before failing. 0 (the default) keeps the previous unbounded wait. #118852 (Sergei Trifonov).
• Added mode = 'exclusive' to the S3Queue engine: file processing is tracked only in the memory of the server, with no coordination through Keeper, for high-throughput ingestion on a single server, such as from a local minio instance. #110552 (Ivan Tkatchev).
• Added the nats_ca_file, nats_client_cert_file and nats_client_key_file settings to the NATS engine, for verifying a broker certificate signed by a private CA and for client-certificate authentication. The feature was first proposed in #69396 by sidhmads. #116117 (Nikolay Degterinsky).
• Added a four-letter-word command to clickhouse-keeper for the backpressure that the leader applies to slow Raft members. #117728 (Kseniia Sumarokova).
• Added the keyValuePairs text index tokenizer, which indexes a Map(String, String) column directly and answers map['key'] = 'value' from the index without reading the map column. #115579 (Jimmy Aguilar Mena).

​

Experimental Feature

• Added Prometheus HTTP API support for the /api/v1/metadata, /api/v1/labels, and /api/v1/label/<name>/values endpoints on TimeSeries tables. The /api/v1/labels endpoint returns the sorted unique label names of the matched time series, and /api/v1/label/<name>/values returns the sorted unique values of one label; both support the optional match[], start, end, and limit parameters. #116518 (Nikita Mikhaylov). #116731 (Nikita Mikhaylov). #117160 (Nikita Mikhaylov).
• Added the /api/v1/format_query endpoint of the Prometheus HTTP API. It parses a PromQL expression and returns its canonical form. #117159 (Nikita Mikhaylov).
• Added an embedded SQL Console UI at /ui on the HTTP server. This is the same console that is available in ClickHouse Cloud - not to be confused with the Web UI. #100406 (Luis Neves).
• Reimplemented the experimental KQL (Kusto) dialect with a dedicated lexer and parser instead of translating it to SQL text and reparsing it. This fixes expression-injection in string operators such as contains and has, several results that disagreed with Kusto such as 7 / 2, substring with a negative start, and bin, and functions that were registered but did nothing. Unsupported syntax is now rejected instead of being mistranslated. #112932 (Alexey Milovidov).
• Added an experimental trino value of the dialect setting (gated by enable_trino_dialect): queries written in Trino SQL are translated to ClickHouse SQL, including Trino-specific syntax (ARRAY[...] literals, TRY_CAST, UNNEST, ROW constructors and types, VALUES tables, OFFSET before LIMIT, FETCH) and a mapping of several hundred Trino function names and argument conventions to their ClickHouse equivalents. Bitwise shifts over untyped integer literals now match Trino semantics. #115383 (Alexey Milovidov). #119430 (Ethan Lin).
• Added the hierarchicalKMeans and assignCentroid functions, building blocks for IVF-style vector search. #112309 (Shankar Iyer).
• Added SELECT support for TimeSeries tables. #115622 (Vitaly Baranov).
• TimeSeries tables now derive their inner engine family from default_table_engine, require all inner tables to use the same replication family, and always pin recent_samples_ttl_seconds into SETTINGS for RECENT SAMPLES targets. #115922 (Vitaly Baranov).
• Added PromQL support for absent and count_values, including range queries and by/without aggregation modifiers. #112795 (Valery Petrov).
• PromQL now supports the range functions sum_over_time, avg_over_time, count_over_time, present_over_time, absent_over_time, quantile_over_time, and predict_linear. #112353 (Valery Petrov). #112842 (Valery Petrov).
• The TimeSeries table engine now supports dictionary-encoded series identifiers: the id column can be declared with a LowCardinality type, and new tables use Tuple(UInt64, LowCardinality(UUID)) by default while existing tables keep their declared id type. The previous layout remains available through TAGS INNER COLUMNS (id Tuple(UInt64, UUID)), and the new layout gives a 1.16x geometric-mean PromQL speedup in the benchmark. #117033 (Nikita Mikhaylov).
• timeSeriesGroupArray is now supported in SimpleAggregateFunction, and the timeSeries*ToGrid functions now also accept the samples as a single Array(Tuple(timestamp, value)) argument. #117340 (Vitaly Baranov). #117341 (Vitaly Baranov).
• TimeSeries tables now expose the outer samples column as samples instead of time_series. INSERT INTO ts_new SELECT * FROM ts_old still works because the column order is unchanged. #119225 (Vitaly Baranov). The METRICS target is renamed to METRIC FAMILIES, which is what that inner table actually holds; METRICS keeps working as an alias. #119227 (Vitaly Baranov).
• Distributed query plans (make_distributed_plan) now stop idle upstream stages promptly after a satisfied LIMIT, instead of continuing to read and compute data that nobody consumes. They also no longer stall when the initiator pipeline has only one execution thread, because waiting for distributed plan stages no longer occupies that thread. Plans with many nodes now honor the query’s max_threads in worker fragments, so they no longer exhaust the server thread pool or outlive max_execution_time; use_concurrency_control is disabled automatically for these queries because it can starve worker tasks while they wait on exchange sockets. Plans with many finished stages also no longer pay up to 100 ms of idle waiting per stage while draining completed stages, which speeds up multi-join queries several times in some cases. The data sent between nodes is now also serialized on all threads of a task instead of one thread per destination, removing a bottleneck when a query runs on a small number of workers. #115690 (Alexander Gololobov). #116017 (Alexey Milovidov). #116300 (Groene AI). #116878 (Alexander Gololobov). #117328 (Alexander Gololobov). #119067 (Alexander Gololobov).
• parallel_replicas_plan_based can now collect runtime statistics and automatically enable parallel replicas for supported queries. #115788 (Igor Nikonov).
• A Merge table can now be read with parallel replicas: reads from the underlying MergeTree tables are coordinated across replicas, and aggregations and joins above them are distributed too. Enabled by the parallel_replicas_allow_merge_tables setting together with parallel_replicas_plan_based. #115001 (Igor Nikonov).
• Added experimental support for the silk fiber runtime, enabled with the enable_silk_runtime server setting. When it is enabled, subsystems that support it can run jobs on fibers instead of occupying an operating-system thread while waiting for I/O. #112667 (Miсhael Stetsyuk).
• Adaptive codec selection (enable_adaptive_codec_selection) for MergeTree now also applies to columns without a specialized candidate codec: when no type-specific codec candidate exists, each block is written with whichever of the table’s default codec and NONE compresses it smaller, so incompressible data is no longer stored larger than raw. For Float32 and Float64 columns it now also considers the ALP codec. #115639 (Raufs Dunamalijevs). #117553 (Raufs Dunamalijevs).
• Added aggregation pushdown below JOIN (eager aggregation) to the experimental Cascades optimizer as a cost-based alternative: a partial aggregation is performed on one join input before the join and merged above it, or, when the join provably neither duplicates nor null-extends the aggregated side and the join keys are a subset of the GROUP BY keys, the aggregation is pushed below the join entirely. Controlled by the new setting cascades_aggregation_pushdown (enabled by default; takes effect only with enable_cascades_optimizer = 1 and make_distributed_plan = 1). #115354 (Dmitry Novik).
• Queries with INTERSECT and EXCEPT now work under the experimental make_distributed_plan setting. #116563 (Alexander Gololobov). When make_distributed_plan is enabled and no distributed plan can be built, the query falls back to local execution. #116487 (Antonio Filipovic). A failing make_distributed_plan = 1 query reports the error of the task that failed, not the closed exchange connection that failure caused; worker task failures reach the client with the worker’s error code. #118390 (Groene AI).
• Fixed Column '__text_index_...' not found in table and Query builder not found for text search query when a query using a text index read with parallel_replicas_plan_based enabled. #116363 (Igor Nikonov).
• parallel_replicas_plan_based now supports RIGHT JOIN and ORDER BY ... WITH FILL; when the analyzer is disabled, the query runs locally instead of falling back to query-based parallel replicas, because plan-based parallel replicas are not supported there. #113515 (Groene AI). #116135 (Igor Nikonov). #118781 (Igor Nikonov).
• Adaptive codec selection (enable_adaptive_codec_selection) now applies to every substream of a column, such as Array and String sizes and null maps, instead of the main data stream only. #117348 (Raufs Dunamalijevs).
• WebAssembly UDFs now split an input block that would not fit into the guest’s linear memory into smaller batches, sized by the input’s measured serialized size. The new setting webassembly_udf_input_split_memory_ratio controls how much of the linear memory one call’s input may occupy. #116552 (Vasily Chekalkin). A UDF using the BUFFERED_V1 ABI that returns a result buffer with a non-zero size but a data pointer at linear-memory offset 0 is now rejected instead of unrelated memory being read. #116548 (Vasily Chekalkin).
• The experimental ReaderExecutor now shrinks its read window and block size under memory pressure, configured via reader_executor_memory_pressure_{elevated,high,critical}_level_pct; its default read window is now 8 MiB. #115635 (Sema Checherinda).
• PromQL: implemented the max_over_time and min_over_time functions. #112354 (Valery Petrov). PromQL now also parses legacy octal numeric literals with Prometheus-compatible semantics. #113586 (Minh Vu). sgn now preserves NaN and negative zero. #113747 (Minh Vu). stddev and stdvar no longer return NaN for large finite values whose variance is representable. #114250 (Minh Vu).
• PromQL no longer counts consecutive NaN samples as changes. #113744 (Minh Vu).
• CREATE TABLE ... ENGINE = DeltaLake(...) can now create a new DeltaLake table, writing the initial commit through delta-kernel-rs, attach to an existing _delta_log, or register an existing table into a Unity catalog. It is gated behind the new allow_delta_lake_create_table setting; creating a partitioned table (PARTITION BY) is not supported yet. Writes to a DeltaLake table now cast values to the Delta write-schema type with an accurate cast that throws when a value does not fit the target type, for example 300 written into a Delta byte, instead of silently truncating; the new delta_lake_accurate_write_cast setting, enabled by default, controls this, and compatibility below 26.9 restores the previous truncating behavior. #106011 (Smita Kulkarni).

​

Performance Improvement

• Reduced memory usage and query-plan optimization time for keyed aggregations over large constant-folded literals by hashing the aggregation hash-table-statistics cache key while it is serialized instead of accumulating a full copy in memory first. For an 80 MB literal, peak memory drops by about 250 MB and optimization time by about 2.5x. #115847 (Groene AI).
• Added the query_plan_aggregation_bucket_top_k setting (enabled by default) to control the optimization that keeps only each two-level aggregation bucket’s best n groups when a final aggregation feeds ORDER BY ... LIMIT n. The optimization is now visible in EXPLAIN output. #116124 (Nihal Z. Miaji).
• A WHERE equality is now merged into the JOIN condition even when its operands have different types but a common supertype, such as Int32 and Nullable(Int32). This lets more CROSS JOINs be rewritten to INNER JOINs. #112630 (Hechem Selmi).
• Improved performance of merges of text indexes. #114525 (Anton Popov).
• Reduce CPU usage of the post-filter used by IN and NOT IN predicates. #111690 (Elmi Ahmadov).
• The arrayJoin function syntax is now planned as the ARRAY JOIN operator internally, so it can use the same ARRAY JOIN optimizations. A filter on a column produced by ARRAY JOIN is pushed into the array before it is expanded, so elements that do not match are never turned into rows and other columns are not replicated for them, which speeds up selective ARRAY JOIN ... WHERE queries over wide rows. A follow-up reduces the fused path’s per-block overhead by avoiding per-window filter copies and computing survivor offsets only when eager replication needs them. Controlled by the query_plan_fuse_filter_into_array_join setting. #115474 (Yarik Briukhovetskyi). #116910 (Yarik Briukhovetskyi). #117660 (Yarik Briukhovetskyi).
• Fixed a performance regression where a query combining a text index filter with ORDER BY ... LIMIT stopped answering the filter from the index and read the whole text column instead. #115586 (Jimmy Aguilar Mena).
• Speed up query pipeline construction for queries with large expressions or many streams by computing each plan step’s output header once instead of once per stream. For a ClickBench Q29-like query on a 96-core machine, this removes about a third of the hot query time. #115677 (Alexey Milovidov).
• Speed up building the hash table of an ANY, SEMI, or ANTI JOIN when the right side repeats the same join keys many times. #115820 (Nikita Taranov).
• Speed up random data generation. rand, rand64, randConstant, randomString, randomFixedString, generateUUIDv4, generateUUIDv7, and dateTimeToUUIDv7 now generate bytes with ARS (reduced-round AES in counter mode) wherever the CPU has AES instructions: on AArch64 chosen at build time, on x86-64 chosen at run time when VAES is present. Targets and CPUs without those instructions keep the previous generator unchanged. A follow-up reduced the fixed setup cost on x86-64, which is noticeable when blocks are small. #116133 (Groene AI). #116158 (Raúl Marín).
• Fix a CPU regression for queries that issue many independent small reads against the same MergeTree part when the table has LowCardinality columns. Such queries on tables with a small index_granularity were up to several times slower since 26.6. #116134 (Groene AI).
• Speed up GROUP BY without aggregates by inlining some functions and by starting prefetching at the same cardinality as GROUP BY with aggregates, which helps when the hash table stops fitting in the L2 cache. #115840 (Nikita Taranov). #116267 (Nikita Taranov).
• Use the bloom_filter skip index for IN predicates on a directly indexed column when transform_null_in = 1 and the IN set contains no NULL values. Previously the index was skipped for such queries, forcing a full scan. #111329 (Groene AI).
• The adaptive GROUP BY aggregation (enable_adaptive_aggregator) now falls back to the ordinary algorithm when rare keys or aggregate arguments are too wide to copy profitably, so aggregations over heavy string columns run at the ordinary algorithm’s speed and memory usage. Adds adaptive_aggregator_freeze_threshold_bytes to bound each thread’s frozen table in bytes. #115755 (Nihal Z. Miaji).
• Lightweight UPDATE queries now write patch parts that carry the table’s sort-key columns, so applying a patch no longer has to hold the whole patch part in memory. #116007 (Alexey Milovidov).
• Added an alternative row-major payload layout for hash joins to speed up join output reconstruction. Controlled by enable_hash_join_row_store (default true) and min_rows_ratio_for_hash_join_row_store (default 5). A probe that emits no right-side columns, such as SELECT count(), no longer publishes a fake zero match count into the row-store cache, so a later full-output run can still use the row-store optimization. Closes #117260. #104884 (Hechem Selmi). #117410 (Groene AI).
• The IEJoin algorithm now chooses which two JOIN ON inequality conditions to use as its key conditions by their estimated selectivity from the column min/max statistics, instead of taking the first two in syntax order. The remaining conditions are evaluated per candidate pair, so picking the most selective pair can speed up such joins by orders of magnitude. #115790 (Vladimir Cherkasov).
• A bloom_filter skip index on an Array column is now used for arrayJoin(column) IN (set), arrayJoin(column) GLOBAL IN (set), and arrayJoin(column) = const, the same way it was already used for hasAny(column, set) and has(column, const). Previously these arrayJoin forms fell back to a full scan. #109536 (Groene AI).
• Reduce cancellation latency for distributed queries that use parallel replicas by not waiting for a replica’s initial announcement after the query is already cancelled. #114862 (Nikita Taranov).
• Speed up reading system.tables by not copying the query context for every row. #115867 (Nikolay Degterinsky).
• Speed up Array(LowCardinality(T)) aggregations by preserving dictionary identity. #115961 (Ivan Babrou).
• Improved the performance of the roundDuration and roundAge functions on AArch64 by rewriting their comparison cascades into a branchless form that the compiler can auto-vectorize. Results are unchanged. #116164 (Groene AI).
• Avoid extremely slow BACKUP of a Memory or KeeperMap table when max_compress_block_size is set to a small value by sizing the temporary file stream with temporary_files_buffer_size. #116233 (Groene AI).
• Speed up server shutdown for databases with many tables by shutting tables down in parallel. Controlled by the new database_catalog_shutdown_table_concurrency setting; 0 means the number of CPU cores. #104969 (Jayme Bird).
• arrayElement and arraySlice now read lazily replicated arrays (produced by ARRAY JOIN and JOIN when enable_lazy_columns_replication is enabled) directly, without materializing them. This reduces memory usage and speeds up queries that index or slice a large array. #112304 (Diego Gomes Tomé).
• A UNION DISTINCT B now deduplicates each branch in parallel before merging them, the same way its SELECT DISTINCT * FROM (A UNION ALL B) rewrite already did. Previously the query funneled every row through a single thread, so it did not scale with max_threads. #115173 (Groene AI).
• Speed up regular-expression searches for simple anchored literals, including prefix, suffix, exact-literal, and escaped-literal matches. #116554 (Elmi Ahmadov). #117331 (Elmi Ahmadov).
• Reduce CPU and memory overhead of TTL deletions during merges: blocks are now filtered in one vectorized pass instead of being rebuilt column by column, and a block with nothing to delete is passed through untouched instead of being copied. #116623 (Shaohua Wang).
• Optimize one-argument date conversions on Date and DateTime columns. #114267 (Ivan Babrou).
• clickhouse-local now attaches the system and information_schema tables on first access instead of at startup. #115489 (Konstantin Bogdanov).
• Speed up GRANT, REVOKE, and SHOW GRANTS for users with many grants by avoiding a deep copy of the access-rights tree when it is only read. #115696 (miao tang).
• Reduce memory usage when inserting into a wide MergeTree part that writes many more streams than it has columns, for example a table with a few bucketed Map columns. The min_columns_to_activate_adaptive_write_buffer setting is now compared against the number of streams the part writes, since one write buffer is allocated per stream. #115759 (Groene AI).
• Fixed a performance regression since 26.6: a query filtering a Nullable column with IN (subquery) read the whole table instead of skipping the read when the subquery returned no rows. Closes #116450. #116538 (Yarik Briukhovetskyi).
• Fixed quadratic complexity of groupConcat over a FixedString argument. The whole argument column was converted to String once per row, so a query over 400,000 rows took 45 seconds instead of 0.1 seconds and could overrun its max_execution_time, which is only checked between blocks. #116795 (Groene AI).
• Push a WHERE condition on a USING column through a RIGHT JOIN so the left MergeTree input can use it as a primary-key index condition, even when the join keys have different types. Applies with enable_analyzer = 1. #103478 (Manuel).
• Reduce memory usage of Parquet dictionary-based row-group pruning (input_format_parquet_dictionary_filter_push_down) several times by hashing integer dictionaries directly and keeping the value set as a sorted vector instead of a hash table. #115744 (Alexey Milovidov).
• Reading dictionary-encoded Parquet data is now faster: dictionary indexes are decoded and resolved to values in one fused pass, turning runs of a repeated value into bulk fills instead of per-row gathers, for about 20% faster scans of dictionary-encoded numeric columns. #115819 (Alexey Milovidov).
• Speed up some PREWHERE queries with text indexes by placing text-index lookups and other zero-size-column predicates before expensive predicates on the indexed column. #116546 (Jimmy Aguilar Mena).
• Speed up the numbers table function, sorting, and other bulk index fills on AArch64 by improving code generation of their sequential-fill loops. Results and x86-64/macOS code generation are unchanged. #116616 (Groene AI).
• Speed up groupUniqArray for 8-bit types (UInt8, Int8, Enum8, Bool) by using a 256-bit bitmap instead of a hash set. #116880 (Manuel).
• Writing rows in JSON output formats that emit object keys, including JSONEachRow and its aliases, PrettyJSONEachRow, JSON, JSONStrings, JSONStringsEachRow, JSONObjectEachRow, and GeoJSON, is now faster because each field name is no longer copied into a temporary string and re-measured for every row. Compact formats that emit no keys are unchanged. #117082 (Groene AI).
• Fixed a performance regression introduced in 26.4 when writing JSON and Dynamic columns in text formats, including Dynamic(JSON) text output. #115852 (Pavel Kruglov). #117655 (Nikita Fomichev).
• Speed up repeated insertion of the same LowCardinality value into an in-memory column, for example during joins. #115020 (Minh Vu).
• Correlated subqueries with equality on columns that differ only in nullability, or on native-number columns with a least common type, are now decorrelated via equivalent-expression substitution instead of falling back to a CROSS JOIN with a post-join filter. This avoids severe memory growth on such queries. Closes #113407. #115493 (Dmitry Novik).
• Speed up division and modulo of Int128, UInt128, and Decimal128 values that fit in 64 bits. Decimal128 rounding with round, floor, ceil, trunc, and roundBankers is 1.5x faster on x86-64, and gains more on architectures that have no 128-bit divide instruction. Results are unchanged. #116199 (Groene AI).
• Functions with several SIMD implementations, including MD5, SHA1, hashing functions, greatCircleDistance, and geoDistance, now choose one from the CPU’s capabilities instead of benchmarking them at runtime. MD5 becomes 2-3.7x faster. The function_implementation setting is obsolete and has no effect. #116501 (Raúl Marín).
• Speed up dotProductTransposed on QBit columns with floating-point element types on hosts where SimSIMD has no vectorized dot kernel, which for Float64 means AArch64 and x86 without AVX-512. The reduction now uses independent partial sums instead of a single accumulator, so its summation order changes. Closes #116784. #116816 (Groene AI).
• Reduce jemalloc fragmentation by lowering lg_extent_max_active_fit to its default value 6. #116836 (Azat Khuzhin).
• Speed up formatDateTime with non-constant time zone arguments by caching resolved time zones locally. #116851 (Robert Schulze).
• PromQL selector scans over TimeSeries tables are now parallelized, which speeds up reads of the narrow, decode-heavy samples table columns. #116932 (Nikita Mikhaylov).
• Speed up conversion of TimeSeries tag arrays by reserving the final output size up front and by building extracted tag values and NULL maps in one pass. #115620 (Minh Vu). #115687 (Minh Vu).
• Speed up timeSeriesGroupArray when the input samples are already sorted by timestamp. #117703 (Vitaly Baranov).
• The Native protocol now batches end-of-query control messages into a single socket write by default, and sync now flushes them immediately. This removes one round trip of latency per query for clients on high-latency connections that idle between queries. #111034 (Kaviraj Kanagaraj).
• Phrase search in text indexes now stores positions in packed blocks and reads only the blocks that cover candidate rows, reducing position-data reads and memory usage. #111550 (Elmi Ahmadov).
• Appending many default values to Array and String columns now pre-sizes the offsets array, reducing reallocations and peak memory usage. #113577 (Groene AI).
• The preliminary per-stream DISTINCT now abandons deduplication when the input is almost entirely unique, freeing its hash tables because the final DISTINCT still deduplicates the output. This roughly halves memory usage and speeds up SELECT DISTINCT on unique data. Controlled by allow_preliminary_distinct_abandoning and observable via DistinctTransformsAbandonedDeduplication. #116508 (Nihal Z. Miaji).
• PromQL instant-vector responses now reuse the formatted evaluation timestamp within each result block. #115567 (Minh Vu).
• Speed up TimeSeries reads by writing sampling keys directly into the result column. #115683 (Minh Vu).
• Reduce lock contention and per-lookup allocations during serialization by sharding the object pool. #116869 (Manuel).
• LIKE ... ESCAPE 'c' and ILIKE ... ESCAPE 'c' can now use skip indexes instead of falling back to a full scan. TYPE text indexes prune granules for both; LIKE ... ESCAPE 'c' can also use the ngrambf_v1, tokenbf_v1, and sparse_grams indexes, and NOT LIKE ... ESCAPE 'c' can use bloom_filter skip indexes and the primary key when applicable, matching plain NOT LIKE. TYPE text indexes still do not support NOT LIKE. #105848 (Groene AI).
• Distributed ORDER BY query plans can now use read in order when the query order matches the table’s sorting key, and projections can reuse that base-table in-order read too. #115244 (Shankar Iyer). #116854 (Mikhail Artemenko).
• Slightly speed up INSERT into JSON columns by making typed-path default insertion cheaper. #107359 (Pavel Kruglov).
• Added the query_plan_lift_predicate_across_join optimization, which copies equi-join filter conjuncts across MergeTree join inputs for better primary-key pruning. It speeds up some queries by 70-80x. #108377 (Yarik Briukhovetskyi).
• Automatic parallel replicas now share prepared sets with the probe plan instead of executing them twice. #114231 (Nikita Taranov).
• Supported LIKE and ILIKE patterns can now use text indexes. #114728 (Elmi Ahmadov).
• Improve generic arrayReverse performance by reserving space for all nested elements. #115011 (Minh Vu).
• LEFT ANTI JOIN and LEFT SEMI JOIN, when no column of the right table is selected, now build a hash table that stores the keys alone and release the right-hand blocks as they are built. This makes them up to 1.5x faster and can halve memory usage. Controlled by the new setting enable_join_key_only_hash_tables, which is enabled by default. #115301 (Nikita Taranov).
• Speed up timeSeriesFromGrid for dense rows by bulk-copying values. #115518 (Minh Vu).
• Answer min, max, and count aggregations without GROUP BY and filters directly from per-part column statistics, collected automatically for suitable columns by the auto_statistics_types MergeTree setting, reading no data for the parts that have statistics materialized. This speeds up ClickBench Q6 (SELECT MIN(EventDate), MAX(EventDate) FROM hits) from tens of milliseconds to a few milliseconds. Controlled by the new setting use_statistics_for_min_max_aggregation, which is enabled by default. #115738 (Alexey Milovidov).
• On AArch64, short runtime-sized copies in hot string, join, and part-reading paths now stay inline instead of being rewritten to a memcpy call. #117353 (Groene AI).
• Improve performance of bitmapAndCardinality, bitmapOrCardinality, bitmapXorCardinality, bitmapAndnotCardinality, and bitmapHasAny by computing the result directly instead of first materializing the intersection bitmap. #117478 (Anand Kr Shaw).
• Reduce memory usage of reading a TimeSeries table directly with SQL (SELECT ... FROM ts_table). The samples are now aggregated in sorting-key order instead of being buffered in a hash table, so peak memory no longer grows with the size of the scanned window, and LIMIT is pushed into the read. #117726 (Nikita Mikhaylov).
• Queries with join-side disjunction predicates can now push the per-side filters down to PREWHERE on MergeTree reads, reducing unnecessary column reads. #117760 (Nikita Taranov).
• Reduce clickhouse-keeper request latency for concurrent RemoveRecursive requests when many uncommitted nodes exist. #117778 (Alexander Gololobov).
• Improve performance of hasToken, hasAnyTokens, and hasAllTokens queries with text indexes in lazy posting-list apply mode. #117905 (Anton Popov).
• Speed up k-way merges: merging sorted streams with multi-column or collated sorting keys uses a sorted array of cursors instead of a heap (7-8% faster), merges of non-intersecting parts detect the whole cursor as one batch in O(1) comparisons, and runs of duplicate keys inside a part are skipped, which makes SELECT ... FINAL from ReplacingMergeTree use up to 40% less CPU and the insert-time deduplication of optimize_on_insert about three times faster when there are many duplicates. #115869 (Alexey Milovidov).
• Reading columns with many substreams, such as JSON, from object storage is much faster and makes far fewer requests: the marks of all streams of a MergeTree part are now loaded asynchronously and in parallel by default (setting load_marks_asynchronously). #118477 (Alexey Milovidov). Marks of streams with an empty data file, such as the SharedVariant substreams of JSON paths with values of a single type, are not loaded at all, which removes about a third of the requests. #118478 (Alexey Milovidov). The metadata streams of JSON and Dynamic columns are no longer read twice when prefetching. #118479 (Alexey Milovidov). A read of a file inside a packed part is bounded to the file’s slice, so it does not transfer the rest of the archive and reuses HTTP connections. #118483 (Alexey Milovidov).
• Improved performance of reading individual paths and subcolumns from JSON columns that use the advanced shared data serialization. Each path and substream now starts in its own compressed block, so a selective read no longer decompresses data belonging to other paths. #118154 (Pavel Kruglov).
• The join reordering optimizer uses the CD-C and CD-A conflict detectors from the SIGMOD’13 paper “On the Correct and Complete Enumeration of the Core Search Space”, so joins of any kind can be reordered, including ANTI and SEMI joins, not only INNER joins. For seven relations, CD-C explores about 12 times more valid plans than before. #117677 (Fisnik Kastrati).
• Decode Iceberg data manifest files concurrently while the list of data files for a query is produced. The setting iceberg_manifest_decode_concurrency (default 4) bounds how many manifests, delete and data alike, are decoded at a time; it replaces iceberg_delete_manifest_decode_concurrency, which is kept as an alias. #116723 (Aaron Harlap).
• Use the lower_bound and upper_bound statistics of the Iceberg manifest list to prune data files. Closes #104567. #115817 (Konstantin Vedernikov).
• Reduce peak memory of the JSONExtract* functions on the rapidjson code path (allow_simdjson = 0): the parser’s DOM allocator is rewound between rows, so the DOM is bounded by one document instead of growing with max_block_size. A 100-level nested document uses 84% less memory and is 30% faster, a 200-key object 61% less and 21% faster. #117618 (Groene AI).
• Faster writing of String values in the JSON output formats (JSONEachRow and its variants, JSON, JSONObjectEachRow, PrettyJSONEachRow and GeoJSON) and in JSONExtractRaw: a run of bytes that needs no escaping is copied in one block instead of one byte at a time. Values in which most bytes need escaping become slower. The output is unchanged. #117755 (Groene AI).
• addDays, addWeeks, subtractDays and subtractWeeks over a DateTime column are up to 12% faster in time zones with UTC-offset changes: a per-row range check that the argument type makes impossible to fail is no longer executed. #118219 (Groene AI).
• toHour and toMinute no longer walk the calendar lookup table in time zones where the result is pure arithmetic: toHour in every zone whose UTC offset does not change after 1970 (including UTC and Asia/Kolkata), and toMinute in zones that keep a constant minute-of-hour, such as Asia/Kolkata, Pacific/Chatham and America/St_Johns. Results are unchanged. #117562 (Groene AI).
• Reduce the per-query cost of applying the compatibility setting and of logging changed settings into system.query_log. A query on a service that sets an old compatibility value can run more than twice as fast. #118114 (Raúl Marín).
• The bound on the cost of estimating the selectivity of col IN (...) from column statistics, controlled by statistics_max_set_size_for_exact_selectivity_estimation, is now also applied when compatibility is set to 26.6 or an earlier version. #118095 (Nikita Fomichev).
• Automatic parallel replicas now skip their planning overhead entirely when no read step of the query is estimated to read at least automatic_parallel_replicas_min_bytes_per_replica times the number of replicas. #116521 (Nikita Taranov).
• Join runtime filters now reach the local probe-side plan under parallel replicas when parallel_replicas_filter_pushdown is enabled, so grouped probe-side queries filter rows before local aggregation instead of after it. #118650 (Nikita Taranov). The join runtime filter no longer builds a Bloom filter that the build-side key count predicts to be too dense to be useful. #118336 (Manuel).
• The bucket-local top-K optimization (query_plan_aggregation_bucket_top_k) no longer materializes every group’s key for a statistic that only automatic parallel replicas consume, so GROUP BY ... ORDER BY count() DESC LIMIT n over high-cardinality keys spends less CPU in the final aggregation. #118374 (Groene AI).
• Skip analyzing the implicit minmax_count projection during query planning when the query aggregates with a function that projection cannot provide. Previously every aggregating query over a MergeTree table interpreted the projection’s definition while optimizing the plan, even for sum, avg, uniq and every other aggregate it can never serve. #118723 (Groene AI).
• A condition on an ARRAY JOIN element, such as ARRAY JOIN tags AS t WHERE t IN (...), is now used for skip index analysis the same way arrayJoin(tags) IN (...) already is. #118582 (Yarik Briukhovetskyi). A WHERE condition that compares an ARRAY JOIN element with other columns of the row is applied before the expansion, so filtered-out elements are never expanded. #118904 (Yarik Briukhovetskyi).
• ARRAY JOIN of a Nested column inside a subquery, CTE, UNION branch or view now reads only the referenced subcolumns, like a top-level ARRAY JOIN does. Fixed an exception Tuple doesn't have element with name when a second ARRAY JOIN referenced a subcolumn of a Nested column joined by the first one. #119055 (Vladimir Cherkasov).
• A bloom_filter skip index is no longer selected for a condition built only from != or NOT IN. A bloom filter can only prove that a value is absent from a granule, so such a condition never skipped a granule while the whole index was still read. force_data_skipping_indices now reports INDEX_NOT_USED for it. #116544 (Groene AI).
• Text indexes on mapKeys and JSONAllPaths can now prune IN predicates even when the right-hand-side set exceeded use_index_for_in_with_subqueries_max_values and was built without stored elements. #118128 (Nikita Taranov). A text index built with the array tokenizer can now serve arbitrary LIKE and ILIKE patterns by scanning its dictionary. #115649 (Elmi Ahmadov).
• Speed up primary key index analysis for conditions with a large IN set by reusing per-thread buffers instead of allocating on every call. #110937 (Manuel). Avoid a redundant LowCardinality type unwrap per mark when evaluating IN conditions on the primary key. #117054 (Manuel).
• Improved performance of aggregate functions that combine a null map with an -If condition (sum, avg, min/max/argMin/argMax with -If, functions with two or more Nullable arguments such as corr and quantileExactWeighted, and the timeSeries* functions with -If) when a block is consumed in increasing sub-ranges, as optimize_aggregation_in_order and window frames do. #118025 (Groene AI).
• Hashing a value that has no contiguous in-memory representation (JSON, Variant, AggregateFunction) no longer allocates and copies a buffer per row. xxHash64 over a Variant column is about 40% faster, sipHash64 over a JSON column about 5%. Counting distinct aggregate states with uniqExact is about 33% faster for 200-byte states; uniq, DISTINCT and arrayDistinct hash states through the same path. #119322 (Groene AI).
• Speed up compression and decompression with the T64 codec on AArch64 and on x86-64 hosts without AVX-512. The byte and bit transposes now use portable vector shuffles instead of a per-byte scalar loop. Compressed output is unchanged. #116206 (Groene AI). Faster DoubleDelta decompression. #118787 (Nikita Mikhaylov).
• Improve LIMIT BY performance by stopping input reads after a constant grouping key reaches its limit. #118770 (Minh Vu).
• Squash small chunks before writing the intermediate table during recursive CTE evaluation. Deep recursions no longer accumulate one tiny Memory block per produced chunk, which degraded per-step reads of the working table. #107068 (Zach Naimon).
• Speed up jaroSimilarity and jaroWinklerSimilarity using AVX2. #104457 (Manuel). Speed up isValidASCII by using SIMD kernels from simdutf. #119062 (Konstantin Bogdanov). Faster base58Encode and base58Decode. #117619 (Groene AI). Improve the performance of string functions in byte mode by avoiding unnecessary ASCII scans. #115003 (Minh Vu).
• Speed up mortonEncode and mortonDecode by resolving each input column and the constant range mask once per call instead of once per row. Results are unchanged. #117727 (Groene AI).
• Speed up the fast Walsh-Hadamard transform used by randomHadamardTransform and by vector quantization with AVX2 (about 28% higher throughput). #119064 (Jingxin (Philip) Li).
• Faster cbrt, fmod, fma and hypot: in release builds the slower implementations from Rust’s compiler_builtins were shadowing the optimized llvm-libc ones, and hypot now uses the libc implementation. #118910 (Konstantin Bogdanov).
• Reduced the time and peak memory of s2GetNeighbors over many rows by reserving its result array once per block instead of once per row. Peak memory at four million rows in a single block drops by about 44%, and the gain grows with max_block_size. #115919 (Groene AI).
• Avoid materializing constant arguments in generic least and greatest evaluation. #119108 (Minh Vu).
• Faster materialization of lazily replicated sparse columns when consecutive output rows refer to the same source row. #118237 (Minh Vu).
• Skip rebuilding column TTL blocks when no row in the block is expired. #119109 (Minh Vu).
• Faster LEFT ANTI joins with the ie_join algorithm when there is no residual condition: empty bitmap scans are skipped. #118584 (Minh Vu).
• PromQL: a math binary operator applied to two aggregations of the same expression with the same grouping, such as sum(x) - max(x) or sum by (job) (rate(x[5m])) / count by (job) (rate(x[5m])), now evaluates the argument once and computes both aggregations in a single GROUP BY, instead of scanning the argument twice and joining the two results by group. #118307 (Nikita Mikhaylov). Use the constant empty group for aggregations with an empty by() modifier. #118132 (Minh Vu). Improved the performance of the deriv function. #118987 (Vitaly Baranov).
• A PromQL selector matching a whole metric on a TimeSeries table no longer re-evaluates its primary-key range on id for every row it reads; measured 23.7% faster on a 105M-row whole-metric query. #119092 (Groene AI).
• Faster timeSeries*ToGrid aggregate functions: per-sample timestamp validation is skipped when the sorted bucket endpoints already prove the range. #118189 (Minh Vu).
• Reduced contention on the global logger mutex in the MergeTree read path, which was noticeable on servers with many cores and queries touching many parts. #117498 (Manuel). Partition values are no longer formatted for a trace-level log message when trace logging is disabled. #119605 (Rory Shanks).
• Push the disk_name filter of system.remote_data_paths into the disk traversal, so a query that names the disks it needs no longer walks every object storage disk on the server. #118400 (Alexey Milovidov).
• Speed up clickhouse-client startup with many options. #118386 (Ivan Babrou).
• Fewer allocations when reporting ProfileEvents snapshots. #118305 (Ilya Yatsishin).
• Speed up reading subcolumns from wide parts and from compact parts with substream marks by avoiding unnecessary full-column serializations. #111020 (Rory Shanks).
• Speed up aggregation batches with non-null states. #117009 (Ivan Babrou).
• Speed up flushing system.opentelemetry_span_log by writing spans directly into its columns instead of boxing each value into a Field. A burst of spans no longer makes SYSTEM FLUSH LOGS opentelemetry_span_log hit its timeout. #117481 (Alexey Milovidov).
• Avoid join-order regressions when the selectivity estimator cannot evaluate every predicate in a WHERE condition: unaffected predicates are still used to estimate cardinality instead of falling back to full-table estimates. #118137 (Nikita Taranov).
• Improved performance of some array functions: generic arrayFirst and arrayLast reserve result rows, arrayExcept no longer copies the input when the exclusion array is empty, single-argument arrayConcat no longer rebuilds the result, and arraySlice(arr, 1) no longer rebuilds the result array. #118188 (Minh Vu). #119665 (Minh Vu). #119697 (Minh Vu). #119817 (Minh Vu).
• Reuse HTTP connections to object storage when a read range ends at the end of the object. Streams reading the same MergeTree part no longer open a new connection for each range. #118480 (Alexey Milovidov).
• SELECT count() can use the trivial count optimization on tables with lightweight-update patches that only change values, avoiding unnecessary table and patch reads. #118882 (Anton Popov).
• Improve Map subcolumn pruning: mapContainsValue can read only the values subcolumn, mapKeys/mapValues filters can read only the matching keys or values subcolumn even when the full Map is also selected, and has/notHas over a Map can read only the keys subcolumn when it is safe to do so. #119172 (Minh Vu). #119261 (Minh Vu). #119190 (Minh Vu).
• The advanced shared-data serialization for JSON now sizes compressed blocks by min_compress_block_size instead of flushing per path and substream, avoiding large part-size regressions on tables with many shared-data paths. #119285 (Pavel Kruglov).
• Reduce decompression overhead for chained compression codecs by constructing codec-description ASTs only when requested. #119490 (Raufs Dunamalijevs).
• Accelerate countBytesInFilter and countBytesInFilterWithNull on AArch64 using direct NEON vector register accumulation, improving filter counting throughput by up to 15x. #119555 (Joshua Dorst).
• Partition pruning now runs before loading column statistics and before calculating skip-index ordering costs for PREWHERE optimization. Query-plan PREWHERE column sizes and statistics are now also estimated from the parts left after pruning, retaining the existing byte-cost estimate when those parts have no per-column measurements. #119558 (Rory Shanks). #119606 (Rory Shanks). #119607 (Rory Shanks).
• Speed up single-argument uniq, uniqCombined, and uniqCombined64 on common scalar types when aggregating many interleaved groups. #119594 (Manuel).
• UTF-8 validation is now vectorized across ClickHouse, not only inside the isValidUTF8 SQL function on x86. Text index conditions, Prometheus query parsing, format escaping and output, detectLanguage, and isValidUTF8 itself now validate long strings much faster, including on AArch64. Closes #116074. #116204 (Groene AI).
• Improved the performance of the splitByNonAlpha and splitByString tokenizers. #119434 (Anton Popov).
• Speed up lowerUTF8 and upperUTF8 for mixed columns with mostly ASCII rows. #119790 (Minh Vu).
• Speed up sorting of wide tables (ORDER BY over many columns) by skipping an unnecessary per-row check in the merge step when no column can be replicated. This removes overhead that was noticeable on ARM. #110627 (Groene AI).
• Speed up query analysis for large parameterized views by avoiding repeated hashing of the substituted view query during query-tree comparisons. Closes #118736. #119517 (Alexey Milovidov).

​

Improvement

• AI functions (aiGenerate, aiClassify, aiFilter, aiExtract, aiTranslate, aiRedact, aiEmbed, aiSimilarity) are now beta and no longer require the allow_experimental_ai_functions setting, which is now obsolete. The ai_function_* settings moved to the beta tier, and ai_function_max_retries now defaults to 1 so that a single transient provider error does not fail the query. #116227 (George Larionov).
• Added dimensional metrics for S3Queue. #115422 (Bharat Nallan).
• Very large max_streams_to_max_threads_ratio values on Merge table reads are now validated correctly: stream-count calculations that overflow size_t or exceed the maximum source count are rejected with PARAMETER_OUT_OF_BOUND instead of hitting undefined behavior. #113382 (Alexey Milovidov).
• The Web UI no longer shows sort and filter controls in column headers when the first page of a result contains only a single row and those controls would only re-run the same query. Results that are already sorted or filtered, or truncated at the display limit, still show them. #116181 (Alexey Milovidov).
• A trace started by sampling (opentelemetry_start_trace_probability) is now written back into the query’s client info, so remote and distributed secondary queries and DDL entries join the same trace instead of starting disjoint ones. #115619 (Diego Gomes Tomé).
• The Empty query syntax error now says which text was empty when the text is not the query itself but a fragment parsed on its own, such as the value of the parallel_replicas_custom_key or additional_result_filter setting: Empty query (parallel replicas custom key). The error for custom-key filtering requested without parallel_replicas_custom_key now names the settings involved. #115959 (Alexey Milovidov).
• Using join_algorithm = 'direct' against a Buffer table now returns the ordinary unsupported-storage error instead of Cannot clone ReadFromPreparedSource plan step (NOT_IMPLEMENTED). #115309 (Groene AI).
• INSERT INTO ... SELECT from a cluster table function such as s3Cluster with parallel_distributed_insert_select now expands aliases in the WHERE condition before pruning files to read, so alias-based filters participate in pruning and no longer write a stack trace to the server log. #115932 (Alexey Milovidov).
• Wrapped object-storage disks now account reads and writes to the workload of the underlying object storage, so workload scheduling uses the intended limits and priorities. #116288 (Mikhail Artemenko).
• A server started without a config file now declares the same single-node default cluster as the packaged config, so cluster('default', ...) and clusterAllReplicas('default', ...) work instead of failing with CLUSTER_DOESNT_EXIST. #116382 (Alexey Milovidov).
• The / commands of the client (/help, /man, /clear) are now offered by the line editor: typing / at the beginning of the input lists them as hints, and Tab completes the command being typed. A mistake in a command name is now reported with similar commands suggested, instead of being parsed as SQL. #115039 (Alexey Milovidov).
• Support Decimal types when writing into Iceberg tables, including decimal partition columns and decimal min/max statistics. #115832 (Konstantin Vedernikov).
• Web UI: the query text selection stays visible after the editor loses focus, so it is possible to see what the Run selected button will run. #116201 (Alexey Milovidov).
• ReplicatedMergeTree mutations now automatically prune affected partitions from the WHERE condition when possible. Mutation queries also now accept multiple values in the IN PARTITION clause. #110968 (Alexey Milovidov).
• Allow JOIN on keys that have no common supertype, such as UInt64 and Int64. Previously such queries were rejected with There is no supertype for types UInt64, Int64. Keys are now converted to the type of the values they have in common, and a value outside that range does not match anything. #112951 (Alexey Milovidov).
• Reduce the volume of logs from the parallel replicas coordinator: it no longer prints every part and all its mark ranges on each announcement, read request, and response. #115833 (Alexey Milovidov).
• The Maybe you meant hint for an unknown column, function, or table is now printed before the query instead of after it, so it stays visible for long queries. When a column cannot be resolved from a table expression with an alias, the message also names the table behind the alias. #116809 (Alexey Milovidov).
• Comparing a number with an empty or truncated string literal, such as WHERE numeric_column <> '', now reports Cannot convert string '' to type Int32 instead of Attempt to read after eof. A correlated subquery returning several columns where one value is expected now also reports that directly. #116812 (Alexey Milovidov).
• A Values row whose values are fine but whose , or ) is missing or misplaced now reports the missing delimiter and the column it should follow, instead of claiming that the last value could not be parsed as its own type. #116877 (Alexey Milovidov).
• TLS CA certificates configured with openSSL.server.caConfig / openSSL.client.caConfig and caConfig of composable protocols are now reloaded on file changes and on SYSTEM RELOAD CONFIG, allowing CA rotation without restart for server, client, and Keeper connections. #117387 (James).
• Bare UNION now tells you to write UNION ALL or UNION DISTINCT, or to set union_default_mode, instead of naming an internal AST class. The EXCEPT and INTERSECT variants are reworded the same way. #116815 (Alexey Milovidov).
• An aggregate function used in WHERE or PREWHERE, often through an alias such as SELECT count() AS c FROM t WHERE c > 1, is now told to use HAVING instead. Values that cannot be parsed as Date no longer claim to be too short when that is not the reason; the error now names the expected format. #116817 (Alexey Milovidov).
• Invalid calls to the internal viewExplain table function now return BAD_ARGUMENTS instead of analyzing the invalid expression and potentially failing with an unrelated error such as UNKNOWN_TABLE. #108384 (Christoph Viebig).
• Unexpected jemalloc warnings are no longer suppressed, and the server now reports disabled per-CPU arenas in system.warnings. #116899 (Azat Khuzhin).
• Enable reading ReplacingMergeTree data in reverse order with FINAL. #115345 (Alexey Milovidov).
• Added the snapshot_zstd_compression_level Keeper coordination setting to configure the ZSTD compression level used for snapshots. It defaults to 3, preserving the existing behavior. #116866 (Antonio Andelic).
• A typo in the key, index, or TTL expression of an ALTER TABLE now suggests the column that was probably meant, the way SELECT already does. Previously it printed a list of available columns that, for a MergeTree table, was dominated by virtual columns and truncated before the real ones. #116915 (Alexey Milovidov).
• Aggregate functions that take parameters and are wrapped with the -Array combinator now preserve those parameters correctly, so expressions such as argMaxArray(2) and state types such as AggregateFunction(groupArrayMovingSum(42), UInt64) no longer drop them. In debug and sanitizer builds this also fixes an internal assertion. #110143 (Groene AI).
• rseq is now enabled correctly on x86-64 builds that use link-time optimization; previously it was silently disabled. #117700 (Azat Khuzhin).
• Take a column’s CODEC into account when estimating how many bytes a query reads from compact parts. The estimate compressed a sample with the default codec (LZ4) no matter how the data is really stored, so a column stored with a stronger codec was overestimated, which inflates effective_max_reading_threads and can make automatic parallel replicas distribute a query that is cheaper to run on a single node. #114535 (Nikita Taranov).
• Iceberg tables stored on S3 can now read data files from a different bucket than the table metadata. #116771 (Konstantin Vedernikov).
• An aggregate function with the -If combinator whose condition has the wrong type, or is missing, now says that the last argument is the condition and must be UInt8, instead of only reporting an illegal type for the last argument. An ASOF join with no inequality predicate now says that one is needed in addition to the equality predicates, instead of ending the message in , in .. #116916 (Alexey Milovidov).
• clickhouse-local now exposes system.warnings, and ClickHouse now checks rseq compatibility at startup. #117717 (Azat Khuzhin).
• The TimeSeries table engine, the PromQL dialect and the timeSeries* aggregate functions are moved from the experimental tier to the private preview tier. Settings allow_experimental_time_series_table and allow_experimental_time_series_aggregate_functions are renamed to enable_time_series_table and enable_time_series_aggregate_functions; the old names remain as aliases. #118124 (Nikita Mikhaylov).
• TimeSeries tables now store an explicit schema version in the version engine setting. The PromQL execution layer supports only a declared range of versions and rejects queries over tables outside that range with an instructive error. #111204 (Nikita Mikhaylov). A TimeSeries table definition copied with CREATE TABLE ... AS now gets the settings of the new table, and the new id_type setting lets CREATE TABLE ... AS src copy only src’s definition without reading its target tables. #118517 (Vitaly Baranov). #119613 (Vitaly Baranov).
• Reads from TimeSeries tables now respect the join_algorithm and the partition-wise aggregation settings of the query instead of overriding them, and the internal join of samples and tags supports full_sorting_merge. #118864 (Nikita Mikhaylov). The promql_evaluation_time setting defaults to auto regardless of the compatibility setting. #119202 (Raúl Marín).
• Join runtime filter settings are now production-tier (previously experimental). #117383 (Robert Schulze).
• Improved syntax error messages. An unclosed bracket is now reported at the place where the parser stopped, which is next to the actual mistake, and every unclosed bracket is listed with its own position; previously the outermost bracket was blamed, which in a large query can be thousands of characters away from the mistake. A mistyped keyword is now named explicitly (Maybe you meant: SELECT.) instead of only being one item in a list of dozens of alternatives. #116806 (Alexey Milovidov).
• The psql-style \d of the interactive client now describes a table when a table name is given (\d hits), like it does in psql; without an argument it still lists the tables. #118281 (Alexey Milovidov).
• clickhouse-local now merges the config.d and conf.d directories of the current directory into its configuration even if there is no main configuration file, so a directory with a config.d no longer additionally requires an empty config.xml. #118410 (Alexey Milovidov).
• Added the clickhouse_version and system_processor columns to the system log tables. #117896 (Miсhael Stetsyuk).
• EXPLAIN ANALYZE now shows more information about the estimation for joins. #114792 (Kirill Kopnev).
• Automatic parallel replicas now support queries with LIMIT and OFFSET. #117402 (Nikita Taranov).
• Object storage writes (INSERT into s3, azure and other object storage tables) now support parallel replicas. #119214 (Konstantin Vedernikov).
• Added the min_partition_age_to_force_merge_seconds MergeTree setting: when every part in a partition is at least this old, the Simple and StochasticSimple merge selectors compact that partition’s parts within the max_bytes_to_merge_at_max_space_in_pool limit. #115547 (Valery Petrov).
• TRUNCATE TABLE on a ReplicatedMergeTree table no longer blocks concurrent DROP, RENAME and other DDL queries on the same table while the data is being removed. #114672 (Nikolay Degterinsky).
• Allow ALTER TABLE ... MODIFY SETTING storage_policy for MergeTree tables when newly added disks contain only a matching format_version.txt and an empty detached/ directory under the table path. #107736 (Maxim Orlovsky).
• Added the s3_upload_checksum_algorithm setting to choose the checksum sent with S3 uploads (MD5, CRC32 or SHA256). The default is empty and keeps the current behavior. Use CRC32 or SHA256 on FIPS builds, where MD5 is unavailable and uploads would otherwise carry no checksum, which buckets with Object Lock reject. #107318 (Sebastian Vercruysse).
• Partition field summaries are written into the manifest list of Iceberg tables. #118796 (Konstantin Vedernikov). Concurrent INSERTs into an Iceberg table through a catalog no longer back off on a commit conflict and retry immediately. #119238 (Konstantin Vedernikov).
• Log when an object storage listing returns an empty page while reporting that more objects follow. This is legal and is handled correctly, but it has twice been the trigger for silently losing objects and nothing recorded that it occurred. #119065 (Sema Checherinda).
• A warning is now logged when a CTE is declared as MATERIALIZED but the setting enable_materialized_cte is disabled, in which case the keyword is ignored and the CTE is inlined at each reference. Previously the keyword was silently ignored. #118032 (Dmitry Novik).
• An unknown profile event name in the trace_profile_events_list setting now gives a readable error message with hints instead of std::out_of_range: unordered_map::at: key not found. #116049 (Alexey Milovidov).
• The dictionary loader now logs never instead of a timestamp in the year 294247 when an object has no automatic update scheduled (for example a dictionary with LIFETIME(0)), which previously looked like a scheduling bug next to a failed-update message. #118443 (Alexey Milovidov).
• Correct the error message for an unsupported attribute in an ip_trie dictionary: it claimed that Array attributes are unsupported, while in fact they are supported and only Nullable attributes are rejected. #118492 (Alexey Milovidov).
• MEMORY_LIMIT_EXCEEDED is reported promptly for parallel_hash joins whose hash-table construction uses deferred memory tracking. #116477 (Konstantin Morozov).
• Less verbose logging of INSERTs into MergeTree and ReplicatedMergeTree tables. #116486 (Azat Khuzhin).
• The hash tables behind the TLD lists (firstSignificantSubdomainCustom and similar functions), the naive Bayes classifier and text indexes use a simpler, packed string representation. #117175 (Surya Teja).
• ClickHouse Keeper now counts the per-entry allocation overhead in latest_logs_cache_size_threshold, KeeperLatestLogsCacheSize, and lgif’s latest_logs_cache_size, so the configured limit tracks cache memory rather than only serialized payload bytes. #119045 (Nikita Taranov). The latest_logs_cache_entry_count_threshold setting is effective again and limits the in-memory latest-log cache by entry count independently of latest_logs_cache_size_threshold. #119151 (Antonio Andelic).
• Show the progress percentage again for queries reading from the file table function and INFILE. #119213 (Raúl Marín).
• Warn when CLEAR COLUMN leaves an EPHEMERAL-derived MATERIALIZED column stale. #115371 (Shaohua Wang).
• Reduced cancellation latency for INTERSECT and EXCEPT queries: KILL QUERY now interrupts the hash-table build and filter loops within a single chunk instead of waiting for the whole chunk. #118046 (Roman Vasin).
• The compatibility setting can now keep selected new defaults in effect: settings changes marked StartUsingNew are no longer rolled back to the previous default when compatibility points to an older version. #114190 (Mikhail Artemenko).
• Added write support for DeltaLake tables on Azure storage. #116698 (Smita Kulkarni).
• Added aliases without the allow_experimental_ prefix for existing settings whose tier already marks them as experimental or beta; the old names continue to work. #119056 (Robert Schulze).
• Bump the bundled libssh from 0.12.0 to 0.12.2. #119262 (Robert Schulze).
• Function replicate now returns a constant column when all its arguments are constant, like arrayWithConstant. #111069 (sakshichitnis27).
• The PROJECTION_NOT_USED and INCORRECT_DATA errors of force_optimize_projection and force_optimize_projection_name now list every projection the optimizer considered and why each was rejected. #119624 (Mikhail Artemenko).

​

Bug Fix (user-visible misbehavior in an official stable release)

• ALTER TABLE ATTACH PARTITION FROM is no longer replicated in Replicated databases. #94835 (MikhailBurdukov).
• Invalid Azure upload-size settings (for example azure_min_upload_part_size = 0 or azure_max_blocks_in_multipart_upload = 0) are now rejected early with a clear error instead of triggering an internal exception (second_size > 0) deep in the write path. This applies to both query settings and Azure disk configuration. #103394 (Statxc).
• Fixed wrong results for comma or CROSS JOIN queries whose WHERE clause compares columns from the two sides through a non-deterministic expression such as rand. Such queries could return rows that contradict their own WHERE clause because rewriting the join to INNER JOIN re-evaluated the expression per row on each side. #112938 (Groene AI).
• S3Queue and AzureQueue tables whose metadata handle is gone no longer fail with Logical error: 'Files metadata is empty' when read, altered, or flushed; they now report TABLE_IS_DROPPED instead. SYSTEM FLUSH OBJECT STORAGE QUEUE on such a table no longer terminates the server. #113695 (Groene AI).
• Fix LOGICAL_ERROR when filtering an Iceberg column that ALTER TABLE ... MODIFY COLUMN made Nullable. Closes #85029. #114521 (Groene AI).
• PromQL grouping modifiers now support quoted label names. #114545 (Minh Vu).
• Fixed reading a table while a mutation is in progress: a MATERIALIZED column could return a stale or incorrectly recomputed value, and selecting it together with an updated column could fail with an error. #114561 (Shaohua Wang). #115372 (Shaohua Wang).
• Fixed an RBAC bypass in the Alias table engine that allowed users without privileges on the target table to reveal its schema, row count, size, and existence. #114596 (Kai Zhu).
• Fixed a column that was renamed, dropped, and added again in one ALTER returning the dropped column’s values instead of its own default while the mutation was still pending. #114601 (Shaohua Wang).
• ANY RIGHT JOIN and SEMI RIGHT JOIN with several OR-ed conditions in the ON clause no longer duplicate some right-table rows and lose others. A written ANY LEFT JOIN was affected too, because the planner may swap the tables and execute it as ANY RIGHT JOIN. #114676 (Vladimir Cherkasov).
• Fix invalid backslash escape handling in text, tokenbf_v1, and ngrambf_v1 indexes. #115634 (Elmi Ahmadov).
• Fix crashes of the Parquet, ORC, and Iceberg readers on malformed input: a DELTA_BYTE_ARRAY page with an empty length stream, a DELTA_BYTE_ARRAY page of a Decimal column read under a filter, an out-of-range enum in an Iceberg manifest file, and an oversized allocation in the ORC library. #115703 (Alexey Milovidov).
• Reject invalid names of projections and indexes. #115824 (Alexander Tokmakov).
• A KeeperMap table with unparsable metadata no longer prevents server startup; it is now quarantined instead. Closes #115881. #115941 (Groene AI).
• Fixed LOGICAL_ERROR: Index with name auto_minmax_index_<column> already exists when a single ALTER TABLE statement combined RENAME COLUMN with another command, such as MODIFY SETTING, on a table with implicit minmax indices (add_minmax_index_for_numeric_columns and friends). Renaming a plain column also no longer leaves a following DROP COLUMN failing with UNKNOWN_IDENTIFIER, nor a following ADD COLUMN without its implicit index. #116063 (Groene AI).
• Fixed a crash of stateless workers with object storage. #116067 (Konstantin Vedernikov).
• Fixed distributed queries that filter or compute on Decimal, DateTime64, or Time64 constants: pushed-down constants could fail with CANNOT_PARSE_DATETIME, lose precision, or change their Dynamic subtype when sent to remote shards. #95055 (Yakov Olkhovskiy).
• Fixed JSON, Object, and Dynamic columns hashing logically equal values differently depending on their physical layout: how JSON/Object paths were split between dynamic subcolumns and shared data, or whether a Dynamic value sat in a typed or shared variant. Hash-based joins on such a key silently missed matches, and a grace_hash spill could raise a LOGICAL_ERROR (Invalid state transition). A JSON key needed no non-default setting; Dynamic also required allow_dynamic_type_in_join_keys. #109450 (Groene AI).
• Setting a Keeper coordination timeout such as operation_timeout_ms, session_timeout_ms, or startup_timeout to a very large value no longer overflows and expires the wait immediately. Builds with the undefined behavior sanitizer no longer abort. #112088 (Groene AI).
• Fixed a bloom_filter skip index dropping granules that contain matching rows when a String or FixedString column is compared for equality against a constant that is wider than the indexed type, so queries such as SELECT count() FROM t WHERE s = toFixedString('abc', 5) no longer silently return too few rows. #112639 (Groene AI).
• Fixed a large max_concurrent_threads_ratio_to_cores value in CREATE WORKLOAD producing an unspecified concurrent-thread limit. The ratio is now clamped before narrowing, so such a value means “no limit”. #113206 (Groene AI).
• Fix wrong results when plan-based parallel replicas shipped a read pinned by select_sequential_consistency. Follower replicas rebuilt the read without the pinned block-number boundary and returned rows the insert quorum never confirmed. Such a read is now executed locally. #114263 (Groene AI).
• Fixed reading Parquet files whose offset index does not start at row 0. Such a file could make the native Parquet reader return wrong rows, or report LOGICAL_ERROR instead of INCORRECT_DATA. The offset index is now validated when it is read. #114530 (Groene AI).
• Fix query condition cache poisoning by JOIN queries with query_plan_join_shard_by_pk_ranges = 1, which made later queries with the same WHERE condition return too few rows. #114827 (János Benjamin Antal).
• Fixed AI functions under-reporting provider usage in system.query_log and in per-query quotas: counters accumulated before a failed request were dropped, and tokens for a response the provider billed for but that failed validation were never counted. #114905 (George Larionov).
• Fixed Coordination mode mismatch for stream <table>: got Default, expected WithOrder when a parameterized view is used as an IN subquery of an ordered query with parallel replicas enabled. A parameterized view used this way no longer plans its inner read with parallel replicas coordination on a replica. #114916 (Groene AI).
• Fix asynchronous inserts into JSON columns ignoring the max_dynamic_subcolumns_in_json_type_parsing setting. Previously, an asynchronous insert (async_insert = 1) could store more dynamic paths than the setting allows, giving a different column layout than a synchronous insert with the same settings. The cap is now preserved when asynchronous inserts aggregate parsed data. #115329 (Mikhail f. Shiryaev).
• Validate the discriminator of deserialized Variant values, so malformed data is rejected instead of being read out of bounds. #115443 (Azat Khuzhin).
• Fix UNKNOWN_IDENTIFIER when creating a materialized view whose query uses an expression alias as the right argument of IN. Closes #94894. #115651 (Alexey Milovidov).
• Fix RESTORE ... AS <new name> leaving the REFRESH ... DEPENDS ON list of a refreshable materialized view pointing at the original database. The restored view refreshed in response to a table it no longer read, ignored refreshes of its own parent, and stopped refreshing entirely once the original database was dropped. #115689 (Groene AI).
• Validate the deserialized states of stochasticLinearRegression, stochasticLogisticRegression, groupUniqArray, groupArrayIntersect, and topK, and of the Resample combinator, so malformed data is rejected instead of being read out of bounds. #115701 (Alexey Milovidov).
• Mask the aws_external_id setting of the DataLakeCatalog database engine in SHOW CREATE DATABASE and system.databases.engine_full. It was previously printed in plaintext, unlike the other AWS credential settings. Closes #114999. #115791 (Groene AI).
• Fixed SELECT count() never terminating, or returning an invented row count, on a corrupted MsgPack or ProtobufList file. The optimize_count_from_files fast path counted rows without requiring the read position to advance, so a two-byte ProtobufList file produced billions of rows and never returned. Such input is now rejected with the same error the ordinary read path gives. #115884 (Groene AI).
• Fixed ATTACH TABLE ... AS REPLICATED and the convert_to_replicated flag file accepting a conversion whose ZooKeeper path is not a safe path. On a server whose default_replica_path substitutes {database} or {table}, a name containing / used to resolve to a path inside another table’s subtree, letting the converted table register a replica under a table it has no privileges on. Such a conversion is now rejected with BAD_ARGUMENTS, the same way an equivalent CREATE already was. Closes #115967. #115991 (Groene AI).
• Fixed wrong results when joining against a view or view whose projected column has the same name as a column of one of its own inner relations, for example SELECT expr(x.c0) AS c0 FROM a AS x, b AS y. The join order optimizer merged such an expression into the flattened join graph and then applied it twice, so the query silently returned skewed values. When the expression also changed the column type, the same cause produced AMBIGUOUS_COLUMN_NAME at plan time instead. Closes #116036. #116152 (Groene AI).
• Listing tlsv1_3 in <openSSL><server><disableProtocols> (or the client equivalent) now actually disables TLS 1.3. Previously the token was dropped with no error or log line and the server kept negotiating TLS 1.3, so an operator who disabled it believed TLS 1.3 was off while it was not. Closes #79876. #116170 (Groene AI).
• Fixed Invalid number of columns in chunk pushed to OutputPort (LOGICAL_ERROR) when a SELECT ... FROM t STREAM ... ORDER BY ... LIMIT n query read a table with a PROJECTION. Lazy materialization is no longer applied to a STREAM read. #116187 (Groene AI).
• Fixed system.detached_tables failing with NOT_IMPLEMENTED when the server has databases such as Dictionary, Filesystem, HDFS, Overlay, or S3. Unsupported database engines are now skipped during iteration. Closes #104868. #107943 (Aditya Kumar).
• Skip the Query Condition Cache for object-storage files whose ETag is not a strong content identifier, such as HDFS. Previously a same-second, same-size overwrite could reuse stale skip marks and silently drop matching rows. Closes #112016. #112051 (Zhang Yifan).
• Fixed text index functions hasToken, hasAnyTokens, hasAllTokens, and hasPhrase outside WHERE/PREWHERE: they now apply the index’s tokenizer and pre/postprocessing in the SELECT list, aggregate arguments, and with use_skip_indexes = 0, so results match the index path. Closes #108808. #113154 (Elmi Ahmadov).
• Fix wrong results for a query that calls the same table function twice with arguments that differ only in a part of the query that was not taken into account by the AST hash, such as UNION ALL against UNION DISTINCT, INTERSECT against EXCEPT, LIMIT against OFFSET, or the bounds of a window frame. The second call reused the result of the first one. #114171 (Alexey Milovidov).
• Fix THERE_IS_NO_COLUMN / NOT_FOUND_COLUMN_IN_BLOCK when a filter that runs after FINAL is a plain column reference. This affected an explicit PREWHERE b under FINAL and, with default settings, a row policy whose expression is a bare column. #114578 (Groene AI).
• Fix loading old S3 tables with globbed paths and PARTITION BY after upgrading from versions before 26.6. #115277 (Bharat Nallan).
• Fix Iceberg identity-partitioned reads: partition values are now taken from partition metadata instead of returned as NULL. Closes #110216. #116129 (Konstantin Vedernikov).
• KeeperMap now writes single-expression primary keys without redundant outer parentheses, so metadata created by newer servers remains readable by ClickHouse 26.4 during supported rolling upgrades and rollbacks. #116369 (Sergey Kuznetsov).
• Fixed Kafka tables using kafka_sasl_mechanism = 'AWS_MSK_IAM' failing authentication once an hour with SASL authentication error: Access denied. The OAUTHBEARER token lifetime advertised to librdkafka is now capped by the remaining validity of the AWS credentials that signed the token, so the next token refresh is always scheduled before those credentials expire. Successful token refreshes are now logged, and unhelpful ClickHouse-side stack traces are omitted from broker-originated SASL errors in system.kafka_consumers. #109125 (kalavt).
• Fixed silent truncation of DateTime64 and Time64 values with a scale above 6 when date_time_64_output_format_cut_trailing_zeros_align_to_groups_of_thousands is enabled. The printed fraction was capped at 6 digits, so a DateTime64(9) value of 00:00:00.000000700 rendered as 00:00:00.000000 even though the stored value was not zero. The documented output range of the setting widens from [0, 3, 6] to [0, 3, 6, 9]. #112854 (Groene AI).
• Fixed the dpsub and dpsize join-order optimizations (query_plan_optimize_join_order_algorithm = 'dpsub' / 'dpsize') silently dropping single-table filter conditions from a JOIN ... ON clause, which could return extra rows. #114629 (Fisnik Kastrati). #118126 (Fisnik Kastrati).
• ZooKeeper multi requests now reject unsupported and mixed read/write operations. #115768 (Antonio Andelic).
• You can now read Iceberg v2 tables whose manifests omit the optional snapshot_id column, such as tables managed by BigQuery; reading one previously failed with Cannot find column snapshot_id. A manifest that omits sequence_number is also no longer read as sequence number 0. Both values are inherited from the manifest list, as the Iceberg spec requires. #116040 (Takayuki Enomoto).
• Fixed _path path normalization for globbed files and object-storage reads, so WHERE _path = '<path>' no longer misses rows because of redundant or leading separators. Globbed reads through symlinks and .. names also no longer read or, with rename_files_after_processing, rename the wrong file. #116244 (Groene AI). #116252 (Groene AI).
• Fixed wrong window function results for PARTITION BY over a floating-point key (also JSON, Dynamic, and these types nested in Nullable, LowCardinality, Array, Tuple, Map, or Variant) with max_threads > 1: values that compare as equal, such as -0. and 0., could hash into different threads, and one logical partition got several window frames. #116315 (Alexander Gololobov).
• A Native-protocol client can now forward an unknown setting to the server by name. Previously, when a client sent such a setting as a custom setting, the server left the real setting unchanged and reported it at its default value in the settings received with the query. #116317 (Miсhael Stetsyuk).
• Fixed h3Line returning uninitialized memory in the tail of the result array when a line between two valid H3 indices cannot be drawn because it crosses pentagon distortion. Such a call now throws INCORRECT_DATA. #116583 (Groene AI).
• CREATE TABLE ... AS <table function> no longer accepts a PRIMARY KEY in its column list. Previously, such a statement wrote table metadata that the server could not parse back on the next start, so the database failed to load with a syntax error. #116645 (Groene AI).
• Fixed wrong row counts and dropped rows when a DateTime64 primary-key column is filtered through toStartOfDay, toRelativeSecondNum, toRelativeMinuteNum, toRelativeHourNum, toRelativeDayNum, toRelativeWeekNum, toMonthNumSinceEpoch, or toYearNumSinceEpoch with values outside the UInt32-seconds range (before 1970 or beyond 2106). These functions claimed to be always monotonic to the primary index, but their standard-precision results wrapped for out-of-range DateTime64, breaking primary-key pruning. #110344 (Groene AI).
• Fix regressions around use_legacy_to_time: MergeTree key expressions using toTime that were created with the legacy semantics now work again, and CREATE TABLE ... AS SELECT, CREATE TABLE ... AS <table function>, and CREATE MATERIALIZED VIEW no longer fail or silently flatten Nested columns when use_legacy_to_time = 1. Closes #117001. #110958 (Groene AI). #116953 (Yarik Briukhovetskyi).
• Fix OPTIMIZE TABLE ... MANIFEST on Iceberg tables: tables whose metadata file omits the spec-optional top-level refs object no longer fail with Poco::NullPointerException, and an experimental Iceberg table whose rows had all been deleted no longer crashes. Closes #112082. #112163 (Groene AI). #111241 (Smita Kulkarni).
• Fix a server crash during a JOIN that spills to disk. When an allocation failed while the join rehashed its buckets, the in-memory join was left either holding released data or absent altogether, and another build thread could dereference it. #113410 (Groene AI).
• Fix wrong statistics-part pruning for negated floating-point predicates with NaN. #113507 (Han Fei).
• Fix a spurious PreconditionFailed (Code: 499) error when creating or committing to an Iceberg table. If the object store drops the response to the metadata write after the object was written, the retry of that conditional write hits its own object and the query fails even though the write succeeded. Also stop a conditional multipart completion from reporting success over an object it did not write, silently losing the write it was meant to reject. #113994 (Groene AI).
• Fixed lightweight UPDATE on ReplicatedMergeTree being rejected with TIMEOUT_EXCEEDED after a few seconds under concurrent updates, long before the configured lock_acquire_timeout elapsed. The Keeper lock acquisition loop was bounded by a fixed number of tries instead of by the timeout, so raising lock_acquire_timeout had no effect. #114125 (Groene AI).
• Fix ORDER BY ... LIMIT queries on Tuple() columns raising a NOT_IMPLEMENTED exception when top-K dynamic filtering is enabled. Direct empty tuples now skip the optimization, while supported nested composite types continue to use top-K filtering. #114182 (Langning Zhang).
• Fix Iceberg tables partitioned by toRelativeDayNum or toRelativeHourNum publishing the non-canonical transform names days and hours in metadata. ClickHouse now writes the spec-defined singular names day and hour, so external readers such as pyiceberg can parse the partition spec. #114864 (yanglongwei).
• Fix the server failing to start when listen_host contains both :: and 0.0.0.0 and arrowflight_port is configured, and fix two gRPC listeners silently sharing one port in the same configuration. listen_try now also covers the gRPC and Arrow Flight protocols, which bind their listening socket when they are started. #115074 (Alexey Milovidov).
• Fix a case where expensive ZooKeeper exception logging could delay session finalization and leave pending requests unresolved. #115396 (Antonio Andelic).
• An invalid http_handlers configuration is now reported with its own error instead of being wrapped into NETWORK_ERROR "Listen [host]:port failed". Previously, when listen_try was in effect, the error was only logged as a warning about listen_host and the server started without serving HTTP at all. The same applies to the Prometheus listener on prometheus.port, to the Keeper control endpoints on keeper_server.http_control, and to SYSTEM START LISTEN HTTP in clickhouse-local, which blamed listen_host and http_port. #115548 (Groene AI).
• Fixed zero-precision DateTime64 and Time64 text conversions returning DateTime or Time instead of DateTime64(0) or Time64(0): toDateTime64OrNull, toDateTime64OrZero, parseDateTime64BestEffort, parseDateTime64BestEffortOrNull, parseDateTime64BestEffortOrZero, parseDateTime64BestEffortUS, parseDateTime64BestEffortUSOrNull, parseDateTime64BestEffortUSOrZero, toTime64OrZero, and toTime64OrNull. This previously prevented conversion of dates before 1970-01-01 and made the Time64 functions silently lose values that toTime64 parses. #115618 (Mark Needham). #116979 (Navneet Kumar).
• Fixed a server startup failure after creating an Alias table whose target did not exist yet and later became an Alias table itself. Such a table was accepted by CREATE, but rejected while loading metadata, which aborted startup with BAD_ARGUMENTS. The Alias target restrictions are now applied only to freshly supplied definitions. #116153 (Groene AI).
• Fix SHOW CREATE TABLE failing instead of describing the table when the table’s storage object cannot be opened, for example a DataLakeCatalog table the catalog reports as unreadable. The Alias target-access check now consults the engine name of the create query instead of opening the table. #116231 (Groene AI).
• Fix ProfileEvents['SelectedRows'] and ProfileEvents['SelectedBytes'] being counted twice for reads from file, object-storage, data lake, queue, and executable table engines and table functions, and for executable dictionaries and UDFs. They now agree with read_rows and read_bytes, as does the Read: line of EXPLAIN ANALYZE. #116427 (Groene AI). #116679 (Groene AI).
• Fixed executable_pool dictionaries, executable_pool user-defined functions, and ExecutablePool tables waiting ten times longer than the configured max_command_execution_time when the process pool is exhausted, while the resulting error still reported the configured value. Closes #116467. #116560 (Groene AI).
• Fixed a server crash when reading an Iceberg table whose manifest list declares a very large manifest_length. The declared value was used as the metadata cache entry weight, overflowing the cache accounting and terminating the server. The size measured while reading the manifest file is now used instead. #116564 (Groene AI).
• Fixed a hang when a query names a database that does not exist, either through the database setting (for example SELECT 1 SETTINGS database = 'no_such_db') or through the database field of a gRPC request. Such a query now fails with UNKNOWN_DATABASE instead of blocking its connection forever; while it was hung it was also invisible to system.processes and to KILL QUERY. #116582 (Groene AI).
• Fixed a use-after-free where a subscriber of workload entity changes, such as an object-storage disk, could be destroyed while its own change callback was still running. Destroying a workload entity subscription now waits for an in-flight callback and stops new ones from starting. Reachable from a CREATE TABLE with an inline disk(...) definition whose startup fails, concurrently with CREATE RESOURCE or DROP RESOURCE from another session. #116607 (Groene AI).
• Fixed an exception when reading an Iceberg table whose manifest marks a delete file as an equality delete without listing any equality field id. Such an entry is now rejected with ICEBERG_SPECIFICATION_VIOLATION instead of an untyped std::out_of_range. #116617 (Groene AI).
• Fixed the file table function reading and writing outside user_files_path when a .. reached the path behind a directory symlink leaving user_files_path: through brace expansion in a glob pattern, or through a .. partition id in INSERT INTO FUNCTION file(...) PARTITION BY. #116767 (Groene AI).
• Fixed corrupted or crafted Gorilla-compressed data being decoded to wrong values instead of being rejected. Such data is now rejected with CANNOT_DECOMPRESS. #116820 (Groene AI).
• Fixed the loop table function disclosing the column names and types of an Alias table’s target to a user who holds no privilege on that target: a query that read no rows received them in its response header. loop now requires table-level SELECT on the target, matching what it already required of the table name it is given. #116873 (Groene AI).
• Fixed incorrect results or an exception for -State aggregates combined with -OrDefault or -OrFill under WITH TOTALS, ROLLUP, or CUBE. #99980 (Ilya Yatsishin).
• Fixed Iceberg INSERT writing manifest-wide aggregate column statistics into every data_file entry when a partition produced more than one data file. Each entry’s null_value_counts, column_sizes, lower_bounds, and upper_bounds now describe only the data file it points at. Previously an entry could report more nulls than the file had rows, so a spec-conforming external reader could prune files that contained matching rows. #112414 (Groene AI).
• Fix a server abort while committing a mutated part when its column or secondary-index sizes had not been computed yet. #112691 (Groene AI).
• Fix reading corrupted ArrowStream files. A malformed per-buffer uncompressed-length prefix in a compressed Arrow IPC record batch could reach the allocator as an enormous allocation size and produce LOGICAL_ERROR; it is now rejected as INCORRECT_DATA. #113006 (Groene AI).
• Fix wrong results for GROUP BY whose only key is an injective function of constants, for example SELECT 1 FROM t GROUP BY materialize(NULL). Such a key could be removed, leaving the query with no GROUP BY and no aggregation, so it returned one row per input row instead of one row per group. Over a Merge table that mixes a local and a remote child, the same cause could also raise the exception Chunk info was not set for chunk in GroupingAggregatedTransform. #114981 (Groene AI).
• Fixed the exception Not-ready Set is passed as the second argument for function 'in' in three cases: when selecting from a Merge table (or the merge table function) whose underlying table has a row policy containing an IN subquery, when a query with partial_result_on_first_cancel = 1 was cancelled while building the set for IN (subquery), and when a correlated subquery whose body contains IN (subquery) runs with correlated_subqueries_use_in_memory_buffer = 0, which make_distributed_plan = 1 also sets. #114058 (Groene AI). #115083 (Alexey Milovidov). #115868 (Groene AI).
• Fix authentication over the PostgreSQL wire protocol for users with several authentication methods. Users who also had an ssh_key method, or another method with no password, were rejected with Invalid user or password even when the correct password for scram_sha256_password was supplied, and the wrong SCRAM salt was sent when scram_sha256_password was not the last method. #115084 (Alexey Milovidov).
• Fixed THERE_IS_NO_COLUMN (Cannot find column in source stream) for SELECT DISTINCT expr(k) ... ORDER BY k on a table with an aggregate projection grouped on k. Closes #112168. #115147 (Groene AI).
• A text index is now used for hasAnyTokens, hasAllTokens, and hasPhrase called with an explicit tokenizer argument when that argument denotes the same tokenizer as the index. Previously the three-argument form always fell back to a full scan. Closes #115275. #115348 (Groene AI).
• Fix an exception in a window function over a Tuple with a sparse element, for example groupConcat(t) OVER (). #115705 (Alexey Milovidov).
• Conversion between named tuples with disjoint element names is now positional instead of silently filling the result with default values, and aliases that only give names to tuple elements can now repeat without error, for example SELECT tuple(1 AS x), tuple(2 AS x). #115717 (Alexey Milovidov).
• Cancel in-flight merges, mutations, and part fetches during graceful server shutdown instead of waiting for the running step to finish. Previously a single slow merge step or a large part download could delay shutdown for its whole duration. #115733 (Alexey Milovidov).
• Fixed a WHERE _table = ... or WHERE _database = ... condition over a Merge table silently returning no rows when the matching child table reads its data from other tables, for example a Distributed, Merge, Buffer, or Alias child. Such children are now always read, and the condition filters their rows. #116371 (Alexey Milovidov).
• Fix corrupted output when framing_output_format is used together with http_write_exception_in_output_format = 1 (the default for compatibility below 25.11). Data packets after the first one repeated the bytes of the first packet padded with zeros, and the query failed with std::length_error. This affected the Web UI, which reported a JSON parse error after the first few rows. #116620 (Alexey Milovidov).
• Fix links between settings, functions, and section headings on the built-in /docs page. References to other settings or functions did nothing when clicked, and # heading anchors put a corrupted section into the URL. #116635 (Alexey Milovidov).
• Fixed an exception when reading from a table after DETACH and ATTACH of a partition with patch parts created by lightweight updates. #116695 (Anton Popov).
• Fixed silently wrong results when reading a Parquet file whose offset index disagrees with the page data on an Array, Map, or nested column. Such a page is now rejected with INCORRECT_DATA instead of returning another page’s values under the requested row numbers. #116788 (Groene AI).
• Fix background cleanup of unused store/ directories on encrypted disks. ClickHouse can now revoke permissions and remove these orphaned directories instead of failing during cleanup. #116919 (Ivan Shelestov).
• Fix Native-protocol memory allocation from truncated input: server-declared sizes no longer make the client allocate memory for data that has not arrived; exception strings are read as they arrive, the decompressed block size is bounded, and the maximum number of rows in a Native block is now one billion. #117011 (Alexey Milovidov).
• Fix the exception max_rows > 0 (and possible incorrect merge results in release builds) in MergingSortedTransform when an input delivers a trailing empty chunk together with end-of-stream, for example under optimize_read_in_order. #117048 (Alexey Milovidov).
• Internal queries derived from an existing QueryPlan now preserve execution limits and shared resources instead of losing settings such as max_threads and concurrency control. #109722 (Groene AI).
• Fix wrong results when an Enum8 or Enum16 constant is compared with a String or FixedString column. The constant was converted to the enum’s underlying number instead of its name, so primary-key and PARTITION BY pruning, the bloom_filter skip index, IN and NOT IN sets, and the values table function could silently use the wrong value. Matching rows could be dropped, NOT IN could keep rows it should exclude, and IN could disagree with = on the same constant. #112647 (Groene AI).
• Fix wrong ORDER BY and DISTINCT results, and CANNOT_CONVERT_TYPE for aggregate functions, when reading a Merge table whose declared column type does not preserve the sort order of a source table’s type, for example UInt64 over a source holding negative Int64. #113242 (Groene AI).
• Fix min and max returning a wrong value when the implicit _minmax_count_projection answered them from a part-level bound that is not a real row value: NULL in a Nullable sorting key, NaN in a floating-point key, or a Tuple bound assembled component-wise. Such a query now reads rows instead of using the projection. #115151 (Groene AI).
• Fixed a settings-constraint bypass: a SETTINGS clause inside a subquery, a CTE, or a view’s inner query was applied without checking the settings constraints, so it could leave readonly mode or override a CONST, MIN, or MAX constraint such as additional_table_filters, breaking row-level isolation. Nested clauses are now clamped to the constraints instead. #115397 (Itamar Tempelhof).
• Fix enable_adaptive_aggregator with max_rows_to_group_by in throw mode: adaptive aggregation no longer silently falls back to the regular algorithm, and the merged total is now checked against the limit in parallel two-level aggregation, including bucket Top-K and spill-to-disk paths. #116451 (Nihal Z. Miaji).
• Fixed wrong results and an exception in primary-key analysis for intDiv and divide over a Decimal key with an integer constant divisor that does not fit the decimal’s native width. Such a query could return no rows at all instead of the matching ones. #116608 (Groene AI).
• Column names such as inf, infinity, nan, true, and false are now quoted when formatted, so a sorting key that refers to them no longer writes metadata the server cannot parse on the next start. Previously the table, and with it the whole database, failed to load. #116802 (Alexey Milovidov).
• Fix a server crash when creating a table whose columns form a long chain of DEFAULT expressions, such as c1 DEFAULT c0 + 1, c2 DEFAULT c1 + 1, .... Planning such a statement no longer exhausts the thread stack. #116978 (Groene AI).
• Fixed a ReplicatedMergeTree replication queue getting stuck forever after a new replica was added by cloning an existing one. The queue could contain two ALTER_METADATA entries for the same ALTER, so subsequent ALTERs waited indefinitely and SYSTEM SYNC REPLICA timed out. #117107 (Alexey Milovidov).
• Detect a column-count mismatch between a tuple and a subquery in the IN clause during analysis, preventing constant folding from silently hiding the error. #97540 (Alexey Milovidov).
• Fix LOGICAL_ERROR (Temporary part ... already added) when OPTIMIZE ... DRY RUN runs concurrently with a real merge, or with another DRY RUN, on the same parts. #110916 (Groene AI).
• Fixed a server startup crash when the background updater for dictionaries and executable user-defined functions applied an empty configuration before the first configuration repository was registered. #113238 (Groene AI).
• Fixed a data race and a use-after-free on the metadata of DeltaLake and Iceberg tables. Concurrent queries reading total_rows or total_bytes from system.tables for the same table could crash the server. Closes #113265. #113427 (Groene AI).
• Fixed arrayROCAUC and arrayAUCPR, and their aliases arrayAUC and arrayPRAUC, returning a different result for the same data depending on the order of the array elements when a score was NaN. A NaN score now makes the result NaN. Closes #104763. #115101 (Groene AI).
• Fix RESTORE of an incremental S3 backup chain whose locator authenticates by assuming a role with extra_credentials(role_arn = ...). The role identifiers are now kept in the <base_backup> locator of the backup metadata, so each backup in the chain can reopen its base; previously chains deeper than one level could not be restored at all. Chains already written without them can still be restored with use_same_s3_credentials_for_base_backup = 1, except when the locator is a named collection. #116223 (Julia Kartseva).
• Fixed missed insert deduplication when a block reaches the sink with sparse column representation, for example INSERT ... SELECT from a part with sparse serialization. Such an insert did not deduplicate against the same data inserted seconds ago from unmerged parts. #116287 (Mikhail f. Shiryaev).
• Functions h3PolygonToCells and h3PolygonToCellsWithContainment now check for query timeouts and KILL QUERY while covering a polygon, and reject vertices whose longitude is outside -180..180 or latitude outside -90..90 degrees. Previously a polygon with such a vertex could keep a thread busy for tens of minutes, ignoring both. Closes #116195. #116335 (Raúl Marín).
• Fixed the exception ... to have 0 rows, but it has 1: while executing function materialize when a dictionary with LAYOUT(DIRECT()) is read through the merge table function. Closes #116269. #116378 (Groene AI).
• Fix a bloom_filter skip index returning fewer rows for IN on an Array column when the set contains an empty array, for example WHERE x IN ([]). Closes #116431. #116449 (Groene AI).
• Fixed wrong results from search functions (hasPhrase, hasToken, hasAnyTokens, hasAllTokens) on a text index whose preprocessor or postprocessor uses IN or NOT IN, such as if(token IN ('the'), '', token). The transform was silently ignored whenever the predicate was evaluated on rows instead of read from the index, so matching rows could be missing. Closes #116476. #116578 (Elmi Ahmadov).
• Fixed a 404 (S3_ERROR) when reading an Iceberg table through a path that is a parent of the table directory, with the metadata document named by the iceberg_metadata_file_path setting. Writes and maintenance operations on such a table are now refused instead of acting outside the table. Closes #116985. #117037 (Groene AI).
• Fixed incorrect results from sum(if(cond, 0, 1)) when cond has type LowCardinality(Nullable(...)) and optimize_rewrite_sum_if_to_count_if is enabled, which is the default. Rows where the condition was NULL were not counted. Closes #116938. #117046 (Pratham Nayak).
• Fixed clickhouse-client and clickhouse-local not emitting the --chime alert for a long-running query that ends with an error. The threshold was checked against the query time last reported by the server, which is not refreshed when a query fails, so it could be arbitrarily behind the real duration of the query. It is now checked against the time the client actually waited. #117179 (Alexey Milovidov).
• Fixed a server crash when executing ALTER TABLE ... MATERIALIZE TTL on a table that is not a MergeTree but whose metadata carries a TTL. This can affect Alias tables whose target has a TTL and MaterializedView tables that declare a TTL on one of their own columns. Closes #117327. #117181 (Groene AI).
• Fixed a LOGICAL_ERROR in KQL timespan parameter handling for LowCardinality interval arguments. #117162 (Groene AI).
• Fix temporary_files_buffer_size and format_avro_schema_registry_* accepting out-of-range values that then made every subsequent query in the session fail with no way to recover. Such values are now bounded when set. Closes #113782. #113787 (Pedro Ferreira).
• Fix allow_feature_tier rejecting merge_tree_-prefixed settings in settings profiles as unknown and rejecting CREATE TABLE because of MergeTree settings applied by compatibility or the merge_tree config section. Also fix constraints on a merge_tree_-prefixed setting being bypassed when the setting is written through its alias, and a statement that rewrites such a setting to its current value being refused by a CONST constraint. #113156 (Raúl Marín).
• Fix a server crash and out-of-bounds write when inserting into a column with a long chain of compression codecs, for example CODEC(FPC, FPC, ..., FPC). Such chains are now rejected with CANNOT_COMPRESS; chains longer than 255 codecs, which previously could write unreadable parts, are rejected up front with BAD_ARGUMENTS unless allow_suspicious_codecs is set. #113877 (Groene AI).
• Fix OpenSSL cipherList being silently ignored when its value is written on its own line, as XML or YAML auto-formatters do. The configured value is now trimmed before it reaches OpenSSL, and an invalid cipher list is reported instead of discarded, so the server no longer keeps OpenSSL’s default cipher list while appearing to honor the configured restriction. Closes #115808. #116053 (Groene AI).
• Fix a server crash in the ReplicatedMergeTree cleanup thread when max_replicated_logs_to_keep = 0. That value is now rejected, matching the documented domain of the setting. #117147 (Groene AI).
• Fix path traversal in RESTORE: backup metadata entries whose <name> is not already normalized are now rejected with INSECURE_PATH instead of being restored outside the part directory while RESTORE reports success. #117371 (Groene AI).
• Fixed memory leaks of abandoned nested aggregate states. #113926 (Groene AI).
• Preserve row counts above the signed 32-bit range on the PostgreSQL wire protocol. #114397 (Minh Vu).
• Fixed ILLEGAL_COLUMN for ARRAY JOIN queries filtered by an empty IN set, including an empty list and an empty Set table. Closes #116018. #116057 (Groene AI).
• Fixed Alias tables ignoring row policies defined on their target tables. An Alias table now combines the row policies of both the alias and its target table. #116290 (Kai Zhu).
• Fixed concurrent writes of the same format_schema_source = 'string' or 'query' schema being rejected with CANNOT_OPEN_FILE or ATOMIC_RENAME_FAIL. #116502 (Groene AI).
• Fixed moduloOrNull, intDivOrNull, divideOrNull, and positiveModuloOrNull returning the raw division result instead of NULL when dividing by a constant zero inside a subquery whose result is read or persisted by an outer query. Closes #116500. #116594 (Groene AI).
• Fixed Iceberg partition pruning when a manifest’s partition tuple contains fields ClickHouse cannot model. Such tables no longer fail with STD_EXCEPTION, and a manifest whose tuple length disagrees with its own spec is now rejected with ICEBERG_SPECIFICATION_VIOLATION instead of returning deleted rows. #116615 (Groene AI).
• Fixed wrong results from sumKahan and sumKahanIf when the aggregated batch does not start at the first row of the block, including window frames, parallel GROUP BY, and optimize_aggregation_in_order = 1. #116990 (Groene AI).
• Fixed GenerateRandom reads with a trivial LIMIT returning no rows because the source-count reduction could overflow to zero. Absurdly large requested stream counts are now also rejected with PARAMETER_OUT_OF_BOUND. #117171 (Groene AI).
• Fixed out-of-bounds access in timeSeriesGroupArray. #117404 (Vitaly Baranov).
• Fixed GROUP BY emitting a group from a zero-row block, returning a row with count() = 0. Closes #116830. #117503 (Yarik Briukhovetskyi).
• Fixed arrayIntersect with mixed Nullable and not-Nullable arguments. Such calls could fail or return spurious default-valued elements for some data types; arrayUnion and arraySymmetricDifference are unaffected. #117577 (Groene AI).
• ATTACH PARTITION FROM, REPLACE PARTITION, and MOVE PARTITION TO TABLE no longer accept a non-adaptive part whose index_granularity differs from the destination table’s. Previously such a part could return wrong row counts in release builds. Closes #117524. #117594 (Groene AI).
• Fixed max_joined_block_size_rows = 0 (documented as unlimited) making a hash join emit one left row per output block instead of honoring max_joined_block_size_bytes when a left row has multiple right matches. Closes #102007. #117604 (Groene AI).
• Fixed hasColumnInTable disclosing whether a column exists on the target of an Alias table when the user lacks SHOW COLUMNS on the target. The function now requires that privilege on the target too. #117621 (Groene AI).
• Fix silent data loss when a TTL expression applies a large interval directly to a Date or DateTime column by widening the arithmetic away from 16/32-bit overflow. #101793 (BoloniniD).
• Fix Cannot insert field with type ... when inserting a row that omits an AggregateFunction column whose argument types contain LowCardinality. defaultValueOfTypeName is fixed for the same type names. #102403 (Enric Calabuig).
• Fixed a privilege-escalation gap where the executable table function did not enforce the TABLE ENGINE ON Executable grant. With table_engines_require_grant enabled, a user denied CREATE TABLE ... ENGINE = Executable could still run the same server-side script via SELECT ... FROM executable(...). It now requires GRANT TABLE ENGINE ON Executable, matching CREATE TABLE. #109331 (Groene AI).
• Fix silently wrong GROUP BY results with optimize_aggregation_in_order = 1. Groups were merged into one another and aggregate values were wrong when the matched sorting-key prefix contained a reverse (DESC) column, and, independently of reverse keys, when a group-by key was a negative monotonic function of a Float sorting-key column around the NaN boundary. This affected a single MergeTree table and a Merge table over such tables. For a Merge table whose underlying tables read the matched prefix in opposite physical directions, aggregation in order is now declined because a single stream direction cannot describe both. #112200 (Groene AI).
• Fix _path / _file filtering for object storage, file, and url reads. A GLOBAL IN subquery over _path or _file of a path without globs no longer throws Not-ready Set is passed as the second argument for function 'globalIn'; an explicitly named archive member is filtered before the archive is opened; Hive partition columns are inferred for an explicitly named key that does not exist; and urlCluster / fileCluster now prune a filtered-out path on the initiator instead of handing it to a replica. #112968 (Alexey Milovidov).
• FileLog tables no longer silently skip files that appear after CREATE TABLE. The directory watch ran as a BackgroundSchedulePool task but never returned, so it held its slot of the per-task-type budget forever; once background_schedule_pool_size times background_schedule_pool_max_parallel_tasks_per_type_ratio tables existed, further tables never watched their directory at all and SELECT returned fewer rows instead of erroring. The watch now runs on its own thread. #113897 (Groene AI).
• A query queued on a full distributed connection pool now responds to KILL QUERY, and to max_execution_time at the default timeout_overflow_mode = 'throw', instead of waiting until another query happens to release a connection, and a positive connection_pool_max_wait_ms now actually expires. Previously the wait ended only on success at every value of the setting, so the configured timeout was ignored, max_execution_time was reported long after its deadline, and the query could not be killed. The wait applies per replica, so an expired one is a soft failure that fails over to the remaining replicas and only then reports ALL_CONNECTION_TRIES_FAILED. #113979 (Groene AI).
• Fix wrong results when a filter on a JSON subcolumn that mixes a literal prefix with a quoted path segment is analyzed against a JSONAllPaths skip index. The subcolumn name was misparsed, so all granules were skipped and the query returned fewer rows than it should. #114873 (Pavel Kruglov).
• Fix wrong results when a SELECT with select_sequential_consistency = 1 on a ReplicatedMergeTree table runs with classic (task-based) parallel replicas. Every replica performed a complete local read clamped to the quorum boundary instead of a coordinator-assigned slice, so rows were returned once per participating replica. #114945 (Groene AI).
• Fixed CREATE DICTIONARY failing with Incorrect type of ID column: must be UInt64, but it is LowCardinality when the simple primary key is declared as LowCardinality(UInt64) and dictionary_validate_primary_key_type is enabled. Closes #104539. #115139 (Groene AI).
• Fix a hang in workload memory reservations when an increase races with an in-flight decrease. #116085 (Azat Khuzhin).
• A Kafka table with kafka_keeper_path set (StorageKafka2) no longer skips messages it had already consumed when a streaming cycle fails part-way through; previously those messages were never delivered to the materialized views and a later cycle committed an offset past them. As a consequence, with the default kafka_handle_error_mode an unparseable message now blocks the table instead of being silently skipped, matching StorageKafka. kafka_handle_error_mode = 'stream', kafka_handle_error_mode = 'dead_letter_queue', and kafka_skip_broken_messages remain the ways to skip such a message. #116485 (Groene AI).
• Fixed the Unexpected return type from materialize logical error when reading a column that has a text index through the index after a metadata-only or not-yet-materialized ALTER MODIFY COLUMN (including lazy JSON type hints) that requires a read-time type conversion. #116549 (Pavel Kruglov).
• A table created with CREATE TABLE ... AS <table function>(...) is no longer resolved to answer a SYSTEM STOP-family lock query or a system.tables size query about it. Previously those statements contacted the endpoint of every such table, so one unreachable endpoint made them fail or stall. Until such a table is first accessed it now reports NULL sizes and an empty data_paths, and DROP TABLE ... IF EMPTY removes it instead of reporting TABLE_NOT_EMPTY. #116774 (Groene AI).
• Fixes wrong results when a minmax skip index, ORDER BY, or PARTITION BY is built over an Array, Tuple, or Map holding a Bool, such as Array(Tuple(Float64, Bool)) or Map(String, Bool), and when a minmax index serves an ORDER BY ... LIMIT over a Bool column. Previously, rows that the predicate accepts could be dropped, because the stored bound and the value compared against it represented the boolean differently. #116897 (Groene AI).
• Fixed an external GROUP BY exceeding max_memory_usage when enable_adaptive_aggregator is on. Intermediate data the adaptive aggregator held and could not release kept a thread on the baseline algorithm flushing its own hash table on every block, so each block wrote a temporary file and the merge then allocated a reader per file. That data is now written out first. #117326 (Groene AI).
• Malformed Native blocks can no longer overflow the allocation of aggregate-function states during deserialization and cause out-of-bounds writes. #117370 (Raúl Marín).
• Fixed the byte width of the Iceberg manifest lower_bounds and upper_bounds written for int and date columns: they were 8 bytes wide where the Iceberg specification requires 4. Spec-conformant readers such as pyiceberg failed with unpack requires a buffer of 4 bytes while planning any scan carrying a filter on such a column. #117571 (Groene AI).
• Remote-write async insert flush timeouts are now handled correctly: they return retryable 503 Service Unavailable instead of 408 Request Timeout, so Prometheus-compatible senders can retry them, and wait_for_async_insert_timeout is now clamped instead of overflowing for large magnitudes. #117632 (Vitaly Baranov). #118040 (Groene AI).
• Fixed base58Encode and base58Decode of 32- and 64-byte values on big-endian builds such as s390x. The fixed-size codec byte-swapped every 4-byte word unconditionally, so on a big-endian host base58Encode returned a string that disagrees with the base58 specification and with every other architecture, and base58Decode rejected well-formed input when given a 32- or 64-byte size hint. Little-endian builds are bit-identical to before. Strings previously produced by a big-endian build will change, because they were wrong. #117649 (Groene AI).
• Fixed format_schema_source = 'query' running the schema query with full privileges when used from a background task, such as a streaming engine consumer (Kafka, NATS, RabbitMQ, ObjectStorageQueue) or a Buffer background flush. The schema query is now rejected there; use format_schema_source = 'file' or format_schema_source = 'string' instead, or pre-populate the schema cache. #117737 (Pavel Kruglov).
• Fix absent and scalar PromQL queries returning no result when empty_result_for_aggregation_by_empty_set is enabled. #117816 (Valery Petrov).
• A bloom_filter index over an Array(Nullable(T)) column is now rejected at CREATE / ALTER time. It used to be accepted and then made every insert, merge, and mutation of the table fail, with DROP INDEX refused until the failed mutation was killed. #117899 (Alexey Milovidov).
• Fix the GROUPING function rejecting constant aliases that are valid GROUP BY keys. #97504 (Héctor Pablos).
• Fixed SQL injection in the PostgreSQL wire protocol emulation. Prepared-statement parameters, in both the extended-query Parse / Bind / Execute path used by psycopg2, JDBC, node-postgres, Npgsql, and others, and the simple-query EXECUTE path, were substituted into the executed SQL by raw string splicing, so a crafted parameter could break out of its value position and inject SQL. Parameters are now bound as data. This also fixes string parameters that previously failed with UNKNOWN_IDENTIFIER when a client sent them unquoted. #108469 (Groene AI).
• Fix a LOGICAL_ERROR (Part ... intersects previous part ... It is a bug or a result of manual intervention) that makes a MergeTree table impossible to load: attaching it, and every later attempt to load it, fails with that exception, and debug and sanitizer builds abort. A merge could span a block gap holding an outdated part whose mutation version is higher than the merge result’s, producing a pair of parts on disk that neither contain nor are disjoint from each other. Closes #113864. #113941 (Groene AI).
• Fix a LOGICAL_ERROR: Next task callback is not set for query with id exception when parallel_replicas_plan_based is enabled and the query reads a system log table. #114022 (Groene AI).
• Fix files_read and bytes_read in system.backups and system.backup_log being permanently undercounted after a successful RESTORE, because a concurrent restore task holding an early snapshot could overwrite a larger count published by another task. #114304 (Groene AI).
• Fix wrong results when a non-boolean condition, such as a bare integer column in WHERE, is left alone after the other conditions are merged into the JOIN condition or pushed down. Rows whose condition value was a non-zero multiple of 256 were skipped. #114533 (Vladimir Cherkasov).
• Fixed a single ALTER with two MODIFY COLUMN subcommands on the same column silently corrupting that column’s default: a default set by one subcommand was dropped, and a REMOVE DEFAULT / MATERIALIZED / ALIAS was undone, whenever a following subcommand restated the column type. #114807 (Groene AI).
• You can now access nested properties on a Nullable(JSON) value, for example col::Nullable(JSON).Id. Previously this failed with First argument for function tupleElement cannot be Nullable(JSON), which made it impossible to combine a cast to Nullable(JSON) with property access on the result. Outer NULL rows now propagate through extracted paths using the existing subcolumn semantics: paths whose type can represent NULL return NULL, while types such as Array and Map return their default values. #114863 (Luís Lizardo).
• Fix an integer overflow in the deserialization of String values in the binary formats, and bound the recursion of the binary deserialization of Dynamic and JSON values. #115702 (Alexey Milovidov).
• PostgreSQL compatibility now handles pg_catalog qualifiers case-insensitively like PostgreSQL does, and the synthetic OIDs exposed through pg_namespace and pg_class are now unique, so psql commands such as \d no longer duplicate tables or list them under the wrong schema. #115960 (Alexey Milovidov).
• Fixed data loss in MergeTree when a merge runs while an ALTER TABLE ... RENAME COLUMN is still pending. The merge applied the rename but did not record it, so the mutation applied it a second time and read the renamed column as missing, filling it with defaults. Columns without a default expression, such as Dynamic, lost their values. #115981 (Alexey Milovidov).
• Fixed a query over an Iceberg table returning fewer rows than the table holds when a manifest declares a lower bound above the upper bound for a filtered column. Such a pair describes an empty range, so min/max pruning dropped the data file holding the matching rows; those bounds are now ignored instead. Closes #116499. #116604 (Groene AI).
• Fixed two Binary Viewer (/binary) bugs: the image was not centered on load if the URL contained parameters such as user, and the connection parameters (host, user) were lost from the URL after moving the view. The password parameter is no longer kept in the address bar or browser history. #117065 (Alexey Milovidov).
• Fix a large performance regression, present since 26.5, for a query that filters on a column with an equality and orders by a later column of the same sorting key. Read-in-order analysis did not look through an ALIAS node on the filter column, so the filtered column was not recognized as constant and the read was not done in order, forcing a full sort of every matching row before the LIMIT applied. This reaches a table with a normal PROJECTION, whose rewrite wraps the query filter in such an alias, and a table-level ALIAS column used as a predicate. #116854 separately made a base table that can serve the ORDER BY itself preferred; this fixes the case where only the projection can. #116835 (Groene AI).
• Fixed wrong results when reading the .null subcolumn of a column that went through a metadata-only T to Nullable(T) ALTER and then had a RENAME COLUMN still pending: IS NULL, IS NOT NULL, count(), and an explicit x.null read reported rows holding real data as NULL. #116901 (Groene AI).
• Fixed a stack overflow when reading hadoop-snappy data containing many blocks that decode to nothing. #116994 (Alexey Milovidov).
• The handshake response of the MySQL wire protocol is now bounded, so a client can no longer make the server grow memory during the handshake. The mysql_require_secure_transport server setting is now enforced from the actual state of the connection instead of a capability bit advertised by the client. #117003 (Alexey Milovidov).
• Fix background Distributed inserts that deleted unsent spool files after a split batch got a transient error. ClickHouse now keeps the files and retries the batch. #117034 (Rory Shanks).
• Fix wrong results for GROUP BY and uniq* over MACNumToString, unhex, unbin, and UUIDStringToNum. All four declared themselves injective, so the default-enabled optimize_injective_functions_in_group_by and optimize_injective_functions_inside_uniq grouped and counted by the raw argument: arguments that map to the same result came back as separate groups and were counted separately by uniqExact. #117210 (Alexey Milovidov).
• Fix wrong results from the default-enabled optimize_uniq_to_count. uniq(x) over a SELECT DISTINCT x or GROUP BY x subquery reported one distinct value too many when the column contained a NULL, and counted the extra super-aggregate rows of WITH ROLLUP / WITH CUBE as additional distinct values. #117222 (Alexey Milovidov).
• Fix NO_SUCH_COLUMN_IN_TABLE for every query using a text index while a mutation is pending with apply_mutations_on_fly = 1. Until the mutation materialized, hasToken, hasAnyTokens, and hasAllTokens on the indexed column could not be executed at all. #117277 (Alexey Milovidov).
• Several aggregate function deserializers allocated memory from a size declared in the serialized state before reading the payload, so a state of a few bytes could request gigabytes. They now read only what the buffer already holds. #117536 (Groene AI).
• getServerSetting, getMergeTreeSetting, and getMacro returned the values of system.server_settings, system.merge_tree_settings, and system.macros to any user, bypassing the SELECT grant that reading those tables requires. They now require the same grant. #117551 (Raúl Marín).
• Fixed a bug where the analyzer_inline_views setting could bypass the SELECT access check on a view, letting a user without a grant on the view read its contents. #117555 (Dmitry Novik).
• ENGINE = EmbeddedRocksDB(ttl, rocksdb_dir, read_only) with an explicit rocksdb_dir now requires the FILE source grant that file requires for the same directory: READ ON FILE to create such a table, plus WRITE ON FILE when read_only = 0. Previously it required no source grant, so a user holding only CREATE TABLE in a database of their own could point a table at another table’s directory and read its rows. ON CLUSTER and Replicated-database DDL are not covered: the replay runs with no user. The argument-less form is unaffected. Backward incompatible for anyone already using an explicit rocksdb_dir without them. #117428 (Groene AI). #117665 (Pedro Ferreira).
• Fix ALTER and CREATE TABLE ... AS dropping the other settings of a TimeSeries table. #117693 (Vitaly Baranov). #117839 (Vitaly Baranov). #118119 (Alexey Milovidov).
• Fix incorrect results for suffix LIKE patterns with escape sequences when optimize_rewrite_like_perfect_affix is enabled. Such patterns are now rewritten to endsWith only when the suffix is an exact literal. #117712 (Alex Prabhat Bara).
• Fix a dictionary with a Redis source and an Int256 key column silently returning the default value from dictGet for keys that exist in Redis. Wide integer key columns are now refused with a clear error instead of being admitted and never matching. #117735 (Pratham Nayak).
• Fixed excessive memory usage when reading from File, URL, object storage, GenerateRandom, or any other storage that uses the generic read path with max_streams_to_max_threads_ratio above 1. The number of output ports created after a read is now bounded by max_threads, so a SELECT over a tiny table no longer allocates tens of gigabytes while the pipeline is still being built. #117772 (Alexey Milovidov).
• Fix ARRAY JOIN producing rows in incorrect order in some cases. #117918 (Yarik Briukhovetskyi).
• Fix a potential out-of-bounds write in formatDateTimeInJodaSyntax when called with a time zone name longer than 32 characters. #117925 (Robert Schulze).
• Fix EXPLAIN QUERY TREE and EXPLAIN SYNTAX exposing column names and types of tables the user has no privileges for: the analyzer now checks that the user may see at least one column of the table at the moment a table identifier is resolved. The check also makes the error kind independent of whether the queried column exists and stops recording column names the user may not see into missing_privileges of the query log. Closes #78938. #117979 (Alexey Milovidov).
• Fixed Keeper aborting when memory usage stays near the limit: moving finished changelogs and snapshots to object storage was refused, so the local log disk filled up. #117983 (Shaohua Wang).
• Fixed a Kafka / RabbitMQ / NATS / FileLog / S3Queue / AzureQueue table in a database with lazy_load_tables silently stopping to feed its materialized views after a server restart. #118009 (Alexey Milovidov).
• Fixed cancellation of a synchronous distributed INSERT (distributed_foreground_insert) waiting for that INSERT to finish instead of interrupting it. KILL QUERY against such an insert, a client disconnect, and server shutdown could block for as long as a shard was slow. #118013 (Groene AI).
• Fixed a LIMIT being pushed into numbers() when the query’s WHERE referenced an arrayJoin through a WITH alias, which could make the source stop before the expansion produced enough rows. #118045 (Yarik Briukhovetskyi).
• Fix Iceberg schema evolution for additional nested Array, Map, and Struct combinations whose inner fields are added, removed, or reordered. #118127 (Konstantin Vedernikov).
• Fixed incremental backups to Azure Blob Storage storing the beginning of each file instead of its tail. Restoring such a backup silently dropped all the data written after the base backup. #114975 (Smita Kulkarni).
• Fixed potential data loss when an INSERT into ReplicatedMergeTree was cancelled or timed out while ClickHouse was recovering an uncertain Keeper commit. #111196 (kakao-tars-case).
• Fixed object storage listing silently skipping objects when S3 returns an empty page together with a continuation token. This could make a plain_rewritable disk appear empty and hide table parts, and could make data lake tables read a stale snapshot. #118615 (Sema Checherinda).
• Fixed a SQL injection in the PostgreSQL metadata queries. The schema of the PostgreSQL database engine, the requested column names and the publication name were formatted into SQL literals without escaping, so a value containing a quote could execute arbitrary SQL on the connected PostgreSQL server as the role ClickHouse connects with. Reaching it through schema needed only CREATE DATABASE plus use of a named collection whose schema is overridable. #111516 (Shaohua Wang).
• Hide the credentials that a setting value can carry, such as the password in format_avro_schema_registry_url or url_base, everywhere a setting value is displayed or logged. system.query_log, system.processes, system.settings, system.settings_profile_elements, system.session_log, system.distributed_ddl_queue, SHOW CREATE USER and SHOW CREATE SETTINGS PROFILE previously showed them in plaintext to any user who could read them. #117523 (Raúl Marín). EXPLAIN now respects format_display_secrets_in_show_and_select as well. #118129 (Kirill Kopnev).
• CREATE USER OR REPLACE and the OR REPLACE form of CREATE ROLE, CREATE QUOTA, CREATE SETTINGS PROFILE and CREATE ROW POLICY now require the corresponding DROP privilege in addition to CREATE, because they overwrite an existing entity. Previously CREATE USER alone was enough to reset the password of any other user, including a privileged one. #117960 (Alexey Milovidov).
• Fix CREATE OR REPLACE TABLE/VIEW/DICTIONARY and REPLACE TABLE requiring privileges on the internal temporary table they create (_tmp_replace_*). Its name is random, so no grant could cover it and the query failed with ACCESS_DENIED unless the user held wildcard grants on the whole database. The required privileges are now checked against the user-visible names only: replacing an object needs the CREATE and DROP grants of its own kind, plus INSERT on it when the statement populates it. #117967 (Alexey Milovidov).
• TRUNCATE DATABASE ... ON CLUSTER and TRUNCATE [ALL] TABLES FROM ... ON CLUSTER ran without checking any privilege: a user holding only SELECT and CLUSTER could destroy every table in a database, and readonly = 1 did not prevent it. Both now require DROP DATABASE, the privilege the non-cluster spellings have always required. #118155 (Groene AI).
• Fixed a privilege escalation in implicit access-rights computation: holding any single child privilege of the ALTER TABLE group (for example ALTER MODIFY COMMENT) on a view implicitly conferred the entire ALTER VIEW group, so such a user could run ALTER TABLE <view> MODIFY QUERY and rewrite the view’s query. ALTER VIEW is now implied only by the complete ALTER TABLE grant. A deployment that relied on a partial ALTER TABLE grant to alter views must now grant ALTER VIEW (or one of its members) or the full ALTER TABLE explicitly. #118194 (Groene AI).
• Expressions supplied through the additional_table_filters and parallel_replicas_custom_key settings are now subject to the column-level SELECT check, and are no longer injected into the body of SQL SECURITY DEFINER or NONE views. Previously a user could evaluate predicates over columns they were not granted. #118063 (Dmitry Novik).
• Forbid CREATE TABLE ... AS viewIfPermitted(...), including viewIfPermitted nested in another table function such as remote, which now fails with BAD_ARGUMENTS, as it already does for eval and the *Cluster table functions. The function chooses between its SELECT query and the ELSE function according to the current user, but a persisted table is resolved with no user and therefore full access: reading one reported the source table’s column names and types to a user with no grant on it, and the ELSE fallback was unreachable. Tables created before this change keep loading; drop them to close the disclosure. #116885 (Groene AI).
• Fixed a Filesystem database serving a table from its internal cache without the READ ON FILE grant. The grant was checked only while a table name was resolved for the first time, so once any user holding it had read a name, every user with SELECT on the database was served the contents of the file. EXISTS TABLE, which requires only SHOW TABLES, answered from that cache and probed the directory without the grant too. Both now require it, and a user without it is refused before the existence of the file is observed. #119029 (Groene AI).
• Apply remote_url_allow_hosts before SQL-supplied S3 and Azure endpoints send schema, metadata, or container requests. #114357 (Rory Shanks).
• Fixed the server deserializing the payload of a Data packet that arrives on a connection before any query. In interserver mode the connection is not authenticated until the Query packet is processed, so such a packet was rejected only after its Native block had been read, which let an unauthenticated peer choose the column type the server constructed. The packet is now rejected before its payload is touched. #118588 (Groene AI).
• Each code for a time-based one-time password is now accepted at most once, as required by RFC 6238: reusing an already used code, or a code for an earlier time step, is rejected. A user with time_based_one_time_password can successfully authenticate at most once per period. #117563 (Vladimir Cherkasov).
• Fixed wrong results when join_algorithm = 'parallel_hash' is combined with a sorted consumer such as optimize_aggregation_in_order, optimize_distinct_in_order or LIMIT BY. With several join slots and a single-level hash map the join does not keep the order of the left table, so the read-in-order-through-join optimization must not advertise the left sort order in that case. #109225 (Groene AI).
• Fixed a LOGICAL_ERROR (Join is supported only for pipelines with one output port) raised by a merge JOIN when its two sides reached the join with a different number of output ports. #109393 (Groene AI).
• Fix wrong results when a doubly negated expression is compared with a constant, such as WHERE (NOT NOT c) > -0.5. Index analysis cancelled the two NOTs and compared the column itself with the constant, pruning parts, granules and partitions that do match. #116381 (Alexey Milovidov).
• Fixed wrong results after ALTER TABLE ... MODIFY ORDER BY on a table with a descending sorting key column (ORDER BY a DESC). The ALTER silently dropped the DESC direction while existing parts stayed physically sorted the old way, so primary key index analysis pruned the wrong granules and queries returned missing rows. Such an ALTER is now rejected. #116473 (Groene AI).
• Fix the Invalid status NotActive for associated output logical error that could happen for queries with a high max_threads, for example a GROUPING SETS aggregation over a remote table. #116478 (Alexey Milovidov).
• Fixed reading uninitialized memory as literal data when decompressing a malformed LZ4-compressed block. #116992 (Alexey Milovidov).
• Fix silently missing rows for a predicate on toMillisecond, toMicrosecond or toNanosecond of a DateTime64 primary key. All three claimed to be monotonic everywhere, although they restart at zero every second, so primary-key analysis treated the predicate as one contiguous key range and pruned granules that contain matching rows. #117246 (Alexey Milovidov).
• Fix wrong row counts from the default-enabled query_plan_merge_filter_into_join_condition when an equality conjunct of the WHERE filter contains a non-deterministic function such as rand(). The conjunct was moved into the join condition, where it is evaluated once per join input row instead of once per output row, and cloned into the build-side runtime filter as a second evaluation site. #117247 (Alexey Milovidov).
• Fix the default-enabled query_plan_merge_expression_into_join returning rows that violate the query’s own JOIN ON condition when a subquery computes a non-deterministic expression that is then used as a join key. The merged expression was computed twice from the raw inputs, drawing independently for the match and for the output column. #117250 (Alexey Milovidov).
• Fix the default-enabled use_join_disjunctions_push_down silently dropping rows when the filter above a join hides a non-deterministic function inside a lambda body, as in arrayExists(z -> rand(z) % 2 = 0, [a]). The extracted pre-filter was pushed below the join while the original filter stayed on top, so the lambda was drawn twice per row and only rows passing both draws survived. #112700 (Groene AI).
• Fixed a filter containing a non-deterministic function being moved below ARRAY JOIN, which evaluated it once per source row instead of once per expanded row and made whole arrays pass or fail together. #117756 (Alexey Milovidov). Fixed rand, generateUUIDv4 and other non-deterministic functions being evaluated before an arrayJoin replicates the rows, which silently duplicated their values across the expanded rows. #118963 (Alexey Milovidov).
• Reject a CHECK constraint whose expression contains arrayJoin. Such a constraint was accepted and then read past the end of the inserted block, surfacing as a nonsensical Array of size 18446744073709551613 is too large, or an abort in a debug build, or blamed a violation on the wrong row. #117267 (Alexey Milovidov). arrayJoin in a table constraint is now also rejected when the definition is copied by CREATE TABLE ... AS src or CLONE AS src, stated by a full-definition ATTACH, or declared as ASSUME. A constraint containing it that is already stored in the metadata of a table is no longer used by the query-time constraint optimizer, which could remove a WHERE that the constraint did not really imply. #117799 (Alexey Milovidov).
• Reject a TTL expression that contains arrayJoin. Such a TTL was accepted and then silently deleted rows whose own TTL was far in the future, and read past the end of the expression column when an array was empty. #117266 (Alexey Milovidov).
• Fix a silent result change for a query over a trivial view over a Distributed table whose WHERE or view-keyed additional_table_filters contains a bare IN some_table predicate. The default-enabled optimize_trivial_view_pushdown_to_distributed shipped that predicate to the shards, where it was evaluated against each shard’s local table of that name instead of the initiator’s, so it disagreed with the IN (SELECT ...) spelling of the same predicate. #117278 (Alexey Milovidov).
• Fixed query_plan_aggregation_bucket_top_k silently dropping result rows when an arrayJoin sits between the aggregation and the ORDER BY ... LIMIT. #117464 (Alexey Milovidov).
• Fixed SELECT ... FINAL with a WHERE on the partition key expression returning a superseded row when the partition key is computed from a floating-point sorting-key column: partition pruning dropped the partition holding the deduplication winner. #117509 (Alexey Milovidov).
• Fixed mutations_sync = 3, lightweight_deletes_sync = 3 and alter_sync = 3 performing no wait at all on ReplicatedMergeTree, so a mutation or ALTER returned as if it had synchronised. As documented, the value now behaves the same as 2 and waits for all replicas. This also affects OPTIMIZE, TRUNCATE, DROP/DETACH PARTITION, REPLACE PARTITION and MOVE PARTITION TO TABLE, which take the same alter_sync wait path. Statements that previously returned immediately now wait, and can raise UNFINISHED when a replica is inactive; set the value to 0, 1 or 2 to choose a different wait explicitly. #117517 (Groene AI).
• Fixed the logical error Cannot add ZooKeeper operation because query is executed when running TRUNCATE TABLE on a TimeSeries table in a Replicated database. #117902 (Raúl Marín).
• Fix a lightweight UPDATE silently not applied to rows of a data part when the patch part is written in the v1 format (patch_parts_version = 'v1') and covers rows of several data parts. One row was lost when a data part’s patch rows began in a later granule of the patch part, which made the result depend on merge_tree_min_read_task_size and max_block_size. All updates for a data part were lost, on disk as well as at read time, when the patch part had no final mark (index_granularity_bytes = 0) and that part’s patch rows began in the patch part’s last granule. #117962 (Groene AI).
• Fix wrong results and a NOT_IMPLEMENTED exception when a lightweight update’s patch part is applied on top of a pending ALTER TABLE ... UPDATE col = <constant> with apply_mutations_on_fly = 1. The constant-folded column was not materialized before the patch was written into it, so the patch was either silently discarded or the read failed. #118545 (Shaohua Wang).
• Fixed system.tables.total_rows/total_bytes overcounting a non-replicated MergeTree table while a patch part from a lightweight UPDATE is alive, and SELECT count() returning that inflated number with apply_patch_parts = 0. #117463 (Alexey Milovidov).
• Fixed a query with a hasToken-family predicate on a text index silently returning no rows when the query also read a column modified by a pending lightweight update. #118496 (Shaohua Wang).
• Fix excessive patch-part reads for lightweight updates on tables with redundant parentheses in sorting-key expressions, such as ORDER BY (k). #118878 (Anton Popov).
• Fixed the LOGICAL_ERROR Reading from materialized CTE '...' before its materialization completed raised by a lightweight UPDATE, or a lightweight DELETE in lightweight_update_force mode, whose predicate contains an IN subquery that defines a materialized CTE, when the set is built at run time. #119489 (Dmitry Novik).
• Fixed a retried INSERT SELECT over a join not being deduplicated when the two attempts differed only in whether a column was carried as a lazily replicated column. #118004 (Alexey Milovidov).
• Fixed a logical error Cannot write to finalized buffer (an abort in debug and sanitizer builds) in the non-replicated MergeTree deduplication log when opening the next log file during rotation failed, for example after a transient I/O error on the deduplication log’s disk. Such a failure no longer fails the insert that triggered the rotation, so a retry of that insert is not wrongly deduplicated. #118089 (Alexey Milovidov).
• Fix silent data loss on upgrade in Distributed background inserts with background_insert_split_batch_on_failure: batch recovery no longer deletes a spool file that was never sent when a later file of the same batch was quarantined into broken/ by a previous server version. #118211 (Alexey Milovidov).
• Fixed SELECT count() returning too many rows for a negated comparison (!=, NOT IN, NOT LIKE) over a Nullable primary key column: the exact-count optimization counted the NULL-key rows that the WHERE condition rejects. #118214 (Alexey Milovidov).
• Fixed AMBIGUOUS_COLUMN_NAME on every read of a table that has a column named like a subcolumn of another column (for example a.size0 next to an Array column a) after ALTER TABLE ... MODIFY COLUMN changed that column’s type. #118228 (Alexey Milovidov).
• Fix unbounded memory growth from time zone names of the form Fixed/UTC±HH:MM:SS. Every distinct name permanently cost ~4.6 MiB that the server never gave back, and 172801 of them were accepted. Only offsets that a time zone can have are accepted now: a whole number of quarters of an hour, no further from UTC than 14 hours. #118286 (Alexey Milovidov).
• Fixed MULTIPLE_EXPRESSIONS_FOR_ALIAS under rewrite_in_to_join for a query that repeats one identical aliased IN (subquery) expression more than once. #118292 (Alexey Milovidov).
• A table created with CREATE TABLE ... AS <table function>(...) is no longer resolved to answer a system.columns size query about it. Previously selecting data_compressed_bytes, data_uncompressed_bytes or marks_bytes contacted the endpoint of every such table, so one unreachable endpoint made the whole scan fail. Until such a table is first accessed it now reports zero sizes. #118302 (Groene AI).
• Fixed an Inconsistent AST formatting logical error for an aliased IN expression whose left operand also has an alias when it is nested inside a function-argument list, e.g. SELECT ifNull(sum((x AS a1) IN [y] AS c3 = TRUE), 0). #110229 (Groene AI).
• Fix a logical error (server abort in debug/sanitizer builds) when a constant-folded comparison appears in the SELECT list together with group_by_use_nulls = 1 and GROUP BY GROUPING SETS. #110703 (Groene AI). Fixed a Bad cast error, and wrong results in release builds, when an aggregate reads a derived table that uses GROUP BY ALL with ROLLUP or CUBE under group_by_use_nulls = 1. #110921 (Groene AI). Fixed a return type mismatch error when group_by_use_nulls is used with GLOBAL IN in a GROUP BY key. #116980 (George Larionov).
• Fixed a LOGICAL_ERROR (“query tree node does not have valid source node”) when a Merge table wrapping a Distributed table is joined and a bare column of the joined-away side is used as a predicate. #111327 (Groene AI).
• Fixes a LOGICAL_ERROR (Column identifier ... is already registered), a MULTIPLE_EXPRESSIONS_FOR_ALIAS error and a NOT_FOUND_COLUMN_IN_BLOCK error for a query whose IN / NOT IN / GLOBAL IN right operand names a table expression that also appears in the FROM section, for example SELECT l.x FROM l JOIN (SELECT x FROM r) AS r ON l.x = r.x WHERE l.x IN r. Also makes ORDER BY <alias> WITH FILL ... INTERPOLATE (<alias> ...) inside a CTE or any other copied subquery raise INVALID_WITH_FILL_EXPRESSION, as the same query already does when it is not copied: the check was silently skipped for the copy, and such a query returned rows instead of the error. #112219 (Groene AI).
• Fixes NOT_FOUND_COLUMN_IN_BLOCK when query_plan_join_shard_by_pk_ranges = 1 is combined with join_algorithm = 'full_sorting_merge' and a join input is filtered on a sorting-key column that is not the join key and is not selected. Closes #112402. #112664 (Groene AI). Fixed rows being silently dropped from a join when query_plan_join_shard_by_pk_ranges is enabled and the primary key is a floating-point column containing NaN. #118918 (Alexey Milovidov).
• Fixed Code: 476 (Can not use WITH TIES without ORDER BY) raised by a remote shard, and the logical error LIMIT WITH TIES without ORDER BY raised on the initiator, when selecting from a Merge table over a Distributed table with a JOIN and ORDER BY ... LIMIT n WITH TIES. The child query kept its LIMIT clause after its ORDER BY was removed. #113594 (Groene AI).
• Fixed LOGICAL_ERROR: There are no events in epoll when a distributed query with use_hedged_requests is cancelled or completes early (for example because of a LIMIT) while a replacement replica is still pending. Closes #113981. #114065 (Groene AI).
• Fixed ... PARTITION <value> statements against a table whose PARTITION BY contains modulo with an unsigned left operand and a signed right operand, such as PARTITION BY (37528 % c0) with c0 Int32. Such a statement raised LOGICAL_ERROR, or silently affected no partition. A value outside the range of the partition key type is now rejected with ARGUMENT_OUT_OF_BOUND instead of silently matching nothing. #116606 (Groene AI).
• Fix BAD_ARGUMENTS for a query that combines several quantile calls where at least one level is written as an integer literal, for example SELECT quantile(1)(x), quantile(0.5)(x). The default-enabled optimize_syntax_fuse_functions rejected the integer level while the aggregate function itself accepts it, so the query failed although each call is valid on its own. #117211 (Alexey Milovidov).
• Fixed wrong results for GROUP BY, LIMIT BY and uniq* over bech32Encode, dictGetHierarchy and reverseUTF8: the three functions declared themselves injective, so the default-enabled optimize_injective_functions_in_group_by and optimize_injective_functions_inside_uniq grouped and counted by the raw argument, although distinct inputs can produce the same result. #117239 (Alexey Milovidov). #117357 (Alexey Milovidov). #118432 (Alexey Milovidov).
• Fix has, indexOf and countEqual over an array of numbers returning a match for a needle that equals does not consider equal: has(arr, -1) matched an Array(UInt64) element equal to the UInt64 maximum, and has(arr, 16777217) matched an Array(Float32) element equal to 16777216. The constant-array path already answered correctly, so the same haystack gave different results depending on whether it was a constant or a materialized column. The same phantom matches were returned for an Array(LowCardinality(T)) column, which also matched a needle its element type cannot represent outside the numeric domain and reported every element equal to the default value as a match for a NULL needle. In the other direction, a needle counting a different unit than the element, such as a DateTime searched in an Array(Date), was found nowhere although equals considers the pair equal; such a pair is now compared in the type the two meet in. #117313 (Alexey Milovidov).
• Fixed uniqTheta and its variants silently returning a value that is too low when reading an AggregateFunction(uniqTheta, ...) state that another Apache DataSketches implementation had written in the compressed Theta sketch form. #117422 (Alexey Milovidov).
• Fixed undefined behavior and an empty result of a SELECT ... ORDER BY <distance function> LIMIT n OFFSET m over a table with a vector_similarity index when n + m, or LIMIT multiplied by vector_search_index_fetch_multiplier, overflows UInt64; such a query now skips the index and runs the exact search. #117713 (Alexey Milovidov). #117892 (Groene AI).
• Fixed a LOGICAL_ERROR (Invalid number of columns in chunk pushed to OutputPort) for a vector search over a column with a Quantized(...) codec when vector_search_use_quantized_codes and the experimental make_distributed_plan were both enabled. #118850 (Groene AI).
• Fixed an external GROUP BY with enable_adaptive_aggregator exceeding max_memory_usage or failing with MEMORY_LIMIT_EXCEEDED inside the aggregator’s own memory valve: the working set of the memory-pressure sweeps was sized from constants unrelated to max_bytes_before_external_group_by, and records staged before the aggregator thawed its tables stayed resident until the merge, so a thawed thread wrote a temporary file per block. Both are now bounded by that threshold. #117880 (Alexey Milovidov). #118474 (Groene AI).
• Fixed a comment ALTER on a KeeperMap table in a Replicated database being applied only to the local replica. Keeper kept the pre-ALTER CREATE statement and digest, so other replicas never saw the change and the next start declared the replica lost, reverting the ALTER and, if no other table shared its KeeperMap path, dropping its data. #118011 (Groene AI).
• Fixes bloom_filter skip indexes on big-endian platforms (s390x), where the index payload was never written or read, so a query using such an index returned no rows for data that is present. #118164 (Vadim Ilves).
• Fixed arrayIntersect returning a spurious NULL when it is missing from one of the input arrays. #118312 (Minh Vu). Fixed arrayIntersect, arrayUnion and arraySymmetricDifference mishandling a Nullable argument whose elements do not fit the common element type: arrayIntersect([toNullable(1)], [toNullable(257)]) returned [1], and arrayUnion([1], [toNullable(257)]) failed with ILLEGAL_COLUMN instead of returning [1,257]. #118485 (Groene AI).
• Fixed a UInt64 constant that does not fit UInt32 matching an unrelated row of a DateTime column in IN (and in an OR chain of equalities, which is rewritten into IN): the constant was truncated modulo 2^32 instead of being excluded from the set. #118367 (Alexey Milovidov).
• Fixed CORRUPTED_DATA (“Cannot get value from Set”) for a FINAL query with ORDER BY ... WITH FILL INTERPOLATE whose filter is an IN set moved into PREWHERE: the set placeholder column is no longer kept alive past the filter. #118420 (Alexey Milovidov).
• Fixed lcm returning a negative result for Int128 and Int256 arguments when exactly one of them is negative, e.g. lcm(toInt128(-6), toInt128(4)) returned -12 instead of 12. #118431 (Alexey Milovidov).
• Fixed text, tokenbf_v1 and ngrambf_v1 skip indexes not being used when the searched constant has type LowCardinality(String) (e.g. hasToken(s, toLowCardinality('rare')) or s = toLowCardinality('rare')), which caused a silent full scan. Such predicates now prune granules like their plain String counterparts. #110017 (Groene AI).
• Fix a text or ngrambf_v1 index silently dropping matching rows when the indexed column is compared for equality against a FixedString value, or searched with hasAny or hasAll over FixedString needles. Happens at default settings, since use_skip_indexes and enable_full_text_index are both on. #117314 (Alexey Milovidov).
• Fix rows silently dropped by an ngrambf_v1 index for a LIKE pattern with a run of escaped backslashes before a wildcard. #118673 (Alexey Milovidov).
• Fixed a bloom_filter skip index dropping granules that hold matching rows for hasAny/hasAll against a constant array whose elements are of a different date/time type than the indexed column, and TYPE_MISMATCH for a Date32 element. #118956 (Alexey Milovidov).
• Fixed wrong results when a table declares both a Map column m and a column named exactly `m.key_<key>` (dots are legal in column names), or has that name claimed by a subcolumn such as a Tuple element or a typed JSON path. Skip-index analysis read the name as the map’s key subcolumn and pruned every granule whose map lacks <key>, so a mapKeys or mapValues index (ngrambf_v1, tokenbf_v1, sparse_grams, bloom_filter, text) dropped rows that matched the predicate. #119457 (Groene AI).
• Fixes incorrect primary-key pruning for has and notHas when array element types differ from the key type, such as String literals against Enum or FixedString keys. #115294 (Alexey Milovidov).
• Fixed rows being silently dropped by partition pruning, primary key analysis and minmax skip indexes when a NaN is involved: a comparison constant that holds a NaN, possibly nested in a tuple or appearing only after conversion to the key type, as in NOT (f < 'nan'). #116763 (Groene AI). A string constant such as 'nan' compared with a floating point column under an injective key transform such as ORDER BY toString(f), so f != 'nan' skipped the granule holding the NaN row. #117501 (Groene AI). A Float primary key holding ±inf where the predicate’s arithmetic maps that endpoint to NaN, e.g. WHERE k / inf = 0. #118468 (Alexey Milovidov).
• Fix wrong results when indexHint is given an argument that index analysis cannot read as a condition. indexHint(256) was narrowed to CAST(256, 'UInt8') = 0 and discarded every part, so the query silently returned an empty result; a wide integer such as indexHint(toUInt256(256)) did the same, and a String, Decimal or NULL argument threw. Such an argument now contributes no filter, consistently for the primary key, minmax, set and bloom_filter analysis. Closes #111684. #114048 (Groene AI). Also fixed a bare key column of a type with no boolean reading (UInt256, Int128, BFloat16, or their LowCardinality forms) being read as column != 0 inside indexHint, NOT indexHint(x) failing with ILLEGAL_TYPE_OF_ARGUMENT, and a negated indexHint in a disjunction keeping its condition un-inverted, so count() could report rows that no row of the table matches. #119209 (Groene AI).
• Fixed wrong results from optimize_functions_to_subcolumns. A Tuple element or a JSON path named like an automatic subcolumn could be mistaken for it, so for example length(c.a) on Tuple(`a.size` UInt64, `a` String) returned the element instead of the length; empty, notEmpty, isNull, isNotNull, count, mapKeys, mapValues, mapContains, and m['key'] were affected the same way. Separately, length on an Array or Map inside a Nullable (for example Nullable(JSON(`a` Array(Int64)))) returned NULL instead of 0 for a NULL row, tupleElement on a member of a Nullable(Tuple(...)) column could fail with LOGICAL_ERROR, and a function is now rewritten only to a subcolumn of the same root column, so a physical column such as `j.m` is no longer confused with the m path of JSON column j. #115777 (Pavel Kruglov). #119150 (Groene AI). #119631 (Groene AI).
• Declaring STATISTICS(...) on an ALIAS or EPHEMERAL column is now rejected when the table is created or altered, instead of being accepted and making every subsequent INSERT fail with NOT_FOUND_COLUMN_IN_BLOCK. Closes #114791. #115882 (Groene AI).
• Fixed CREATE TABLE accepting a subcolumn as a UNIQUE KEY element, of either a user column (for example UNIQUE KEY (`c.null`)) or a virtual one (UNIQUE KEY (`_partition_value.1`)), which produced a table that could never be inserted into, failing every INSERT with NOT_FOUND_COLUMN_IN_BLOCK. Such a definition is now rejected with BAD_ARGUMENTS. Closes #114470. #115854 (Groene AI).
• Fixed an in-process Keeper client (KeeperOverDispatcher) reporting a failed multi transaction as successful, because a rolled-back batch’s aggregate error was left ZOK. #117135 (Sergei Trifonov).
• Fixed OOM when reading a Parquet file with an invalid num_children field in the schema. #117797 (Michael Kolupaev).
• Fixed a Parquet reader bug where a filter on a UUID column could silently return fewer rows than the file holds, or reject a valid file with INCORRECT_DATA: Statistics have min_value > max_value. Parquet sorts uuid statistics by big-endian bytes while ClickHouse sorts UUID by its second half, so those statistics are no longer used for row group or page pruning. Closes #118371. #118387 (Groene AI).
• Fixed wrong results when reading a Parquet file with a nullable struct column without an explicit schema. Schema inference pushed the nullability down to the tuple elements and returned Tuple(Nullable(...), ...), which cannot hold a struct-level NULL, so a NULL struct was read as a tuple of NULLs. Such a column is now inferred as Nullable(Tuple(...)) when allow_experimental_nullable_tuple_type is enabled. That setting was also added to the Parquet schema inference cache key, so a schema cached under one of its values is no longer reused under the other. #115653 (Groene AI).
• Fix reading Arrow and ArrowStream files containing values hidden by NULLs or unused nested ranges, which could cause validation errors or expose undefined values. Fix nullable tuple conversions and destination defaults, dictionary decoding under requested types, and NULL dictionary entries for complex types. Accept compatible shared dictionaries with differing field annotations and reject replacement dictionaries in Arrow files. Validate nested row counts before allocation and read bufferless batches incrementally without materializing their entire declared row count. #110531 (Nihal Z. Miaji). Fix reading a flattened Tuple subcolumn (e.g. `s.v`) from an Arrow/ArrowStream file whose column is a nullable struct returning column defaults instead of the actual data when allow_experimental_nullable_tuple_type is enabled. Also fix ORC reads resolving a column or struct field name to the first case-folded match under input_format_orc_case_insensitive_column_matching, which returned another field’s data when two names differ only by case. #109741 (Groene AI).
• Fix a crash and wrong results when reading the DWARF format with column types other than the ones the format produces, for example SELECT * FROM file('a.elf', DWARF, 'tag String'). Requesting a column that does not exist in the format now reports a proper error instead of std::out_of_range. #118794 (Raúl Marín).
• Fix data silently dropped when reading a hadoop-snappy stream whose block has several subblocks and the input is refilled in the middle of that block, which happens when the file is read over a pipe or HTTP. #118705 (Alexey Milovidov).
• Fixed two defects in the legacy binary settings format, used for native-protocol clients declaring a revision below 54429. Signed integer settings (for example max_partitions_to_read or iceberg_snapshot_id) were silently ignored, and were reported as changed while holding their default, so they overrode the table-level value with that default. Settings required to be greater than zero (for example max_block_size) were not validated when read from that format. #118610 (Groene AI).
• Fixed a query over an Iceberg table returning rows that had been deleted, when a position-delete manifest declares a lower bound above the upper bound for the reserved file_path column. Such a pair describes an empty range, so the delete file was skipped for every data file it covers; those bounds are now ignored instead. Closes #118113. #118181 (Groene AI).
• Fixed reading an Iceberg table failing with ICEBERG_SPECIFICATION_VIOLATION after a column was dropped while its data is still present in the data files. Closes #113324. #116488 (Konstantin Vedernikov).
• Fixed write_full_path_in_iceberg_metadata writing an unresolvable local:// URI, with a doubled root slash, into the metadata of an IcebergLocal table, so that no external Iceberg reader could open it. A local table location is now spelled file:///path, and a backend without a namespace no longer contributes an empty authority segment. #117828 (Groene AI).
• Fixed writing Iceberg statistics on top of Azure storage. #119369 (Konstantin Vedernikov).
• ATTACH TABLE of a data lake table no longer reads anything from object storage, and such tables no longer report row statistics in system tables. #117584 (Konstantin Vedernikov).
• Reading a table backed by local object storage (paimonLocal, icebergLocal, deltaLakeLocal, or a disk with object_storage_type: local) whose metadata references a file that is missing on disk now fails with a FILE_DOESNT_EXIST exception naming the file, instead of a raw std::filesystem error surfaced as the opaque STD_EXCEPTION (code 1001). #115542 (JIaQi Tang).
• Fixed a SELECT from a table with paimon_incremental_read = 1 failing with REPLICA_IS_ALREADY_ACTIVE (“Failed to activate Paimon replica after Keeper reconnection”) on the first read after the server’s Keeper session was lost and re-established. #118457 (Groene AI).
• Fixed a spurious Code: 76 CANNOT_OPEN_FILE error (Parent directory doesn't exist) on any write to an hdfs disk. Removing a blob deleted its emptied object-key prefix directory, which could race a concurrent write that had just created the same directory and not yet opened its object. Emptied prefix directories are now kept. #118654 (Groene AI).
• Fixed a server refusing to start after remote_url_allow_hosts was tightened while a database with the HDFS engine existed: the host check ran during metadata loading and took every other database down with it. #118964 (Alexey Milovidov).
• Fixed a server that refused to start after the directory backing a Filesystem database had been removed. Such a database is now loaded without tables and can be inspected and dropped, while a CREATE DATABASE or ATTACH DATABASE issued by a user is still rejected when the path does not exist. #118684 (Groene AI).
• With enable_analyzer = 1 (the default), fixed INCOMPATIBLE_TYPE_OF_JOIN and NOT_FOUND_COLUMN_IN_BLOCK errors when an ALL INNER JOIN against a table with the Join table engine carried an ON condition referencing only that table, such as ON t.key = j.key AND j.value > 100. The condition is now applied after the join, which is what the equivalent WHERE spelling already did. Closes #116541. #116612 (Groene AI).
• Fixed uneven partition distribution in the Kafka engine with kafka_keeper_path (StorageKafka2) when kafka_num_consumers is greater than 1. A node running N consumers could claim N times its fair share of the topic’s partitions and hold them indefinitely, leaving other replicas idle. #116843 (Aparajita Pandey).
• clickhouse-local now preserves detailed Poco::Exception messages from local query execution instead of reporting them as generic STD_EXCEPTION errors. This includes errors returned by the mysql table function. #116884 (Mark Needham).
• Fixed HashJoin right-side memory accounting when a memory limit interrupts a join build. A block could be stored without being counted, so the join under-reported its size to the memory limits and the spilling thresholds, and a Join engine table held the memory of a failed INSERT for the rest of its lifetime. #116991 (Groene AI).
• Fix wrong aggregate values from the default-enabled optimize_rewrite_aggregate_function_with_if when the constant is in the first branch of the if and the condition can be NULL, for example sum(if(c, 0, x)) or avg(if(c, NULL, x)). Rows whose condition was NULL were silently dropped from the aggregate. #117221 (Alexey Milovidov).
• Fixed a signed integer overflow when the effective table lock timeout, the smaller of lock_acquire_timeout and a non-zero max_execution_time, was above roughly 9.2e9 seconds or a large negative value. Reading or writing a Log, TinyLog, StripeLog or File table built the lock deadline from it without bounding it first, so the wait was no longer bounded by the setting. #117974 (Groene AI).
• divideDecimal no longer returns 0 instead of the correct value when the dividend reduces to a single digit after being rescaled to the result scale, for example divideDecimal(toDecimal256(5, 0), toDecimal256(1, 0), 0) or divideDecimal(toDecimal32(5.6789, 4), toDecimal32(2, 0), 0). The bug was present in every release since 22.12, so any stored result of such a division is wrong and should be recomputed. #118346 (Minh Vu).
• Fixed ClickHouseDictionarySource table cannot be dictionary table being raised for a dictionary whose source is a table on another ClickHouse server that happens to share the dictionary’s database and table name. The check against a dictionary using itself as its own source now applies only when the source is the local server. #118441 (Alexey Milovidov).
• Fix arrayElementOrNull over an array of tuples, and arrayElement over an array of Nullable tuples, returning a tuple of default values instead of NULL for a constant UInt64 index close to the maximum of UInt64, such as toUInt64(18446744073709551615). #118447 (Minh Vu).
• Fixed a query parameter being reported as unset (UNKNOWN_QUERY_PARAMETER) when it is the entire INTERPOLATE expression, as in ORDER BY x WITH FILL INTERPOLATE (value AS {p:Int64}), while the same parameter nested in a larger expression worked. Closes #118458. #118568 (Groene AI). The same for a parameter that is the entire predicate of a lightweight DELETE, as in DELETE FROM t WHERE {p:UInt8}. #118667 (Groene AI).
• Fix partition pruning and primary key filtering being silently disabled when parallel_replicas_min_number_of_rows_per_replica is set inside a subquery or CTE instead of at top query level. #116529 (Yarik Briukhovetskyi).
• Fixed wrong results and a LOGICAL_ERROR for SELECT ... FINAL under make_distributed_plan = 1 when a PREWHERE or a row policy is deferred until after FINAL (apply_prewhere_after_final, or the default apply_row_policy_after_final). #116626 (Groene AI).
• Fixed wrong results from JIT-compiled conversions of a floating point value to an integer or Decimal type when the value does not fit the destination: a comparison against the converted value could be evaluated against the pre-conversion value, so the same query could flip between repeated executions. Such conversions are now evaluated by the interpreter, which also means they raise CANNOT_CONVERT_TYPE for NaN or Inf and DECIMAL_OVERFLOW on Decimal overflow where an arbitrary value was previously returned. Closes #117442. #117590 (Groene AI). Fixed wrong results from a JIT-compiled CAST of a Nullable value to a non-Nullable type: converting a NULL must raise CANNOT_INSERT_NULL_IN_ORDINARY_COLUMN, but the compiled expression returned the nested value instead, so the same query silently started answering with wrong data once min_count_to_compile_expression was crossed. Such a conversion is no longer compiled. #119079 (Groene AI).
• A simple Query in the PostgreSQL wire protocol now destroys the unnamed prepared statement and the unnamed portal, as PostgreSQL does, so a later Bind or Execute on them is rejected instead of silently reusing stale extended-query state. #118210 (Groene AI).
• portRFC now returns the explicit port of a URL whose host is an IPv6 address in brackets, such as http://[2001:db8::1]:8080/, instead of the default port. #118632 (Alexey Milovidov). port and portRFC now return the explicit port for valid URLs with a query string or fragment after the port. #118482 (Minh Vu).
• fuzzBits and generateULID now return an independently generated value per row when all of their arguments are constants, instead of one value repeated for every row. #118643 (Alexey Milovidov).
• Fix EmbeddedRocksDB, KeeperMap and Redis returning no rows for WHERE key = <literal of another date-family type>, such as a Date key compared with a DateTime literal. #118689 (Alexey Milovidov).
• Fixed the row numbers shown in the Web UI for a page other than the first one: they restarted at 1 on every page instead of counting from the page’s position in the whole result. #116614 (Alexey Milovidov).
• Fix delayed ClickHouse Keeper shutdown when idle client sessions are blocked in socket polling. #118014 (Antonio Andelic). Fix graceful shutdown of clickhouse-keeper hanging for up to 300 seconds when a client connection on keeper_server.tcp_port_secure is stalled in a TLS read. #119450 (Groene AI).
• Fixed a Keeper crash when a member that was removed from the cluster is added back without its process being restarted. Such a member could also replay already-applied entries. #118740 (Shaohua Wang).
• Fix handling of regular expressions ending with an unescaped backslash in match, extractAll, countMatches and splitByRegexp. Such invalid patterns are no longer incorrectly treated as trivial substring matches. #118229 (Alex Prabhat Bara).
• Fixed match and countMatches returning wrong results for a pattern whose character class starts with a literal ] ([]a]b) or contains a literal [ next to an alternation (abc[[]|b): the required-substring prefilter demanded a substring the pattern does not require. The same defect made optimize_rewrite_regexp_functions turn such a replaceRegexpAll into replaceRegexpOne. #118342 (Alexey Milovidov).
• Fix startsWith* array functions returning suffix matches for some variants. #118672 (Minh Vu).
• Fixed an out-of-bounds access that could happen with the getURLHostRFC function. #119026 (Groene AI).
• Fix the HTTP connection pool comparing disk_connections_store_limit, storage_connections_store_limit, and http_connections_store_limit against the total number of connections in the group instead of the number of stored connections. A server whose concurrency exceeded the limit cached no connections at all, opened a new TCP connection for every request to object storage, and could exhaust the ephemeral port range of the host. The limit now bounds only the idle connections kept for reuse. #119078 (Alexey Milovidov).
• Fixed the hash-family join algorithms treating a non-NULL Nullable(Tuple(...)) key that contains a NULL element as a NULL key, which made the result of a join depend on the join_algorithm setting. IN with a subquery over such a key was affected in the same way. #119081 (Alexey Milovidov).
• Fix wrong results when analyzer_compatibility_join_using_top_level_identifier = 1 and a JOIN ... USING key resolves to a SELECT-list alias whose name is a virtual column of the left table expression, such as SELECT upper(t._part) AS _part FROM t JOIN r USING (_part). The raw virtual column value was silently used as the join key instead of the aliased expression. #119102 (Groene AI).
• Fixed a LOGICAL_ERROR (Got read request from replica N for unknown stream <table>) when parallel_replicas_plan_based is enabled and a join is distributed across replicas. #118084 (Groene AI).
• Fixes a case where date_time_overflow_behavior = 'saturate' clamped an out-of-range value to the whole second and discarded the sub-second component. Closes #113152. #116122 (Yarik Briukhovetskyi).
• Fix system.dictionaries.bytes_allocated misreporting memory for HASHED_ARRAY dictionaries: the key containers contributed their element count instead of their buffer size in bytes, so the reported figure came out roughly 2-5x lower than the memory the dictionary actually held, and the bit-packed per-attribute null masks were counted in elements, which over-reported that term by 8x. #118641 (George Viamontes). Fixed the same figure over-reporting memory for nullable RANGE_HASHED dictionaries, whose bit-packed nullable mask was counted one byte per flag. #118675 (Minh Vu).
• Fixed max_execution_time_leaf and timeout_overflow_mode_leaf being ignored for queries executed with parallel replicas (including INSERT SELECT with parallel_distributed_insert_select = 2). #69540 (Igor Nikonov).
• Fix CREATE TABLE ... CLONE AS SELECT ... and CREATE TABLE ... CLONE AS <table_function> silently ignoring the CLONE keyword and creating a plain CREATE ... AS SELECT instead. These shapes are now rejected with BAD_ARGUMENTS. #105004 (Tuan Pham Anh).
• Fix a logical error (a server abort in debug builds) when a non-transactional DROP PARTITION, DETACH PARTITION, DROP PART, REPLACE PARTITION FROM or MOVE PARTITION TO TABLE covered a part whose creating transaction had not committed yet. Such an operation now fails with SERIALIZATION_ERROR and can be retried once the creating transaction finishes; it no longer removes or replaces any part of the affected batch. #105116 (Tuan Pham Anh).
• Fix a logical error when a correlated MATERIALIZED CTE is referenced more than once inside a JOIN. Such a CTE is now rejected at analysis time with UNSUPPORTED_METHOD, consistently with the single-reference and IN-subquery paths. #108883 (Groene AI).
• Fix a logical error Invalid cast from Nothing to native type (an abort in debug/sanitizer builds) when compile_expressions = 1 compiled an expression with a Nullable(Nothing) operand, e.g. and(Nullable(Nothing), <native comparison>). #111508 (Groene AI).
• Fix wrong results with enable_parallel_replicas = 1 when a join tree contains a join other than the leftmost one that cannot be applied independently per replica: a strictness other than ALL outside a LEFT join, for example t1 INNER JOIN t2 ON t1.c = t2.c ANY INNER JOIN t3 ON 1, which returned the row once per replica instead of once; a FULL join, which emitted every row twice; or a PASTE join, which pairs rows by position. Parallel replicas are now disabled for such a join tree, as they already are for the equivalent two-table join. #112317 (Groene AI).
• Fix a segfault and silently wrong results when reading a sizeN subcolumn of a nested Array(JSON) path. Closes #113686. #113752 (Groene AI).
• Fixed wrong results and a server abort when a JOIN ON condition evaluated during an ie_join join contains arrayJoin. Such a condition is now rejected with INVALID_JOIN_ON_EXPRESSION. Previously SEMI/ANTI/LEFT/RIGHT/FULL joins using the ie_join algorithm could return duplicated or missing rows, or abort the server. #114325 (Groene AI).
• Fix reading a Log or TinyLog table whose array or map elements are missing while their sizes are present: such a table returned rows with elements that are not there and made the server read out of bounds. It is now reported as INCORRECT_DATA. #115528 (Alexey Milovidov).
• Fix a segmentation fault when reading in order from a Merge table over MergeTree tables whose sorting key is shorter than the key prefix chosen for the query. #116364 (Groene AI).
• Fix LOGICAL_ERROR: GroupingAggregatedTransform pushed bucket N twice when make_distributed_plan picks the Shuffle aggregation strategy while the aggregation must produce results in bucket order (for example with distributed_aggregation_memory_efficient). It is now planned as a partial aggregation plus a merge. #116446 (Groene AI).
• Fix wrong row order when sorting a LowCardinality column of a floating-point type whose dictionary holds two entries that compare equal, such as -0.0 and +0.0. A multi-column ORDER BY did not apply the remaining keys to those rows, and the MergeTree writer’s stable sort did not keep them in insertion order, changing which row ReplacingMergeTree keeps. #116870 (Groene AI).
• Fixed wrong results from the default-on optimize_time_filter_with_preimage for toYear/toYYYYMM predicates at the column type’s maximum (Date at 2149, DateTime at 2106), a DECIMAL_OVERFLOW exception for the same predicates on a DateTime64(9) column at 2262, and wrong results for the same predicates on DateTime/DateTime64 columns in time zones whose year or month boundary is not local midnight (America/Lima, America/Managua). #117451 (Alexey Milovidov).
• Fixed NOT_IMPLEMENTED (“Cannot decorrelate query, because ‘DelayedCreatingSets’ step is not supported”) for a correlated subquery whose WHERE builds a set, for example an IN (subquery), or a dictGet(...) >= const comparison that optimize_inverse_dictionary_lookup rewrites into one. #118412 (Alexey Milovidov).
• Fixed a race between SYSTEM RESTORE DATABASE REPLICA and the DDL worker of a Replicated database. The restore re-initialised the replica while the previous DDL worker was still running its own startup recovery, so the recovery could mark the wrong range of replicated-DDL log entries as finished, or move local tables to _broken_replicated_tables. The worker is now stopped before the restore touches Keeper. #118481 (Groene AI).
• Fixed WITH RECURSIVE losing its recursiveness in the second and later branches of a UNION, INTERSECT or EXCEPT. When a later branch had no WITH clause of its own, the first branch’s WITH list was copied into it without the RECURSIVE marker, so WITH RECURSIVE src AS (SELECT 1 AS id UNION ALL SELECT id + 1 FROM src WHERE id < 3) SELECT sum(id) FROM src UNION ALL SELECT sum(id) FROM src was rejected with UNKNOWN_TABLE, and returned a wrong value for the copied branch when an enclosing CTE of the same name existed. CREATE VIEW also stored the copied branch without RECURSIVE. #118510 (Groene AI).
• Fixed a GLOBAL RIGHT JOIN over Distributed tables returning unmatched rows of the right table once per shard: the original left table is now broadcast and the preserved right side stays sharded. Closes #95728. #118611 (Anton Ivashkin).
• Fix an endless ORDER BY ... WITH FILL when the STEP is too small to change the value it is added to. #118633 (Alexey Milovidov).
• Fix the Invalid local IN function name globalIn error for a query with IN over a UNION of Distributed tables when distributed_product_mode = 'global' or prefer_global_in_and_join is enabled. #118639 (Alexey Milovidov).
• toLastDayOfMonth and toLastDayOfWeek of a Date argument now saturate at the Date maximum (2149-06-06) instead of wrapping around to 1970. The wrapped result also silently dropped matching rows through primary-key pruning. #118642 (Alexey Milovidov).
• Fixed a crash and a LOGICAL_ERROR in the mergeTreeIndex table function when reading a key column whose ALTER TABLE ... MODIFY COLUMN between T and LowCardinality(T) has an unmaterialized mutation. The key column is now emitted in the type the current metadata declares. #118709 (Groene AI).
• Fix SYNTAX_ERROR when loading views, materialized views, or projections whose SELECT list starts with a column named TOP. #118754 (Kirill).
• Fixed optimize_or_like_chain returning extra rows for OR chains of LIKE/ILIKE patterns that do not end in %: a value followed by a trailing newline wrongly matched the filter. #118766 (Jimmy Aguilar Mena).
• Fixed a LOGICAL_ERROR (“Expected CommonSubplanReferenceStep to reference CommonSubplanStep”) raised at query planning time when query_plan_convert_join_to_in = 1 was used with a correlated aggregating scalar subquery. The optimization now keeps the JOIN instead of rewriting it when either join input carries a common subplan node. #118797 (Groene AI).
• Fix a wrong result for SELECT * APPLY (x -> f(x)) over JOIN ... USING (a, b): every USING column was named after the lambda parameter (f(x)) instead of the matched column, so two output columns shared one name, and wrapping the query in SELECT * FROM (...) returned one USING column’s values under the other column’s position; a type-returning lambda such as toTypeName failed with AMBIGUOUS_COLUMN_NAME instead. The same happened for COLUMNS(a, b) APPLY (x -> f(x)), and for two or more ARRAY JOIN columns. The lambda form of APPLY now names these columns like APPLY f and a plain * already did. #118825 (Groene AI).
• Fixed a server crash that could happen when a WORKLOAD setting limiting concurrency (e.g. max_concurrent_queries or max_bytes_inflight) was lowered below the number of currently in-flight resource requests while the workload was active. #118836 (Sergei Trifonov).
• Fixed a logical error (Method == not implemented for AST) raised by a SET of a custom setting that already held an AST value such as disk(type = 's3', ...). Assigning such a setting a second time in one session, or clamping a nested SETTINGS clause against it, reached a comparison that was never implemented. #118993 (Groene AI).
• Fixed a broken part in detached/ never being renamed with the broken_ prefix by the ATTACH_PART executor, which left it occupying a valid part name and failing every later attach attempt. #118999 (Alexey Milovidov).
• Fix TOO_DEEP_SUBQUERIES being thrown for GLOBAL IN and GLOBAL JOIN subqueries in two analyzer cases: when max_subquery_depth is set, the depth is now counted in nested subqueries instead of query tree nodes; and when a UNION is on the path to the GLOBAL subquery, UNION nodes no longer count toward that depth. Closes #79345. Closes #119623. #119033 (Vladimir Cherkasov). #119650 (Groene AI).
• With prefer_column_name_to_alias = 1, an unqualified column name that is ambiguous between joined tables now resolves to the alias with the same name, as the old analyzer did, instead of throwing AMBIGUOUS_IDENTIFIER. #119192 (Vladimir Cherkasov). #119906 (Vladimir Cherkasov).
• Fix a wrong-order merge (logical error Virtual row does not cover sort column in debug builds) when reading in order with virtual rows and ORDER BY contains constants, e.g. SELECT DISTINCT a, b FROM t ORDER BY 'x', a, 'd' COLLATE 'cs', b on a table with key (a, b). #119195 (Vladimir Cherkasov).
• Fixed wrong results when a table’s sorting key expression resolves to a different result type in the query than in the table, for example a key of CAST(json.b, 'String') read with cast_keep_nullable = 1. Read-in-order treated the two as interchangeable, so rows could come back out of order, optimize_aggregation_in_order could split groups and report wrong aggregates, and read_in_order_use_virtual_row = 1 could fail with LOGICAL_ERROR: Virtual row has different type. #119385 (Groene AI).
• Fixed BACKUP of a Memory table reading an unrelated file named tmp_sizes_json from the server’s working directory: the sizes.json shipped inside the backup is now built in memory, so such a file can no longer make the backup fail or add foreign entries to it. #119330 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK when a MergeTree skip index uses mortonEncode or hilbertEncode directly and a query filters on one of the curve arguments. #119355 (Minh Vu).
• Fixed an ALTER on a view or materialized view with SQL SECURITY DEFINER deleting the auto-created <user>:definer account when the ALTER did not change the definer. Every later SELECT from the view, and every INSERT into a materialized view’s source table, then failed permanently with There is no user '<user>:definer' in 'user directories' (UNKNOWN_USER). #119357 (Groene AI).
• Fixed a materialized view whose SELECT contains a UNION, INTERSECT or EXCEPT in a subquery or CTE breaking after DETACH TABLE + ATTACH TABLE, or after RESTORE: every subsequent INSERT into its source table failed with UNION mode UNION_DEFAULT must be normalized, and under enable_analyzer = 0 it was silently evaluated as a plain union, returning wrong results. ATTACH TABLE (short form) and RESTORE now re-run set-operation normalization on the stored definition. #119365 (Groene AI).
• Fix an ALIAS column that references another column (b UInt64 ALIAS a) reading back as the default value of its type when the table is read through a Merge table over a Distributed table with a remote shard. #119561 (Alexey Milovidov).
• Fix arrayShiftLeft and arrayShiftRight when an explicit NULL default is used with a nullable array. #119626 (Minh Vu).
• Fix ALTER TABLE ... MODIFY SETTING name = DEFAULT being ignored. Closes #49613. #118526 (Vitaly Baranov).
• Fix arrayNormalizedGini rejecting valid rows with different array sizes in the same block. #118631 (Minh Vu).
• Fix substringIndex and substringIndexUTF8 rejecting UInt64 counts above Int64::max() instead of interpreting them as negative counts. #118645 (Minh Vu).
• Fix a logical error for an asynchronous INSERT combined with a non-parallel quorum (insert_quorum greater than one or 'auto' together with insert_quorum_parallel = 0) when the data is sent as blocks over the native protocol. Such a query is now rejected with UNSUPPORTED_PARAMETER, as it already was when the data was inlined in the query text. #118968 (Alexey Milovidov).
• Fixed a graceful shutdown silently discarding the rows of a Buffer table whose destination lives in another database or is another Buffer table. #118985 (Alexey Milovidov). #116898 (Jayme Bird).
• Fix a use-after-free in the gRPC server when a query with a streaming input finishes while a read of the next message is still in flight. #115707 (Alexey Milovidov).
• Custom object metadata written to Google Cloud Storage over the S3 API without HMAC credentials, which is how ClickHouse Cloud on GCP accesses object storage, now takes effect: the headers that GCS spells with the x-goog- prefix are translated for every request, not only for server-side copies. Previously GCS discarded the x-amz- spelling silently, so the write succeeded and the metadata was not applied. #119042 (Sema Checherinda).
• Fixed a crash when reading a subcolumn of a Variant, Dynamic, or JSON column whose own type is Nullable or LowCardinality(Nullable) from a MergeTree table, for example SELECT value.`Tuple(a Nullable(UInt32), b String)`.a over a Dynamic column. Previously such a read hit a LOGICAL_ERROR in debug builds and crashed in release builds. #112338 (Groene AI).
• With shutdown_wait_backups_and_restores = 0, a BACKUP or RESTORE whose I/O did not return could keep the server from terminating on a single SIGTERM. That wait is now bounded by shutdown_wait_unfinished, and new backups and restores are refused once shutdown has begun. #115377 (Groene AI).
• Fix wrong results when query plan optimizations treat non-deterministic expressions as deterministic, including lambdas and WebAssembly UDFs, which could silently drop rows. #117248 (Alexey Milovidov). #117249 (Alexey Milovidov).
• Fix Column identifier ... is already registered (LOGICAL_ERROR) for a correlated subquery over a VIEW on a Distributed table. Such a query is now refused with Correlated subqueries are not supported with remote tables, the same as a correlated subquery over the Distributed table itself. The same applies to parameterized views, Merge tables over such views, and materialized views whose target is such a view. #117335 (Alexey Milovidov).
• Fixed transform and the simple CASE expr WHEN form returning the default value instead of the matching one when the expression is a Nullable(Enum) column. Only the member whose numeric value was 0 used to match. Closes #117206. #117520 (Groene AI).
• Fix SYSTEM RELOAD DICTIONARY <bare-name> ON CLUSTER and SYSTEM UNLOAD DICTIONARY <bare-name> ON CLUSTER so normal distributed DDL clusters resolve the bare dictionary name against the initiator’s current database before queuing the task, matching existing table-like ON CLUSTER DDL behavior. Closes #114322. #117596 (Alexander Yoon).
• Fixed silently missing rows for predicates on toUnixTimestamp of a Date or Date32 primary key or column with statistics: the conversion wraps around UInt32 for dates after 2106-02-07, but index analysis assumed it was monotonic everywhere. #117709 (Alexey Milovidov).
• Fixed a query getting stuck when the server fails to start a thread for a parallel task, for example when a thread pool is exhausted. #117781 (Alexey Milovidov).
• Fixed wrong results for a distributed query whose folded constant contains a DateTime value. The constant is now sent to secondary servers as its raw Unix timestamp wherever the receiving side re-applies the literal’s declared type, so DST fall-back overlaps no longer reinterpret it as another instant. #117837 (Groene AI).
• Fixed queries failing with AUTHENTICATION_FAILED after RENAME DATABASE on a Replicated database that uses cluster_secret. The cached cluster object kept the old database name, which participates in the interserver handshake. #117874 (Kirill Shokhin).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK when transform_null_in = 1 and a distributed or parallel-replicas query aggregates over a predicate that an analyzer pass rewrote into IN, such as SELECT sumIf(id, has([1, 2, 3], id)) FROM t. #117897 (Groene AI).
• Fixed THERE_IS_NO_COLUMN (Cannot find column ... in source stream) when a query shipped as a serialized query plan (serialize_query_plan = 1) reads a Merge table that spans a local table and a Distributed table. #117969 (Groene AI).
• Fixed a SELECT returning fewer rows than it should when a JSON subcolumn is compared with an Enum constant whose label is the empty string, and the table has a JSONAllPaths skip index (bloom_filter, tokenbf_v1, ngrambf_v1, or sparse_grams). #117975 (Groene AI).
• An Enum constant wrapped in Variant or Dynamic and compared against a String or FixedString column no longer uses the enum’s underlying number instead of its name, which previously made IN, NOT IN, and key analysis use the wrong value. #117998 (Groene AI).
• Fixed a Replicated database being unable to recover a lost replica if it held a Remote or RemoteSecure table whose target is a table function such as merge(...) that no longer resolves. #118221 (Groene AI).
• Fixed Not found column ... in block for a query that combines a constant ORDER BY key, LIMIT BY, and LIMIT ... WITH TIES. #118222 (Alexey Milovidov).
• Fixed two overflow cases where memory_usage_overcommit_max_wait_microseconds lost the wait it was asked for and timed out immediately instead of waiting for memory to be freed. #118227 (Groene AI).
• Fixed CREATE VIEW being refused with SUSPICIOUS_TYPE_FOR_LOW_CARDINALITY or ILLEGAL_COLUMN when the view’s inferred column type is gated by settings such as allow_suspicious_low_cardinality_types or enable_time_time64_type. Views store no data, so those storage-type checks no longer block them. #118298 (Groene AI).
• Fixed Not-ready Set is passed as the second argument for function 'globalIn' when a PREWHERE or row-level filter expression containing an IN subquery is evaluated by a query planned to the FetchColumns stage. #118359 (Groene AI).
• SELECT count() FROM <set-table> now works for the Set engine, using the table’s stored row count instead of failing with Method read is not supported by storage Set. Reading regular columns from a Set table is still unsupported. #118491 (Alexey Milovidov).
• Fixed sequenceMatchEvents returning an event chain that the pattern does not authorize, for example one that skips a (?t...) time gate between two of the returned events. The returned chain now always matches a prefix of the pattern. Closes #118120. #118497 (Groene AI).
• Fixed toSecond, toStartOfMinute, toStartOfHour, toStartOfInterval, and the other toStartOf* functions returning wrong values for pre-1970 DateTime64 values in time zones whose historical offset had a sub-minute component, such as Europe/Amsterdam before 1937 or Asia/Kolkata before 1906. #118501 (Alexey Milovidov).
• Fix array aggregation functions returning non-default values for empty arrays when the lambda result is constant. #118636 (Minh Vu).
• Conversions of 128- and 256-bit integers to DateTime, DateTime64, and Time64 now saturate like the 64-bit conversions instead of wrapping or storing an out-of-range value. This also fixes silent row loss through primary-key pruning. #118717 (Alexey Milovidov).
• Fix a possible heap buffer overflow on malformed blob-storage responses. #119054 (Arsen Muk).
• Preserve input sorting for groupArrayInsertAt, stochasticLinearRegression, and stochasticLogisticRegression, so an ORDER BY feeding them is not optimized away. #119142 (Ivan Babrou).
• Fixed a crash and a LOGICAL_ERROR for statements that write trim(BOTH '' FROM x) where a function is expected, such as in a column’s STATISTICS(...) or CODEC(...) list, INDEX ... TYPE, ALTER TABLE ... ADD STATISTICS, a table function, a named collection, or SETTINGS. #119317 (Groene AI).
• Fix CHECK TABLE permanently reporting Found unexpected projection directories after a mutation when the part carries a projection that the table has since dropped. #119348 (Alexey Milovidov).
• Fix a wedged mutation queue after ATTACH PARTITION of a partition that was detached before a DROP COLUMN: mutations of a Compact part carrying the dropped column no longer fail forever with UNKNOWN_IDENTIFIER. #119354 (Alexey Milovidov).
• Fix a bug where unfiltered TRUNCATE ALL TABLES FROM queries permanently disabled merges on retained MergeTree tables, causing later OPTIMIZE TABLE ... FINAL queries to fail with ABORTED. #119425 (Paul Annesley).
• Fix Unexpected value 0 in enum thrown by string functions such as LIKE, match, and position when applied to a Nullable(Enum) value that is NULL, including NULLs produced by a LEFT JOIN with join_use_nulls = 1. #119513 (Alexey Milovidov).
• Fix arrayJaccardIndex when its input arrays contain duplicate elements. #119560 (Minh Vu).
• modulo, positiveModulo, moduloOrZero, and moduloOrNull no longer return wrong results when one operand is a signed integer and the other is an unsigned integer at least as wide. #119616 (Serhiy Bzhezytskyy).
• A Filesystem database no longer keeps pointing at a removed directory, which previously made the server refuse to start on the next restart. #119618 (Alexey Milovidov).
• Fixed a standalone UPDATE ... SET reading the wrong table when a mutation expression names a table that a same-named WITH element hides but the subquery does not see because its own SETTINGS clause disables enable_global_with_statement. #119630 (Groene AI).
• Fixed a stack buffer overflow when reading a BIT column from MySQL through the mysql table function, the MySQL table and database engines, or MySQL dictionaries. #119680 (Alexey Milovidov).
• Fix sparseGrams ignoring max_ngram_length when min_cutoff_length is specified. #119713 (Minh Vu).
• Fix BACKUP and RESTORE of a Log or TinyLog table when one column occupies zero bytes on disk while the table has rows, such as a column of countResample states over an empty range. Previously the restored table could come back empty or fail with FILE_DOESNT_EXIST. #119722 (Groene AI).
• Fix a race between ALTER TABLE and RENAME TABLE / EXCHANGE TABLES. #103044 (Nikolay Degterinsky).
• Fix BACKUP of EmbeddedRocksDB tables. Previously only the table metadata was stored in the backup and the actual data was silently dropped, so RESTORE produced an empty table. The backup now includes all key-value pairs and restores them correctly. #109327 (Groene AI).
• Fix toStartOfWeek and toLastDayOfWeek with a Sunday-first week mode returning a wrapped day number for a Date32 value outside the representable calendar, for example after toDate32('1970-01-01') + INTERVAL 2147483647 DAY. The week boundary is now computed at the calendar boundary, consistently with the Monday-first week modes. #111923 (Groene AI).
• Fixed mergeTreeCodecBlockCounts disclosing whether a source table exists, and even its engine type, to a user without any privilege on that table. Resolving the table function, for example with DESCRIBE, now checks access before consulting the catalog. #116647 (Alexey Milovidov).
• Messages of the PostgreSQL wire protocol are now parsed within their declared size, so a malformed message can no longer make the server allocate memory for a payload that never arrives. #116998 (Alexey Milovidov).
• Fixed CREATE TABLE dst AS src copying the schema of an ENGINE = Alias table’s target for a user who holds SHOW COLUMNS on the alias but no privilege on its target. The statement now requires table-level SHOW COLUMNS on the target, matching DESCRIBE of that target. #117702 (Groene AI).
• Fixed the query condition cache serving a “no matching rows” verdict to a query that differs from the cached one only in a setting that changes how a function evaluates, such as the formatdatetime_* / parsedatetime_* settings or function_locate_has_mysql_compatible_argument_order, which silently dropped all matching rows. #117803 (Alexey Milovidov).
• Fixed a CAST AS Tuple can only be performed between tuple types or from String error, and silently wrong results for a Tuple(String, ...) key, when a MergeTree sorting, partition or skip-index key wraps a single Tuple-typed column in a deterministic function (for example ORDER BY toString(k)) and the query filters that column with k IN (tuple(...)), k NOT IN (tuple(...)) or a multi-column subquery set. #118899 (Groene AI).
• Fix queries that use a higher-order function together with ARRAY JOIN: a lambda that captures nothing is no longer moved across the ARRAY JOIN step away from the function that applies it. #119186 (Yarik Briukhovetskyi).
• Fix TYPE_MISMATCH in the selectivity estimator when a predicate on a DateTime64 column uses an unconvertible string literal. Closes #119126. #119265 (Han Fei).
• Fix unstable ordering of equal sorting keys during inserts, which could cause CollapsingMergeTree to retain stale rows when cancellation and replacement rows were inserted together with optimize_on_insert enabled, and ReplacingMergeTree to retain the wrong row when versions were equal. #119474 (Anton Kovalenko).
• Fixed cleartext credentials in system.asynchronous_insert_log.query. Secrets of table functions such as s3, url or mysql are now masked as [HIDDEN], the same way system.query_log masks them. #119487 (Sema Checherinda).
• Fix a crash when reading a table with active lightweight updates whose sorting key contains a column that is missing in the data part and has a DEFAULT expression. #119540 (Groene AI).
• Fixed Keeper replaying its local changelog tail one entry at a time after a restart when the changelog is stored on object storage. The commit read-ahead used to discard its own prefetched entries whenever it missed the serve deadline, so on storage slower than that deadline it could never catch up and every entry cost two object reads. The default of log_readahead_commit_window_bytes is now 16 MiB instead of 500 MiB, since it bounds a prefetch window rather than a cache. #119627 (Shaohua Wang).
• Fix unbounded memory growth from time zone names of the form libc:<suffix>. These are internal interfaces of the cctz library that ClickHouse never intended to expose: any suffix was accepted, each distinct name permanently cost ~4.6 MiB that the server never gave back, and the time zone silently behaved as UTC instead of the name being reported as unknown. They are rejected now. #119634 (Alexey Milovidov).
• Fix incorrect pruning for space-filling-curve indexes with UInt64 arguments. #119663 (Minh Vu).
• Fixed sum(x + 1) OVER (...) returning a single row holding the overall total instead of the windowed result. optimize_arithmetic_operations_in_aggregate_functions, which is enabled by default, rewrote the windowed sum into sum(x) + 1 * count(x), an expression over plain aggregates that cannot carry the OVER clause, so the window definition was discarded. Both enable_analyzer = 1 (the default) and enable_analyzer = 0 were affected. #119699 (Groene AI).
• Fix substringUTF8 returning too many code points when a negative offset starts before the beginning of a string. #119840 (Minh Vu).
• Fix URLHierarchy, URLPathHierarchy, and URLHash failing to recognize valid URL schemes containing a, z, 0, or 9. #119844 (Minh Vu).
• Fix mapSubtract subtracting duplicate keys in its first argument. #119845 (Minh Vu).
• Fixed the server not validating the schema of the second and subsequent data blocks that a client sends for an external table over the Native protocol. Such a block could make the data be interpreted as a type other than the one it was sent as. #119849 (Alexey Milovidov).
• Fixed AST formatting of the APPLY column transformer: a column-name prefix containing a single quote or a backslash, or a function name that needs back-quoting, was emitted without escaping, so formatQuery returned a query that either could not be parsed or parsed back to a different query. #119851 (Groene AI).
• Fixed a crash of the server caused by a table with ENGINE = NATS whose credentials the broker started to reject, for example after a password rotation or on an expired token. Such a table now recreates its connection and resumes consuming once the credentials are accepted again, instead of staying silently idle. #119853 (Alexey Milovidov).
• Fixed timeSeriesTagsToMap treating FixedString tag names and values differently between the array form and the separate-argument form. FixedString padding is now trimmed consistently, so the same logical tags produce the same map and lookups such as mapContains(..., 'a') work in both forms. #119859 (Minh Vu).
• Fixed wrong results when join_algorithm = 'direct' reads the right-hand table of a join from a MergeTree table and plan-based parallel replicas (parallel_replicas_plan_based = 1) are enabled. Such a join could silently return only part of the rows that should have matched, with no error reported. The lookup read that the join performs is no longer distributed across replicas. #119861 (Groene AI).
• Fixed data loss where a MergeTree part whose columns.txt was left empty (zero bytes) by an interrupted metadata rewrite was detached as broken on load, discarding all of its rows. For a wide part the column list is now rebuilt from the part’s own columns_substreams.txt and the table metadata, as it already was for an absent columns.txt; when the rebuilt list cannot be validated, the part is left untouched for recovery instead. #111597 (Groene AI).
• Integers in the TSV, TSVRaw, TSKV and CustomSeparated formats can now be zero-padded (007) or carry a redundant leading + (+7), spellings that previously failed to parse but that CSV already accepted. The same reader serves other conversions from text to an integer, so WHERE i = '007' on a numeric column, query parameters and CAST('[007]' AS Array(Int64)) accept them too. A field that is a lone + remains an error, now reported as CANNOT_PARSE_NUMBER on every one of these routes instead of a route-specific code. Schema inference is unchanged: 007 still infers as String. #113910 (Groene AI).
• Fixed permanent data loss in ALTER TABLE ... EXECUTE remove_orphan_files on Iceberg tables with iceberg_use_version_hint = 1. When metadata/version-hint.text named an older version than the newest committed metadata file, every object of the newer committed snapshot, including its data files, was deleted as an orphan. Reachability is now computed from the authoritative current metadata, and a metadata listing this command cannot order unambiguously is refused instead of resolved by listing order. #114324 (Groene AI).
• Fix buffer overflows and a server abort on malformed responses from external servers: an Azure Blob Storage endpoint that returns more data than was requested or omits the ETag header, and a MySQL server that returns an oversized BIT value. Azure reads, copies and deletes, and S3 copies made by backups and by S3Queue post-processing, are now pinned to the object generation they selected, so an object overwritten in place is reported instead of silently mixing two generations. Also fix the decoding of a MySQL BIT value shorter than 8 bytes on big-endian hosts. #115706 (Alexey Milovidov).
• Fixed a crash inside the crash handler itself when a fatal signal arrives during startup of clickhouse-server, clickhouse-keeper or clickhouse-local. The handler dereferenced the global context, which does not exist yet at that point, so right after logging the stack trace it took a second fault, hung for about five minutes, and never sent the configured crash report. #116245 (Groene AI).
• Fixed CLEAR COLUMN silently leaving a MATERIALIZED column derived from the cleared column at its old value. This happened whenever the mutation had already built a stage before the recompute was appended: another pending mutation coalescing with the CLEAR into one pass, which needs no special settings, or an enabled enable_block_number_column / enable_block_offset_column. The wrong value was permanent; later merges did not repair it. #117296 (Shaohua Wang).
• Fixed ILIKE on a text index returning rows the predicate does not match, and skipping rows it does match, when the pattern or the data contains a non-ASCII code point that case-folds onto an ASCII letter. Affected indexes with a lowerUTF8/upperUTF8 preprocessor and patterns whose needle contains k. #117354 (Jimmy Aguilar Mena).
• Fixed wrong results in five cases where a WHERE predicate on a join key was pushed to the other side of a JOIN whose two keys have different types. Such a key is no longer substituted when the keys’ common type is a float, because -0.0 and +0.0 are join-equal but differ bit for bit; nor when it is Dynamic or JSON, whose runtime contents the static type does not describe; nor when the key is unstable within the query or changes the row count; nor in a conjunct that reads a column’s representation rather than its value, as isConstant and toColumnTypeName do, including when such a read or an unstable function is hidden inside a lambda body. Separately, with enable_analyzer = 1, a predicate on the wider of two differently typed join keys is now also used as an index condition on the other table of a plain INNER JOIN, so it is read through its primary key instead of in full; previously only ANY INNER JOIN, LEFT JOIN and RIGHT JOIN got this. #117973 (Groene AI). Also fixed a LOGICAL_ERROR (Unexpected return type from _CAST) when such a predicate was pushed down to a side that join_use_nulls widens to Nullable. #120109 (Groene AI).
• Fix rows silently dropped by index analysis for has(array_of_named_tuples, key) when the tuple fields are declared in another order and the key is a function of the column. #118706 (Alexey Milovidov).
• Fixed ClickHouse Keeper returning ZNOAUTH for every request on a node whose ACL list is empty but is stored under a nonzero internal ACL id, which happens for Keeper data converted from a ZooKeeper snapshot or written by a very old Keeper. Such nodes are unrestricted again, as they were before 26.6. Closes #117603. #118965 (Groene AI).
• Fixed YYYYMMDDhhmmssToDateTime64 returning incorrect fractional seconds for timestamps before 1970-01-01. #119240 (Raúl Marín).
• Fixed a crash when a disk is configured with metadata_type = web or metadata_type = web_index on top of a non-web object storage. Such a configuration is now rejected with INVALID_CONFIG_PARAMETER instead of interpreting an unrelated object storage as a web one. Closes #119798. #119818 (Groene AI).
• Iceberg tables recovered from stored metadata no longer enable background compaction or old-file cleanup unless those settings are explicitly stored in the table definition. #119977 (János Benjamin Antal).
• Fixed a NATS table whose settings all come from a named collection (ENGINE = NATS(collection) with no SETTINGS clause) being written to its metadata file with a dangling SETTINGS keyword. The stored definition could not be parsed again, so the table failed to load and aborted the load of its whole database, making unrelated tables in it inaccessible. Closes #119120. #120025 (Groene AI).
• Fixed non-idempotent formatting of a function named SOME: the name was printed unquoted, so the server read its own output back as the SOME array quantifier. formatQuery('SELECT 1 < SOME(a, b -> 1)') returned a query that re-parses to arrayExists, and a user-defined function whose body compared against `some`(...) was persisted without the quotes and then failed to load after a restart. #120096 (Groene AI).
• Fix reading a column that has two streams whose names differ only by escaping, for example a Tuple element named size0 inside an Array, or a typed JSON path named like a substream of a sibling path. Such columns returned wrong values, failed with CANNOT_READ_ALL_DATA or another exception, and made the table unmergeable. #115670 (Pavel Kruglov).
• Fix least, greatest, bitShiftRight, midpoint, and avg2 returning wrong results when a JIT-compiled expression has an Int128 result. With the default compile_expressions = 1, a negative Int128 result could be treated as unsigned after a few executions. #118363 (Groene AI).
• Fix ttl_only_drop_parts preventing expired columns from ever being cleared. A column TTL requires rewriting the part, so with this setting enabled an expired column could be kept forever instead of being cleared. Column-TTL rewrites are now scheduled separately and still run when ttl_only_drop_parts is enabled. #118428 (Alexey Milovidov).
• Fix mergeTreeTextIndex bypassing row policies: reading the text index is now denied whenever a row policy applies to the table, regardless of which columns the policy uses. #119032 (Elmi Ahmadov).
• Fix force_primary_key incorrectly rejecting some queries with INDEX_NOT_USED when optimize_extract_common_expressions (enabled by default) simplifies the filter to a trivial primary-key condition wrapped in an implicit cast. #119083 (kasimtj).
• Fix groupNumericIndexedVector returning a value that is too large when an index repeats in the input, and returning different results depending on max_threads. Also fix numericIndexedVectorPointwiseMultiply and numericIndexedVectorPointwiseDivide dropping zero-valued indices, and numericIndexedVectorPointwiseDivide by a vector of all ones keeping the value of an index the divisor is missing instead of zeroing it. A groupNumericIndexedVectorState already stored in a table is not repaired by the upgrade and has to be recomputed. #119564 (Alexey Milovidov).
• isNull and IS NULL now return 1 for every NULL row, including when the column’s NULL mask holds a byte other than 0 or 1, which the Native format permits and which if can produce from its condition column. Previously such a row could contribute its raw byte, so sum(isNull(e)) returned 30 where 20 is correct. if(cond, x, NULL) with a Nullable x no longer returns NULL for a row whose condition byte was neither 0 nor 1. If an older server stored a value computed from isNull on such a column (sorting key, skip index, partition key, projection, MATERIALIZED or DEFAULT column), rebuild it. #119785 (Groene AI).
• Preserve explicit time zones in timeSeriesRange DateTime64 results. #119858 (Minh Vu).

​

Build/Testing/Packaging Improvement

• Added a CMake file for compiling ClickHouse with the nvcc compiler. Related: #116807. #117734 (Konstantin Vedernikov). Added the cudf contrib library. #117869 (Konstantin Vedernikov).
• Added a build option to select 4 KiB or 16 KiB jemalloc page size on AArch64 instead of the default 64 KiB. #115424 (Azat Khuzhin).
• Fixed several portability issues on illumos, including missing termios.h, register-macro conflicts, and floating-point environment casts. #117363 (Joshua Carp).
• ClickHouse is now built as a position independent executable on Linux amd64 and aarch64 with glibc, including the release builds, so the text segment is subject to ASLR. Addresses in the bare stack trace of a fatal log are now file offsets relative to the object that contains them, annotated with that object for frames outside the main executable, so they can be resolved with addr2line -e <object> <hex> regardless of where the binary was loaded. For the same reason system.errors.last_error_trace, system.error_log.last_error_trace and system.crash_log.trace now store a frame of the main executable as its file offset, keeping runtime addresses for frames in other objects. #117449 (Groene AI). #119284 (Alexey Milovidov).
• Fixed compilation of benchmark_statistics with ENABLE_BENCHMARKS=ON. #116292 (Groene AI).
• Use cctz tzdata 2026c. #116539 (Konstantin Bogdanov).
• Restored the bundled aws-c-http library to v0.11.0 after it had been accidentally reverted to an older submodule revision. #116911 (Alexey Milovidov).
• Fix building ArrowFlight protobuf files with the CMake Unix Makefiles generator by creating the protoc output directory before invoking protoc. #116879 (Alexey Bakharew).
• Reduced compile time and object file sizes by removing redundant template instantiations in comparison functions, binary arithmetic, arrayDistance, hashed dictionaries, and argMin/argMax. #117540 (Konstantin Bogdanov).
• Fix building unit_tests_dbms when ClickHouse is configured with -DENABLE_LIBRARIES=OFF (and -DENABLE_RUST=OFF). #105684 (Ivan Babrou).
• Harden official builds by enabling full RELRO: the global offset table is now mapped read-only after startup. #119037 (Raúl Marín).
• Cross-compilation sysroots updated: FreeBSD builds now target FreeBSD 14 (13 is end-of-life), riscv64 and loongarch64 builds use current Debian glibc and kernel headers. Unused files were removed from the sysroot submodule. #119049 (Konstantin Bogdanov).
• Fixed startup of the shell-based Docker server images when an absolute filesystem_caches path is a missing directory below a directory the server does not own. The entrypoint now prepares these cache directories with the same ownership handling as storage_configuration disk paths. #118842 (thien-ch).
• Update openssl to 3.5.8. #118927 (Actuele AI).
• Update xxHash to v0.8.3. #117597 (Actuele AI).
• Update libxml2 to 2.15.4. #118649 (Actuele AI).
• Update libdivide to v5.4.0. #118660 (Actuele AI).
• Update Jieba-CPP to v1.0.0. #118889 (Actuele AI).
• Update minizip-ng to 4.2.2. #118886 (Actuele AI).
• Update miniselect to 0.4.0. #118887 (Actuele AI).
• Update openldap to 2.6.15. #118888 (Actuele AI). #119359 (Actuele AI).
• Update sparsehash-c11 to v2.11.1. #118891 (Actuele AI).
• Bump croaring from v4.5.1 to v5.1.1. Serialized bitmaps are unchanged on little-endian platforms; on big-endian platforms the on-disk layout of groupBitmap states changes to the portable little-endian format. #118917 (Actuele AI).
• Update xz to v5.8.4. #119360 (Actuele AI).
• Update librdkafka to v2.15.1. #119467 (Actuele AI).
• Update simdjson to v4.6.11. #119508 (Actuele AI).
• Fixed building ClickHouse with clang 23, which emitted a lifetime-safety diagnostic that was suppressed only for clang 24 and later. #118005 (Navneet Kumar). #119211 (mosya415).

​

ClickHouse release 26.8 LTS, 2026-08-27. Presentation, Video

​

Backward Incompatible Change

• The X-ClickHouse-Format HTTP request header now explicitly selects the format of the response: it is an alias for the output_format setting, so it overrides the FORMAT clause of the query and the path extension. In previous versions, it only set the default when the output format was unspecified, like the default_format parameter. This provides an “out-of-band” option to request data in a desired format. It never changes how the request body of an INSERT is parsed - use input_format or format for that. #105249 (Alexey Milovidov).
• Unquoted JSON numbers for DateTime and DateTime64 columns in JSONEachRow and similar formats are now interpreted as Unix timestamps with optional sub-second precision, consistent with Values, CAST, and toDateTime64. Previously, a fractional number such as 1703363853.035 was rejected, and a bare integer such as 1703363853 was read as the raw scaled value of DateTime64, producing a 1970-... timestamp. Quoted strings and ClickHouse’s own JSON output are unaffected. #108091 (Alexey Milovidov).
• The default value of max_insert_threads changed from 1 to auto, which resolves to the number of CPU cores available to the server. This parallelizes INSERT SELECT by default, and max_insert_threads can now also parallelize the writing side of a plain INSERT when the destination write path can safely fan out. This can change the number of parts created by such queries and the order of inserted rows. To restore the previous behavior, set max_insert_threads to 1, or set compatibility to a version below 26.8. #109000 (Alexey Milovidov). #109006 (Alexey Milovidov).
• The PostgreSQL and MaterializedPostgreSQL database engines now respect remote_url_allow_hosts, like the PostgreSQL table engine, the postgresql table function, and DDL-created dictionaries already do. With remote_url_allow_hosts configured, CREATE DATABASE and user-issued ATTACH DATABASE now reject PostgreSQL and MaterializedPostgreSQL databases that point at disallowed hosts with UNACCEPTABLE_URL. Existing databases still load at server startup. #111135 (Alexey Milovidov).
• The experimental ALP(STD) codec now performs Float32 scaling arithmetic in Float64. This improves compression ratios and eliminates exception-heavy compression of decimal data, but Float32 values written by earlier versions may now decode 1 ULP differently. #111627 (Raufs Dunamalijevs).
• Removed the library dictionary source. SOURCE(LIBRARY(...)) now fails with UNKNOWN_ELEMENT_IN_CONFIG, and the dictionaries_lib_path server setting is obsolete and has no effect. #111980 (Alexey Milovidov).
• Removed the Apache Arrow library-based reader and writer for the Arrow and ArrowStream formats. The native ClickHouse implementation, which has been the default since 26.7, is now the only one. The settings input_format_arrow_use_native_reader and output_format_arrow_use_native_writer are obsolete: they are still accepted, but have no effect, so a query that set them to 0 to force the Apache Arrow implementation now uses the native one. #111996 (Alexey Milovidov).
• Reject timespan setting (milliseconds/seconds) values that overflow Int64 microseconds. #112757 (Azat Khuzhin). This might be an incompatibility only when you used absurdly large values that overflowed.
• TLS credentials of a MySQL source (ssl_ca, ssl_cert, ssl_key) can no longer be specified as file paths from SQL, such as in CREATE NAMED COLLECTION, query arguments, or CREATE DICTIONARY, because the server opens those files with its own privileges. Paths remain supported in the server configuration file. Elsewhere, pass the certificate or key contents in the new ssl_ca_pem, ssl_cert_pem, and ssl_key_pem parameters, which are masked in logs and in SHOW queries like passwords are. #112070 (Alexey Milovidov).
• Privilege checks for the SYSTEM ... CACHE ON CLUSTER commands now use each command’s own privilege instead of the SYSTEM DROP CACHE privilege group. This lets holders of a single granular cache privilege run its matching command, and prevents a holder of the group from running SYSTEM SYNC FILESYSTEM CACHE ON CLUSTER without SYSTEM SYNC FILESYSTEM CACHE. #114042 (Groene AI).
• For views whose body is a plain SELECT over a single Distributed table, the whole outer query is now pushed to the shards (new setting optimize_trivial_view_pushdown_to_distributed, enabled by default). This changes observable behaviour for such views: FINAL and SAMPLE written on the view reference are now propagated to the shard-local table instead of being ignored, and extremes is not reported on single-shard clusters. Set optimize_trivial_view_pushdown_to_distributed = 0 to restore the previous behaviour. #101791 (simonmichal).
• Lightweight UPDATE patch parts now use a new v2 on-disk format sorted by (sorting_key..., _block_number, _block_offset) and applied with a new merging algorithm. Peak memory is bounded by the largest equal-sort-key run instead of the full patch, and updates that cross merge boundaries no longer fall back to in-memory Join apply. Old-format patch parts remain readable. During a rolling upgrade from a version before 26.8, keep patch_parts_version = 'v1' or use the compatibility setting until all replicas are upgraded. #103182 (Anton Popov).
• Added the VALID FOR <interval> clause to CREATE USER and ALTER USER as a shorthand for VALID UNTIL. The expiration deadline is computed as the current time plus the given interval at query execution time and stored in the VALID UNTIL form. The valid_until column of the system.users table now has the type Array(DateTime64(0)) instead of Array(DateTime), so that deadlines beyond the year 2106 are represented exactly; tooling that reads this column should handle the new type. #110171 (Alexey Milovidov).
• EXPLAIN SYNTAX now returns the reformatted query as a single String record (with embedded newlines) instead of one record per line, so its output is a recoverable single row that is directly usable (for example, SELECT count() FROM (EXPLAIN SYNTAX ...) returns 1). This is controlled by the new single_record option, which defaults to 1; set single_record = 0 to restore the historical one-record-per-line output. Other EXPLAIN kinds (PLAN/PIPELINE/AST) keep their per-line tree output. #110479 (Alexey Milovidov).
• Extended the supported range of Date32 from [1900-01-01, 2299-12-31] to [0000-01-01, 9999-12-31], matching DateTime64. Parsing and conversions now accept the extended range instead of silently clamping to the old boundaries. Backward compatibility notes: in the numeric conversion toDate32(N), values in [120530, 2932896] are now interpreted as day numbers (dates from 2300-01-01 to 9999-12-31) instead of Unix timestamps in early 1970, matching the rule that a number that fits into the day-number range is a day number; numbers below the day number of 0000-01-01 and timestamps after 9999-12-31 saturate to the new boundaries. #111534 (Alexey Milovidov).
• arrayIntersect and arraySymmetricDifference no longer treat a value repeated inside a single argument as if it appeared in several arguments. Queries relying on the previous behavior can now return different results: arrayIntersect([1], [2], [1, 1]) returns [] instead of [1], arrayIntersect([1, 2], [2], [1, 1, 2]) returns [2] instead of [1, 2], and arraySymmetricDifference([1], [2], [1, 1]) returns [2, 1] instead of [2]. For arraySymmetricDifference two arguments are already enough: arraySymmetricDifference([1], [2, 2]) returns [2, 1] instead of [1]. A value is now counted for an argument only when it was present in every argument before it, so the result contains exactly the values present in all of the arguments. As part of the same change, arrayIntersect builds its hash table from the smallest argument rather than from all of them, which makes it up to 1.85x faster and use a third less memory when the arguments differ a lot in size. arrayUnion is not affected. #113021 (Alexey Milovidov).
• disable_insertion_and_mutation now prevents background consumption from Kafka, RabbitMQ, and NATS tables while still allowing direct writes to external storage. Gated Kafka2, NATS, and RabbitMQ tables do not initialize consumers for direct SELECT. The recently introduced message_queue_disable_insertion setting now requires a server restart. #113660 (Christoph Wurm).
• A window PARTITION BY or ORDER BY over an AggregateFunction column is now rejected with ILLEGAL_COLUMN, as top-level ORDER BY over such a column already was. Previously such a query was accepted by at least one analyzer, and window PARTITION BY partitioned differently depending on max_threads. The refusal also covers a state nested in Array, Tuple, Map, Variant or SimpleAggregateFunction. A SimpleAggregateFunction over an ordinary type, QBit, and GROUP BY and DISTINCT over a state, are unaffected. #113878 (Groene AI).
• The include_from setting no longer defaults to /etc/metrika.xml. Previously that file was used for configuration substitutions whenever it existed, even though nothing in the ClickHouse configuration referred to it. If you relied on that implicit substitutions file, add <include_from>/etc/metrika.xml</include_from> explicitly to each affected config file; separately loaded users.xml and XML dictionary configs need their own include_from element. #114161 (Alexey Milovidov).
• The NATS table engine accepts credentials inline in the new nats_credentials setting (the same payload as a .creds file), and no longer accepts nats_credential_file from SQL: the path is a reference to a file on the server filesystem, which the server opens with its own privileges, so it can only be specified in a named collection defined in the server configuration file, or as nats.credential_file in the server configuration itself. A query may replace such a configured path with inline nats_credentials, unless the operator pinned it with <nats_credential_file overridable="false">. Tables created before this restriction keep working after an upgrade. #114644 (Alexey Milovidov).
• Asynchronous metrics can now have the Map data type, and the per-CPU-core and per-device metrics were converted to single key-value metrics: OSUserTimeCPU0, OSUserTimeCPU1, … became a single OSUserTimeCPU metric with a map from the CPU core number to the value (and similarly for the other OS*TimeCPU* metrics, CPUFrequencyMHz_*, Temperature*, EDAC*, Block*_*, Network(Receive|Send)*_*, Disk*_*, *BlobsQueueEstimate, and AsyncLogging*QueueSize). system.asynchronous_metrics has a new key_values Map(LowCardinality(String), Float64) column (the value column is NaN for such metrics), system.asynchronous_metric_log logs them as one row per key using a new key column, the Prometheus endpoint exports them with a label (e.g. ClickHouseAsyncMetrics_BlockReadBytes{device="sda"}), and the Graphite MetricsTransmitter sends them as <prefix>.<Metric>.<key>. If your monitoring reads the old metric names, set the new asynchronous_metrics_key_values_mode server setting to legacy_names to publish every key as a separate scalar metric under its previous name again, or to both to publish both forms at the same time while the monitoring is being migrated; the setting is applied by SYSTEM RELOAD CONFIG, without a restart. #115333 (Alexey Milovidov). #116791 (Alexey Milovidov).

​

New Feature

• Added CREATE HANDLER, ALTER HANDLER and DROP HANDLER statements to define custom HTTP handlers from SQL, persisted in a local or Keeper storage. Added the currentHandler and currentRequestURL functions, the system.handlers table, and the http_handler_name / http_request_url columns in system.query_log. #106231 (Alexey Milovidov).
• Added a way to access tables and databases via URL paths in the HTTP interface (e.g. `/database/table.format.gz?filter=a>0`), plus new settings (`http_allow_database_as_path`, `http_allow_table_as_file`, `http_allow_filters_as_path`, `http_allow_filters_as_unrecognized_url_parameters`, `select`, `order`, `sort`, `filter`, `page`, `compression`, `format`, `input_format`, `output_format`, `default_format`, `database`) that compose with one another and with the existing `query` URL parameter. #105249 (Alexey Milovidov).
• Add framing formats, selected by the new setting framing_output_format: they multiplex different response parts of the query in a single HTTP response stream — chunks of data, totals and extremes, progress packets, profile events, server logs, and exceptions. Implemented framing formats: None (default, everything works as before), EventStream (HTTP server-sent events), JSONEachPacketBase64, and JSONEachPacketString (a JSON object per packet with base64-encoded or string data). #110127 (Alexey Milovidov).
• Added the default_session_user server setting: the user to authenticate as when a client connects without specifying a user name (previously hardcoded to default; the native, MySQL, and PostgreSQL protocols now accept an empty user name instead of rejecting it, and Arrow Flight honors the setting instead of its own hardcoded default). Composable protocol endpoints in the protocols section can override it per endpoint with a default_session_user key, so different listening ports can serve different anonymous users. Setting it to an empty string prohibits connections without a user name. Interserver connections never use the default session user. #110179 (Alexey Milovidov).
• Support pipe operators in SQL queries: FROM t |> WHERE x > 1 |> AGGREGATE count() AS c GROUP BY y |> ORDER BY c DESC |> LIMIT 10, similar to the pipe syntax of GoogleSQL. Any SELECT query can be followed by a chain of |> operators, and each operator wraps the query before it into a subquery, so the resulting AST is the same as for the equivalent query with nested subqueries. In a query that starts with the FROM clause, the SELECT clause is now optional and defaults to SELECT *. #111151 (Alexey Milovidov).
• Added the parseQueryToJSON and formatQueryFromJSON functions to serialize and deserialize ASTs as JSON, and an experimental clickhouse_json dialect enabled by enable_json_ast_dialect. #100412 (Alexey Milovidov). #113480 (Nikita Fomichev).
• New setting run_query_in_background. The server accepts the query, immediately returns an empty result, and runs it to completion regardless of what happens to the connection. The result is discarded. Track the query by its query_id in system.processes and system.query_log. Intended for long queries like INSERT ... SELECT, CREATE TABLE … AS SELECT, or CREATE MATERIALIZED VIEW … POPULATE that must not die with a dropped connection. #112816 (Miсhael Stetsyuk).
• Added a new system table, system.user_query_log, which shows every user their own query log records without requiring access to the query log table. The original implementation and the idea belong to Yue Ni (@niyue), #104462. #110156 (Alexey Milovidov).
• Implement the remaining items of the URL unification (issue #59617). The new s3_base setting resolves relative URLs in the s3 table function and the S3 table engine (and the functions sharing their configuration: s3Cluster, gcs, oss), similarly to url_base. The new URL database engine treats table names as URLs, resolves them against an optional base URL, and dispatches by the URL scheme, so files, web and object storage are handled uniformly. It replaces Filesystem in the default (Overlay) database of clickhouse-local with a file:// base URL, so plain file names keep working, and you can also write SELECT * FROM 'https://example.com/data.csv'. #111512 (Alexey Milovidov).
• Added the bigquery table function and the BigQuery table engine for reading from and writing to Google BigQuery tables, including public datasets. The table structure is inferred from the BigQuery table schema. Authentication supports an OAuth access token, a service account key in JSON format, and an OAuth client with a refresh token. #110166 (Alexey Milovidov).
• Plain CREATE MATERIALIZED VIEW ... POPULATE is now locally atomic: the view is subscribed to new inserts of the source table and the existing data is snapshotted together under a brief exclusive lock on the source, so rows inserted through the same server concurrently with the population are no longer missed or duplicated (new setting materialized_views_populate_atomically, enabled by default). The guarantee covers the local insert path only; inserts arriving on another replica or through a distributed write path are outside the cut, and it requires a source that can provide a pinned snapshot (the MergeTree family and Memory); other sources, as well as CREATE OR REPLACE / REPLACE and views created in Replicated databases, keep the legacy non-atomic population. POPULATE can now also be used together with TO to backfill the target table. #108715 (Alexey Milovidov).
• Keeper can store data on disk now, in a custom LSM tree. It has similar performance to the previous storage. Use coordnation setting use_lsmt_storage = true to enable, storage_memory_only = false to store files on disk, add data_storage_path or data_storage_disk to config to specify where to put the files. To write to S3, point data_storage_disk to a disk of type s3_plain. #113903 (Michael Kolupaev).
• The “S3 tables” engine catalog for data lakes, the same as #103220 but with a working INSERT. #113505 (Konstantin Vedernikov).
• Add support for Snowflake Horizon. You can now query Iceberg table in Iceberg behind the horizon catalog. You can also write to the Iceberg table via the catalog. #114547 (Melvyn Peignon).
• Added support for the Puffin file format. It can be used with the file, url, s3, and similar table functions. #103936 (Konstantin Vedernikov).
• Support the GROUPS frame mode for window functions (SQL:2011), e.g. any(price) OVER (PARTITION BY symbol ORDER BY ts GROUPS BETWEEN CURRENT ROW AND 1 FOLLOWING). In a GROUPS frame the boundaries count whole peer groups — sets of rows that are equal on the ORDER BY key — so N PRECEDING/N FOLLOWING mean N peer groups before/after the current row’s peer group, rather than physical rows (ROWS) or ORDER BY value distances (RANGE). #108653 (Nihal Z. Miaji).
• Added the mergeTreeCodecBlockCounts table function, which reports for each part, column, and substream of a MergeTree table how many compressed blocks use each codec. #109623 (Raufs Dunamalijevs).
• Added the MultiPoint geo data type, stored as Array(Point), and included it in the Geometry type. #109951 (David Meng).
• Added the _etag virtual column for HDFS storage. #108255 (Zhang Yifan).
• The native ORC reader and output format now support uniontype, mapping it to the ClickHouse Variant type. Previously, reading such columns failed with Unsupported ORC type, and writing them failed with ILLEGAL_COLUMN. #110078 (Alexey Milovidov). #110085 (Alexey Milovidov).
• Added a japanese tokenizer for text indexes and the tokens, hasAnyTokens, and hasAllTokens functions, using the MeCab morphological analyzer. The dictionary is loaded at runtime from a location set in the server configuration and verified against a SHA-256 checksum before use. #111420 (Jimmy Aguilar Mena).
• Added a chinese tokenizer for the tokens function and MergeTree text indexes. It segments Chinese text into words using a dictionary and a Hidden Markov Model (the algorithm follows jieba), with coarse_grained (default) and fine_grained granularities. Continues #80174. #89945 (Amos Bird).
• Added a new icu tokenizer for text indexes and the tokens, hasAnyTokens, and hasAllTokens functions, providing locale-aware Unicode word segmentation (including dictionary-based segmentation for languages such as Chinese, Japanese, and Thai). #109940 (Jimmy Aguilar Mena).
• Added a splitByRegexp tokenizer for text indexes and the tokens function. It splits the input into tokens using a regular expression as the separator, for example tokenizer = splitByRegexp('[^\p{L}\p{N}#+]+'), which allows preserving tokens containing special characters such as C++ or C# that other tokenizers would break apart. #110002 (Jimmy Aguilar Mena).
• The Keeper HTTP dashboard now includes a Cluster tab that shows Raft membership as a topology graph with health colors, lag, and priority, so you can see which nodes are healthy, lagging, or unreachable and open the leader’s dashboard directly. #111655 (yanglongwei).
• A new server configuration section, create_union_system_log_tables, requests the creation and automatic maintenance of all_... tables (e.g. system.all_query_log) that query the union of a system log table, its rotated versions and/or the same tables across all replicas of a cluster. The remote, remoteSecure, cluster and clusterAllReplicas table functions now accept a trailing SETTINGS clause with the settings of the Distributed storage they create, e.g. clusterAllReplicas('default', system.query_log, SETTINGS skip_unavailable_shards = 1). #112670 (Alexey Milovidov).
• Add nats_credentials setting to the NATS table engine, allowing users to specify NATS credentials inline as a string (matching the payload of a .creds file). #110733 (addshore).
• Added the sort-based IEJoin algorithm for joins whose ON clause has two inequality comparisons (<, <=, >, >=) between the joined tables, enabled by adding ie_join to the join_algorithm setting. Supported kinds are ALL INNER/LEFT/RIGHT/FULL JOIN and SEMI/ANTI LEFT/RIGHT JOIN. Previously such queries were executed as a CROSS JOIN with a filter (INNER only), which is much slower on large tables. #109920 (Vladimir Cherkasov).
• Added the aggregate function gini, which computes the Gini coefficient of finite, non-negative numeric values, including grouped and distributed aggregations. #112280 (Amirreza Akhondi). #114643 (Groene AI).
• Added the HiveText output format, which writes Apache Hive LazySimpleSerDe text. Fields are separated by \x01; rows by the format_hive_text_rows_delimiter setting (default \n); and nested Array, Map, and Tuple values use Hive’s separator list. The output targets Hive’s default LazySimpleSerDe and is not symmetric with ClickHouse’s HiveText input for nested values or custom row delimiters. #107582 (Alexey Milovidov).
• Added support for constant labels on the Prometheus metrics endpoint: a new <labels> section inside <prometheus> adds the configured labels to every exposed metric, which helps distinguish multiple clusters scraped by the same Prometheus. Label values support standard config substitutions such as from_env. Closes #85969. #111672 (Valery Petrov).
• Added the always_fetch_mutated_part setting. When enabled, a replica can download mutated parts from another replica instead of executing mutations locally. #113020 (Rory Shanks).
• Added the aggregate function mergedJSONPatch for RFC 7396-style merge-patch aggregation over JSON values, including state-merge support and AggregatingMergeTree usage. Users can aggregate a JSON column using a timestamp or another ordering column as the sort key to compute the latest state. #108349 (larryluogit).
• Added the aiSimilarity function, which computes the semantic similarity between two texts using an embedding model. It returns a Nullable(Float32) in [-1, 1], where 1 means the texts are identical, or NULL if an operand is NULL/empty or its embedding failed. #110777 (David Meng).
• Add restore_table_data, restore_access_entities, and restore_functions restore settings for granular control over what gets restored. These can override structure_only for individual categories — e.g. structure_only=true, restore_access_entities=true restores table definitions and access entities without table data. #102402 (Nikita Fomichev).
• The array subscript operator supports an array of integers as the index: arr[indexes] returns the elements at all of the given positions, equivalently to arrayMap(i -> arr[i], indexes). #108371 (folly).
• Add function notHas, the negation of has for arrays, maps, and JSON. When the haystack is a constant array, notHas(constant_array, x) is rewritten to x NOT IN constant_array by optimize_rewrite_has_to_in (enabled by default), so it executes via a set lookup and can prune by the primary key index like NOT IN. #109926 (Nihal Z. Miaji).
• Functions arrayElement (the vec[n] operator) and arraySlice now work for the QBit data type: qbit[n] returns the n-th vector element at full precision, and arraySlice(qbit, offset, length) returns a projection to a subset of dimensions. Both read only the bit planes of the stride groups they need, and slices aligned to stride-group boundaries reuse the stored streams without copying. This closes #109942. #109953 (Utkal Singh).
• Support animated PNG in the PNG output format. A t column turns the result of a query into an animation, with the time scale of t set by output_format_image_time_multiplier_seconds and output_format_image_time_divisor_seconds, and output_format_image_streaming_animation to write the frames out as the query produces them instead of buffering them in memory. #112846 (Alexey Milovidov).
• New function aiRedact that detects and redacts personally identifiable information (PII) in text using an LLM provider. Specify the categories to redact (e.g. ['email', 'name']) or pass an empty array to use a default set of PII categories. Matched values are replaced with a token ([REDACTED] by default). #110464 (David Meng).
• Add aiFilter function that evaluates a natural-language condition against text with an LLM and returns UInt8 for use in WHERE, PREWHERE, and JOIN ... ON. #110594 (yanglongwei).
• Support TLS/SSL connections to PostgreSQL for the PostgreSQL table engine, the postgresql table function, the PostgreSQL and MaterializedPostgreSQL database engines, and PostgreSQL dictionaries: sslmode plus the certificates and the key, given either as literal contents (sslrootcert_pem, sslcert_pem, sslkey_pem; masked like passwords) or as paths (sslrootcert, sslcert, sslkey; accepted only from a named collection defined in the server configuration file). #110615 (Alexey Milovidov).
• ClickHouse server now has an introspection port. This is a native protocol TCP listener that starts before the server begins attaching tables and stops only after the tables’ detach completes. During these windows, an operator can connect to it with clickhouse client and run queries such as SHOW PROCESSLIST, SELECT * FROM system.stack_trace, or SYSTEM INSTRUMENT ADD 'QueryMetricLog::startQuery' SLEEP ENTRY 0.5. Additionally, the default value of the shutdown_wait_unfinished server setting is increased from 5 to 120 seconds. The previous default was smaller than the connection poll interval, so the server could force exit in the middle of shutdown instead of waiting for outstanding connections to finish. #110838 (Miсhael Stetsyuk).
• Introduce emulated ZooKeeper metrics in Keeper mntr command: zk_leader_uptime, zk_sum_election_time, zk_cnt_election_time, zk_sum_leader_unavailable_time, zk_cnt_leader_unavailable_time; Also introduce related leader-oriented metrics for Keeper only: KeeperLastLeaderElectionTime, KeeperLastLeaderUnavailableTime. #113334 (Maxim Orlovsky).
• Support ALTER TABLE … MODIFY PROJECTION to change projection-level settings (e.g. index_granularity) of an existing projection without rebuilding it. The new settings apply lazily via merges. #113343 (Diego Gomes Tomé).
• Added a finish_time column to system.mutations that records when a mutation was completed. Unfinished mutations and mutations whose completion time is unknown report zero. #113474 (Nikita Mikhaylov).
• Added a new private preview feature tier between experimental and beta, so feature tiers are now ordered experimental < private preview < beta < production. The server setting allow_feature_tier gains a level: 0 allows all tiers, 1 excludes experimental, 2 additionally excludes private preview, and 3 allows production settings only. Configurations that previously pinned allow_feature_tier = 2 to allow production settings only must now use 3. Closes #113481. #113581 (Groene AI).
• The built-in Web SQL UI (/play) can now sort a result by its columns, filter it by their values, and page through it, without editing the query. Sorting arrows and a filter input appear in every column header, a selected cell offers the comparisons that fit its value, and a pager with an adjustable page size appears under a result that does not fit on one screen. It is all performed by the server, using the order, filter, limit and page query-construction settings, so the whole result is shaped rather than the rows that happen to be displayed; the chosen sort, filters and page travel in the page URL, so a shared link reproduces the result. #115540 (Alexey Milovidov).
• chdig v26.8.1: ratatui, tmux panes, tmuxinator sessions, sharing logs with colors, feature tour. #115216 (Azat Khuzhin).

​

Experimental Feature

• Added an experimental Cascades cost-based optimizer for distributed query plans, enabled by enable_cascades_optimizer = 1 together with make_distributed_plan = 1. It chooses between shuffle, broadcast, replicated, and local join strategies, two-phase, shuffle, and local aggregation, two-stage distributed top-N, and parallel and replicated reads by estimated cost, inserting exchange operators as needed. #86353 (Alexander Gololobov).
• Added an experimental MergeTree setting enable_adaptive_codec_selection. When enabled, merges and mutations pick the smallest-output codec per block for columns that use the default codec. #111834 (Raufs Dunamalijevs). * Adaptive codec selection now compresses with each candidate codec at most once per block instead of measuring all candidates and then recompressing the winner. #113511 (Raufs Dunamalijevs).
• The clickhouse binary builds for WebAssembly (wasm64, via Emscripten), and clickhouse local runs queries, including MergeTree tables, under Node.js ≥ 24 and in the browser. A new Build (wasm64) CI job pins the build and the Node.js execution path (the browser path is not covered by CI yet) and provides the WebAssembly module as an artifact. #113404 (Alexey Milovidov).
• parallel_replicas_plan_based now chooses the local/remote boundary by optimizing the query plan, can parallelize queries over views with a UNION without parallel_replicas_allow_view_over_mergetree, falls back to the local plan for IN (subquery) queries instead of failing, and now also distributes eligible JOINs (INNER, LEFT, RIGHT, including SEMI/ANTI). #111063 (Igor Nikonov). #112268 (Igor Nikonov). #112443 (Igor Nikonov).
• Distributed query plans (make_distributed_plan) now also support queries whose source is known at planning time to produce no rows, such as reads from an empty MergeTree table. #111941 (Groene AI).
• Added the use_projection_index_in_read_pools setting, disabled by default. When enabled, ranges fully filtered out by a projection or skip index are dropped inside MergeTree read pools before read tasks are created for them, instead of being skipped granule by granule during reading. This avoids creating read tasks and setting up readers for data that will not be read. #110291 (Nikolai Kochetov). #112296 (Anton Popov).
• PromQL: fixed operations on instant vectors without tags, e.g. vector(1) + vector(2), topk(1, vector(1)), label_replace(vector(1), ...) and binary operators with on(), which failed with a type error. #111872 (Valery Petrov).
• PromQL: implemented the changes and resets functions. #112352 (Valery Petrov).
• Report the error that cancelled a distributed query (make_distributed_plan = 1) instead of a bare QUERY_WAS_CANCELLED. A failing worker task’s real error was previously visible only in the server log. #112590 (Alexander Gololobov).
• When make_distributed_plan = 1, ClickHouse now disables features that distributed query plans do not support yet instead of letting them fail later. Closes #109476. #112463 (alesapin).
• Distributed query plans (make_distributed_plan) now use lazy materialization for ORDER BY ... LIMIT k queries. #113111 (Shankar Iyer).
• Support non-partitioned window functions (OVER (ORDER BY ...)) in distributed query plans (make_distributed_plan). Previously any query with a window function failed because WindowStep could not be serialized for remote execution; windows with PARTITION BY are still rejected for now. #109802 (Vighnesh Pathrikar).
• Added the setting use_query_condition_cache_for_top_k, which gates the query condition cache for ORDER BY ... LIMIT n (TopK) queries. #111492 (Alexey Milovidov). The query condition cache is now enabled by default for queries that use the ORDER BY <column> LIMIT n (TopK) optimization. It can be turned off again with the setting use_query_condition_cache_for_top_k. #114539 (Alexey Milovidov).
• Added static partition-to-shard affinity for StorageKafka2 via the kafka_partition_shard_num and kafka_shard_count settings, allowing multiple ClickHouse shards to deterministically consume disjoint subsets of Kafka partitions. #108886 (johnjing).
• Harden the experimental AI functions: deny insecure (http) endpoints to remote hosts by default (new setting ai_function_allow_insecure_endpoint), bound outbound API calls per query by default (ai_function_max_api_calls_per_query now defaults to 1000), and sanitize provider error responses written to the logs. #111286 (George Larionov).
• PromQL: quantile with a quantile level outside [0, 1] now returns -Inf/+Inf/NaN like Prometheus instead of throwing an error, and histogram_quantile now ignores input series whose le label is missing or unparsable, like Prometheus. #111871 (Valery Petrov).
• Allow multi-component identifiers in TimeSeries tables. #112799 (Vitaly Baranov).
• Added TimeSeries settings samples_index_granularity and tags_inner_granularity to control the index granularity of inner tables. #113075 (Vitaly Baranov).
• PromQL: added the deriv function and fixed a bug where its backing aggregate function returned values 1000x too small with millisecond-resolution timestamps. #112360 (Valery Petrov).
• PromQL: reject invalid cross-delimiter escapes in strings and invalid Unicode surrogate code points. #113587 (Minh Vu). #113203 (Minh Vu).
• PromQL: preserve finite values when using modulo with an infinite divisor. #113746 (Minh Vu).
• PromQL: date/time functions hour, minute, month, year, day_of_week, day_of_month, day_of_year, and days_in_month can now be called without arguments, defaulting to vector(time()) as in Prometheus. #111869 (Valery Petrov).
• PromQL now supports the Prometheus HTTP API lookback_delta parameter for instant and range queries. #113971 (Minh Vu).
• Distributed query plans (make_distributed_plan) now support IN (subquery) without rewriting it to a JOIN: the set is built once on the initiator and its values are shipped with the worker tasks. The forced rewrite_in_to_join override is removed; the rewrite remains available as an explicit setting. #113826 (Alexander Gololobov).
• Streaming queries support watermarks. #106169 (Mikhail Artemenko).
• Added a read-through filesystem cache to the experimental ReaderExecutor read path (use_reader_executor, disabled by default). #110029 (Sema Checherinda).
• PromQL: implemented functions clamp, clamp_min, clamp_max, and round. #111870 (Valery Petrov).
• Correctly collect insert stats and queries through the Prometheus remote-write endpoint. #112825 (James Cunningham).
• PromQL aggregation operators now accept mixed letter casing, matching Prometheus. #114076 (Minh Vu).
• TimeSeries: store all tags in the tags column. #114300 (Vitaly Baranov).
• Apply the read-in-order optimization to ORDER BY queries when parallel_replicas_plan_based is enabled: the sort is now shipped to the replicas and merged on the initiator, so each replica reads in the sorting key order. Also fixes Unknown function __topKFilter for ORDER BY on a non-primary-key column with this setting enabled. #114315 (Igor Nikonov).
• Reject literal line feeds in ordinary quoted PromQL strings. #114381 (Minh Vu).
• Fix duplicated GROUP BY keys in the partial aggregation strategy of make_distributed_plan, and allow this strategy for aggregations over a distributed read. #114523 (Alexander Gololobov).
• Throw exception on violating the allowlist of hypothetical index types for WHATIF. #114646 (Yarik Briukhovetskyi).
• Restores quantized sub-column read for vector search (introduced in 26.7). #114763 (Shankar Iyer).
• The experimental TimeSeries table engine no longer applies the Gorilla codec to auto-created value columns of the samples inner table; they now default to CODEC(ZSTD(1)). Explicitly declared columns keep whatever codecs the user wrote. The timestamp column default is unchanged (CODEC(DoubleDelta, ZSTD(1))). #114790 (Nikita Mikhaylov).
• A distributed plan query (make_distributed_plan) now stops its upstream stages when a LIMIT is satisfied, instead of computing the full result and discarding it. #114806 (Alexander Gololobov).
• Added a userspace page cache tier to the experimental ReaderExecutor read path (use_reader_executor, off by default). On a miss the executor now populates the userspace page cache and serves subsequent reads from it, zero-copy. #114890 (Sema Checherinda).
• Support ROLLUP, CUBE, GROUPING SETS and the grouping function in queries executed with make_distributed_plan = 1. #114950 (Alexander Gololobov).
• Fix the error Illegal type Decimal(18, 0) of argument of function toIntervalNanosecond when a PromQL query uses offset inside a range selector, e.g. rate(m[2m] offset 5m). #114997 (Nikita Mikhaylov).
• A very large distributed_plan_workers_num no longer makes the server abort with a std::length_error when the experimental make_distributed_plan and enable_cascades_optimizer are enabled. Such a node count is now rejected with INVALID_SETTING_VALUE. #115398 (Groene AI).
• Support GROUPING SETS aggregation in distributed plans built by the Cascades optimizer (make_distributed_plan = 1, enable_cascades_optimizer = 1). #115426 (Alexander Gololobov).
• TimeSeries tables can now keep a “recent samples” table: when the new engine setting recent_samples_ttl_seconds is non-zero, the engine creates a TTL’d copy of the samples table (partitioned by recent_samples_partition_by, by day by default) written on every insert, and PromQL queries whose time range fits in the TTL window automatically read from it, making alert/recording rules and short-range dashboards several times faster on large tables. The preference can be disabled with the new query-level setting time_series_prefer_recent_samples_table. #115441 (Nikita Mikhaylov).
• Implement SELECT query from TimeSeries. #115622 (Vitaly Baranov).
• Implement the /api/v1/series endpoint of the Prometheus HTTP API for TimeSeries tables: match[] series selectors (repeated values are a union), an optional start/end time range, and an optional limit parameter. #115676 (Nikita Mikhaylov).
• Distributed query plans (make_distributed_plan) now stop idle upstream stages promptly after a satisfied LIMIT: an idle exchange sink reads NoMoreDataNeeded as soon as it arrives, and an idle exchange source notices its closed output and forwards the stop upstream. Previously such stages kept reading and computing data that nobody consumes. #115690 (Alexander Gololobov).
• Automatic parallel replicas can now collect runtime statistics and enable parallel replicas for supported queries when parallel_replicas_plan_based is enabled. #115788 (Igor Nikonov).
• Register an external recent samples table of a TimeSeries table as a referential dependency. #115944 (Groene AI).

​

Performance Improvement

• New adaptive algorithm for parallel GROUP BY (controlled via setting enable_adaptive_aggregator, enabled by default): each thread aggregates into its own hash table until it holds adaptive_aggregator_freeze_threshold keys and then freezes it, so frequent keys keep updating the small cache-resident tables with no coordination, while rare keys are routed by their hash into per-bucket backlogs and aggregated exactly once, inside the bucket-parallel merge. #111459 (Nihal Z. Miaji).
• Reduced memory usage of MergeTree tables with many data parts by sharing the schema-derived per-part metadata (column list, column descriptions, serializations, and columns substreams) across parts of a table instead of copying it into every part. The effect is largest for wide tables with many small parts. #109816 (Raúl Marín).
• Optimize GROUP BY ... ORDER BY ... LIMIT and GROUP BY ... LIMIT queries by maintaining a bounded heap during aggregation to prune groups that cannot appear in the result, significantly reducing memory usage and execution time for high-cardinality grouping. Controlled by the enable_group_by_top_k_optimization setting. Based on the initial implementation by Dmitriy Terenichev (#78553). #96630 (Konstantin Bogdanov).
• Reduce hash-table cell size for single-String aggregation keys, improving string-heavy GROUP BY workloads. The new enable_packed_string_keys_in_aggregation setting is enabled by default; disabling it, or setting compatibility below 26.8, restores the legacy StringHashTable method, which can still be faster for very low-cardinality aggregation with keys longer than 11 bytes. #110573 (Harikrishnan Prabakaran). #112089 (Harikrishnan Prabakaran).
• Parallelize the final merge of single-level aggregation hash tables with the new enable_parallel_single_level_merge setting, enabled by default. This speeds up GROUP BY queries whose per-thread hash tables stay below the two-level threshold but whose merge dominates the runtime. #110395 (Nihal Z. Miaji).
• Improve performance of DISTINCT when the input stream is sorted by a prefix of the distinct columns, for example SELECT DISTINCT a, b FROM table ORDER BY a. Such queries run up to 2.4 times faster. #110170 (Nihal Z. Miaji).
• Speed up and reduce the memory usage of DISTINCT queries on partitioned MergeTree tables by keeping each partition’s rows within a single stream, so each stream’s preliminary DISTINCT works on a disjoint set of keys, instead of the same set of keys appearing in multiple preliminary DISTINCTs. This applies when the partition expression is a deterministic function of the DISTINCT columns, and is gated by a cost heuristic on partition count and skew (which can be toggled by setting force_distinct_partitions_independently, disabled by default). The same per-partition disjointness is propagated through Expression, Filter and ARRAY JOIN steps to the final DISTINCT, LIMIT BY and GROUP BY consumers so they can skip their merge too. Controlled by the new settings allow_distinct_partitions_independently (enabled by default) and max_number_of_partitions_for_independent_distinct to tune the heuristic. #108326 (Nihal Z. Miaji).
• Optimized functions uniq, uniqExact, and uniqHLL12 in aggregation without key. #110150 (Anton Popov).
• The Parquet V3 reader can now skip whole row groups using the dictionary page, in addition to min/max statistics and bloom filters, for equality and IN conditions when a column chunk is fully dictionary-encoded. Controlled by the new input_format_parquet_dictionary_filter_push_down setting. #106952 (Alexey Milovidov). #110567 (Alexey Milovidov).
• Lazy materialization for reading Parquet files from object storage, including Iceberg tables: for ORDER BY ... LIMIT n queries, the columns that are not needed for sorting and filtering are read only for the n rows that survive the LIMIT. On a 200 MB Parquet file on S3, SELECT k, s ... ORDER BY k DESC LIMIT 10 reads 3.3 MB instead of 171 MB (51x less I/O, 8x faster). Controlled by the new setting query_plan_optimize_lazy_materialization_for_object_storage (enabled by default, and also requires query_plan_optimize_lazy_materialization). #110970 (Alexey Milovidov). * Lazy materialization for ORDER BY ... LIMIT n queries (#110970) now also applies to local Parquet files read with the file table function and the File table engine: the columns that are not needed for sorting and filtering are read only for the n rows that survive the LIMIT. The second read of a surviving file fails close with the new FILE_CHANGED_DURING_READ error if the file was modified between the two passes. Controlled by the new setting query_plan_optimize_lazy_materialization_for_file (enabled by default). Also fixes lazy materialization for object storage failing with Not found column or subcolumn ... in block when a requested subcolumn (e.g. of a JSON column) is deferred. #114262 (Alexey Milovidov).
• Column statistics are now materialized on INSERT by default when the table’s current active size plus the written block size is at most the new materialize_statistics_on_insert_max_table_size setting (default 25 GiB); the check is per written block, so a first bulk load into an empty table may still materialize statistics for each written block. This gives the cost-based join optimizer accurate estimates for freshly-loaded dimension tables and avoids pathological join orders (for example, TPC-H Q5 and Q8 no longer time out at scale factor 40), while large established fact tables keep materializing statistics during merges. #109454 (Alexey Milovidov).
• Speed up window functions on partitioned MergeTree tables by reading each partition through its own stream and evaluating the windows per partition, skipping the hash scatter that ordinarily reshuffles every row across threads before the window sort. This applies when the table partition expression is a deterministic function of the window PARTITION BY columns and is gated by the same cost heuristic as per-partition GROUP BY / DISTINCT (enough partitions relative to max_threads, no dominant partition). Controlled by the new settings allow_window_partitions_independently (enabled by default), force_window_partitions_independently, and max_number_of_partitions_for_independent_window. #114783 (Nihal Z. Miaji).
• Reduce memory usage when writing packed parts by streaming the final archive directly into the destination file instead of holding a second full copy in memory during part commit. #111995 (Alexey Milovidov).
• Added a new join_algorithm value parallel_full_sorting_merge: for hash-compatible equality joins, it shards a full sorting merge join by the hash of the join keys into independent per-shard merge joins running on all threads. It keeps the low, streaming memory usage of a merge join while parallelizing it (in a benchmark, ~2.4x faster and ~3.3x less memory than parallel_hash). ASOF joins fall back to a single full_sorting_merge; hash-incompatible key types (floating-point, JSON, Object, Dynamic) skip only the hash-scatter rewrite and can still be sharded at the source by primary-key ranges when query_plan_join_shard_by_pk_ranges is enabled. The result is not ordered. #109005 (Alexey Milovidov).
• Add GeoParquet spatial pruning at row-group and page levels to the Parquet reader. Row-group pruning skips entire row groups whose bounding box does not overlap the query geometry. Page-level pruning uses the covering.bbox column index to skip irrelevant pages within row groups, and spatial predicate pushdown now also applies during Parquet row reading. #104435 (Vasily Chekalkin).
• Parquet filter push-down (row group pruning and column index) supports more type conversions: DateTime64(x) <-> DateTime64(y), Decimal(x) <-> Decimal(y), Float32 <-> Float64. #110457 (Michael Kolupaev).
• Reduce duplicated read requests on remote disks when mark ranges are fragmented, for example by primary or skip indexes. #113584 (Anton Popov).
• Speed up execution of IN over clustered (e.g. sorted by primary key) columns by reusing the previous row’s result for equal consecutive rows. #114853 (Nikita Mikhaylov).
• An aggregation without GROUP BY keys no longer fans its single-row output out to max_threads streams, so queries whose plan continues after such an aggregation get a much smaller pipeline (fewer processors to create, schedule and profile) without losing any parallelism. #115170 (Nikita Mikhaylov).
• Keeper now reads changelog entries with less lock contention: catch-up reads plan under a shared lock and do disk I/O lock-free, and both follower catch-up and the commit path can pre-decode entries in a background thread pool, reducing redundant disk reads and CPU decoding on the leader. #108473 (Antonio Andelic).
• Speed up Keeper startup by reading multiple changelog files concurrently instead of serially, controlled by new settings log_startup_read_max_streams and log_startup_read_buffer_size. #109881 (Antonio Andelic).
• Add a new bucketed schema type for system.metric_log, which stores all metrics in a single Map(Enum16(...), Int64) column using the bucketed Map serialization with 128 buckets, plus a per-metric ALIAS column for compatibility: the table consists of a few columns instead of thousands, zero values are not stored, and reading a single metric reads only one of the 128 buckets. In addition, a Map with Enum keys can now be indexed by the string name of the enum value, e.g. map['name']. #115380 (Alexey Milovidov).
• Added a query plan optimization that pushes volume-reducing functions (length, lengthUTF8, empty, notEmpty) below the Sorting and Filter steps, replacing the wide String / FixedString argument with the fixed-size result, so it is neither buffered by a sort nor copied by a filter. Controlled by the new setting query_plan_push_down_volume_reducing_functions, enabled by default. #106199 (Peng).
• Enable read_in_order_use_virtual_row by default. When reading in order of the primary key (e.g. ORDER BY primary_key LIMIT n) over a table with many parts, only the parts that can actually contribute to the result are read (plus a read-ahead window of at most max_threads parts that keeps reading parallel), which significantly reduces peak memory consumption. #106215 (Alexey Milovidov).
• Aggregation queries without aggregate functions now use HashSet-based methods instead of HashMap (for supported key types). Speedups up to 1.8x times were observed. #108862 (Nikita Taranov).
• Avoid JOIN runtime-filter overhead when the probe side is estimated to be small. The threshold is controlled by the new join_runtime_filter_min_probe_rows setting (default 1000). #104860 (Vladimir Cherkasov).
• Queries with predicates like nullIf(key, sentinel) = const can now use primary-key, partition, and skip-index pruning. #107308 (Aditya Kumar).
• Added the jemalloc_merge_tree_arenas server setting to control dedicated jemalloc arena(s) for MergeTree part and table metadata. On many-core servers, a per-CPU pool can reduce allocator lock contention between concurrent merges, and the dedicated arena now only holds long-lived metadata. #109490 (Raúl Marín).
• Sped up ZSTD decompression on x86_64 for columns with small match offsets, such as fixed-width integer columns, by vectorizing short-offset overlapping copies with SSSE3 pshufb. For example, decompression of a UInt64 column is up to ~1.2x faster on AMD Zen 5. #110909 (Konstantin Bogdanov).
• Speed up groupUniqArray for numeric types by up to 3x by using CRC32 hashing, as in uniqExact, and by inlining per-row insertion. #110917 (Manuel).
• Faster estimateCompressionRatio with the NONE codec. #111107 (Raufs Dunamalijevs).
• Lowered the default max_snapshot_commit_thread_pool_size to 16, matching backup_threads, to reduce CFS throttling when committing snapshots for tables with many parts. #111417 (Han Fei).
• Reduce CPU usage when committing snapshots for tables with huge numbers of files. #112297 (Han Fei).
• Skip async-insert deduplication hash prewarming when deduplication is disabled. #111549 (Sema Checherinda).
• A JOIN whose ON condition is always false, for example ON 1 = 2, ON NULL, or a.t = 'A' AND a.t = 'B', no longer reads the non-contributing side. Controlled by the query_plan_short_circuit_constant_false_join setting (enabled by default). #110234 (Groene AI).
• Speed up reading of Iceberg tables by prefetching the next manifest file while the current manifest is being parsed, overlapping I/O with CPU work. #108543 (Asya Shneerson).
• Queries with FINAL that read in primary-key order with a small LIMIT no longer use split_intersecting_parts_ranges_into_layers_final. #110431 (Nikolai Kochetov).
• Improve performance of RIGHT and FULL JOIN with the parallel_hash algorithm when the right side contains NULL join keys or rows filtered by the ON expression, by avoiding null-map rebuilding. #111258 (Hechem Selmi).
• Speed up countSubstringsCaseInsensitiveUTF8 by fixing quadratic complexity that could make the function take minutes on a haystack of a few megabytes with many matches. #112003 (Groene AI).
• The Prometheus query API (/api/v1/query, /api/v1/query_range) of the experimental TimeSeries engine now executes queries in parallel instead of on a single thread. #112108 (Nikita Mikhaylov).
• Auto-created timestamp and value columns of the TimeSeries samples inner table now default to CODEC(DoubleDelta, ZSTD(1)) and CODEC(Gorilla, ZSTD(1)) respectively. #112110 (Nikita Mikhaylov).
• Fix a performance regression in system.query_log ingestion. Building the Settings and asynchronous_read_counters map columns was about six times slower than before, which could make SYSTEM FLUSH LOGS query_log exceed its 180-second timeout on servers that log many queries with many changed settings. Closes #112191. #112210 (Groene AI).
• Improve performance of RIGHT and FULL hash JOIN with multiple OR disjuncts in the ON section, and of RIGHT/FULL JOIN with residual (inequality) conditions. #111590 (Hechem Selmi).
• Reduce memory usage of UNLOCK SNAPSHOT by not loading per-data-part metadata that the unlock path does not use, avoiding out-of-memory for snapshots with a very large number of parts. #111849 (Julia Kartseva).
• Restore memory pre-allocation for dynamic paths of a JSON column during asynchronous INSERT flushes, avoiding repeated reallocations while batched rows are appended. #112222 (Groene AI).
• Re-enable the consecutive-keys optimization for TTL ... GROUP BY aggregation during merges. It had been accidentally disabled on that path, so results stayed correct but merges did more work than necessary. #112314 (Groene AI).
• Speed up reading Nested columns from binary-encoded types such as Native and RowBinary when input_format_binary_decode_types_in_binary_format is enabled, and report malformed nested Tuple type names as a syntax error instead of spending exponential time parsing them. #112560 (Alexey Milovidov).
• Improved performance of squashing and merging Dynamic and Object columns with many dynamic variants or JSON paths. #110583 (Christoph Viebig).
• Speed up min and max on 128-bit and 256-bit types (Int128, UInt128, Int256, UInt256, Decimal128, Decimal256). Up to 1.8x faster for wide integers and up to 3.4x faster for wide decimals. #111965 (Manuel).
• Enabled by default (and promoted to beta) two on-disk features: 1) packed_skip_index_max_bytes = '1M': skip-index substreams up to 1 MiB are now written into a single skp_idx.packed archive per part, reducing object count and write requests on object storage. Compatibility: only newly written parts are affected and they stay fully readable on older versions; versions before 26.6 simply do not use the packed indices for pruning (they are treated as not materialized). Set packed_skip_index_max_bytes = 0 to keep the standalone per-file layout. 2) compute_exact_num_defaults_for_sparse_columns plus optimize_trivial_count_with_sparsity_filter: exact per-column num_defaults counters are persisted and used to answer SELECT count() FROM t WHERE <pred> without a data scan when the predicate splits rows into defaults vs non-defaults. Compatibility: the exact_num_defaults flag in serialization.json is ignored by older versions, so parts remain fully readable after a downgrade; the count rewrite only applies to parts that carry exact counters, so results stay correct on tables with mixed parts. #111816 (Raúl Marín).
• Add optional caching for tokens missing from text indexes. #112742 (Rory Shanks).
• Fixed quadratic reallocation when replicating an Array whose elements hold JSON values in shared data, for example a constant Array(JSON) passed to an aggregate function. Such queries copied a number of bytes that grew quadratically with the row count. #112897 (Groene AI).
• Use the primary key index for pointInPolygon when the point argument is a whole key column of type Point (or another Tuple of two numeric elements), e.g. pointInPolygon(coord, [...]) for a table ordered by coord. Previously only the pointInPolygon((x, y), [...]) form with two scalar key columns was analyzed. Closes #54805. #112956 (Alexey Milovidov).
• Skip the data-reading step entirely for TTLDrop merges to reduce memory usage, especially for read and prefetch buffers. Closes #105639. #105859 (Pavel Kruglov).
• Reduce memory allocations when reading backup metadata by moving file entries into in-memory maps instead of copying them. #111718 (Julia Kartseva).
• Use the text index when its expression contains an empty-string comparison such as arrayFilter(s -> s != '', ...), even when optimize_empty_string_comparisons = 1 rewrites it to notEmpty. Closes #111788. #112115 (Jimmy Aguilar Mena).
• Restore query condition cache pruning after a lightweight DELETE: repeated selective queries over a table that had ever been touched by a lightweight DELETE no longer fall back to reading every mark. #112947 (Nikita Fomichev).
• Higher-order functions no longer physically copy captured columns for every array element when enable_lazy_columns_replication is enabled, reducing memory usage and copying time for lambdas that capture large columns. #111581 (Diego Gomes Tomé).
• Fixed a ~1.5x performance regression introduced in 26.5 in GROUP BY over Dynamic keys and JSON path subcolumns. #112316 (Utkal Singh).
• text_index_posting_list_apply_mode now defaults to lazy, so text index queries decode posting lists on demand at packed-block granularity with a cursor instead of eagerly materializing them into bitmaps for text indexes with posting_list_codec set. #113521 (Anton Popov).
• Speed up PromQL queries over TimeSeries tables: timeSeriesIdToGroup now reuses the previous row’s group when consecutive rows carry the same series id, skipping the per-row id serialization and hash lookup. #113580 (Nikita Mikhaylov).
• PromQL aggregation operators topk, bottomk, and limitk now use a streaming, parallel execution plan with bounded memory instead of collecting all series into a single row. Queries over high-cardinality metrics that previously failed with MEMORY_LIMIT_EXCEEDED now run in bounded memory. #113656 (Nikita Mikhaylov). #114409 (Alexey Milovidov).
• Speed up division of 256-bit integers (Int256, UInt256, and Decimal256), affecting intDiv, modulo, Decimal256 arithmetic, and conversion of these types to text. #112477 (Konstantin Bogdanov).
• Improve MergeTree read performance for JSON subcolumns by reusing resolved subcolumn metadata during required conversions. #113008 (Rory Shanks).
• Speed up building text and bloom-filter-based skip indexes with the sparseGrams tokenizer. #110541 (UnamedRus).
• Snapshot creation now uses a dedicated thread pool, so a heavy concurrent BACKUP no longer starves it. #113509 (Han Fei).
• Evaluate PromQL subexpressions shared by multiple plan steps (the topk/bottomk/limitk operand, the left side of or) once instead of twice: the SQL generated from PromQL now materializes shared subqueries. This removes a duplicated scan of the samples table; topk over rate gets ~2x faster cold, and kubernetes-mixin rule queries improve by 10-29%. #113772 (Nikita Mikhaylov).
• You can now enable early short-circuit folding for builtin and and or during analysis with the enable_function_early_short_circuit setting. This can skip executing dead single-row count() scalar subqueries while preserving normal type inference and semantic validation. Resolves #83017. #83505 (fhw12345).
• Combine I/O cost with selectivity in PREWHERE condition ordering. This fixes a performance regression of PREWHERE execution in some cases introduced after #101275. Part of #110462. #110695 (Pavel Kruglov).
• Improved the performance of levenshteinDistance (and editDistance) for medium-length strings on ARM by avoiding per-row heap allocations in the blocked Myers code path. #108185 (Groene AI).
• Add the merge_use_batch_sorting_queue MergeTree setting to optionally use the batch sorting queue for ordinary MergeTree merges, reducing CPU and wall-clock time for merge workloads where sorted merging is a significant cost. #108468 (Rory Shanks).
• Added the merge_tree_min_bytes_per_read_stream setting to reduce excessive read streams and pipeline overhead for ordinary local unordered narrow-column MergeTree scans on high core-count servers. #109035 (Jiebin Sun).
• Speed up function toStartOfInterval (aliases time_bucket, date_bin) by up to 5x for SECOND, MINUTE and HOUR intervals. #109729 (Manuel).
• The native protocol now sends String columns as a separate stream of cumulative byte offsets (the same layout Array uses for its offsets) followed by the concatenated data, instead of a per-value varint length prefix, once both peers are on protocol revision 54489 or newer. A client can then read String columns with exact buffer preallocation and a single bulk copy, about 4 times faster than the per-value layout. Old clients and servers are unaffected: the layout is negotiated through the protocol revision. The same layout is available in the Native and Buffers formats through the new settings output_format_native_write_string_with_size_stream / input_format_native_read_string_with_size_stream (off by default, so the formats stay portable). #110320 (vahid-sohrabloo).
• Avoid eager JSON serialization of Iceberg manifest entries when iceberg_metadata_log_level is below the call site’s threshold. #110437 (Andy Bradshaw).
• Improve full sorting merge join performance up to 1.5-3x times on sparse scenarios. #111200 (Vladimir Cherkasov).
• Improve efficiency of minmax and basic column statistics for floating point columns. #111221 (Christoph Viebig).
• Enable join_runtime_filter_from_fixed_hash_table when the probe and build join key types differ (for example, a Nullable probe key with a non-Nullable build key). #111244 (Hechem Selmi).
• Returns COUNT() queries directly from the text index cardinality metadata. #111494 (Elmi Ahmadov).
• Sped up row-wise hashing of JSON/Object column values stored in shared data (used e.g. by uniq/uniqExact over a JSON column and by hashed GROUP BY/DISTINCT keys) by avoiding a per-value ColumnDynamic reconstruction. Hash values are unchanged. #111501 (Valery Petrov).
• Functions mapContainsKeyLike, mapContainsValueLike, mapExtractKeyLike, and mapExtractValueLike with a non-constant pattern use less memory and time when enable_lazy_columns_replication is enabled. The pattern is no longer eagerly copied per map entry at capture time; the copy is deferred to the execution of the internal lambda and is freed sooner, so one fewer expanded copy of the pattern exists at peak. #111749 (Diego Gomes Tomé).
• Speed up sorting and other comparison-based operations over JSON columns whose paths are stored in shared data. Previously each value comparison materialized a temporary Dynamic column and ran a full binary deserialization for both sides, making ORDER BY over such columns pathologically slow. #111894 (Groene AI).
• clickhouse-client no longer pings the server before every query, which removes a network round trip per query. It also no longer re-establishes the connection - silently starting a new session, and so losing temporary tables, the current database and session settings - just because the server was too slow to answer a ping. #111990 (Alexey Milovidov).
• Use the transitive conditions optimization more wisely: the additional filter created by transitivity is now used only for index analysis where the whole filter cannot be applied. See the optimize_and_compare_chain setting. #112076 (Yarik Briukhovetskyi).
• Added two settings, shrink_over_allocated_columns_min_waste_ratio and shrink_over_allocated_columns_min_waste_bytes, that make INSERTs shrink over-allocated columns (whose reserved memory exceeds used memory due to power-of-two growth of variable-length columns) to fit before materialization and part writing, reducing peak memory usage. Disabled by default.. #112161 (Pavel Kruglov).
• Lazy column replication for ARRAY JOIN is no longer lost when a query plan is serialized. With serialize_query_plan = 1, the node executing a remote plan fragment used to ignore enable_lazy_columns_replication and always materialize every replicated copy, even though the setting defaults to enabled. #112330 (Groene AI).
• Formatting Decimal values as text is much faster — up to 190 times for a Decimal256 with a large scale, and about 1.6 times for a Decimal64. On processors with AVX-512 IFMA, converting 64-bit integers to text is about 1.4 times faster. Also fixed the rounding of toDecimalString, which dropped a carry out of the fractional part: toDecimalString(toDecimal64('9.995', 3), 2) returned 9.00 and now returns 10.00. #112457 (Konstantin Bogdanov).
• A WHERE equality is now merged into the JOIN condition when its operands have different types but a common supertype, such as Int32 and Nullable(Int32). This allows a CROSS join to be converted to an INNER Join. #112630 (Hechem Selmi).
• Queries on Iceberg tables with many delete files now start faster. ClickHouse reads and decodes the delete manifest files concurrently instead of one at a time, so their storage reads overlap. The new setting iceberg_delete_manifest_decode_concurrency (default 4) controls how many are decoded at the same time. #112679 (Asya Shneerson).
• Asynchronous logging now uses a bounded lock-free queue: threads writing log messages never block on a mutex or wake up the logging threads. #112803 (Alexey Milovidov).
• randomHadamardTransform of a constant vector is now evaluated once instead of once per row. This speeds up vector search queries that rotate the query vector with randomHadamardTransform before comparing it against a QBit column. #112921 (Alexey Milovidov).
• Restores set sharing between the part tasks of one mutation filtering on IN (subquery) over a primary key column. enable_sharing_sets_for_mutations is on by default, but the set built for primary key analysis stopped participating in the shared cache, so a table with N parts materialized the same set N times. #112941 (Groene AI).
• Fixed cubic complexity of query planning for a JOIN with a Merge table over many tables with different structures. Such queries could previously spend minutes in query planning without responding to max_execution_time or KILL QUERY. #113140 (Alexey Milovidov).
• The bucket-region cache of the S3 client now also works for data lake catalogs; previously every request to them resolved the region again. #113330 (Konstantin Vedernikov).
• Fixed a performance regression where an aggregation without GROUP BY keys serialized its whole input plan subtree, including the full contents of constant-folded literals, on every execution to compute a hash-table-size cache key it can never use. A query such as SELECT quantileMerge(arrayJoin(arrayMap(x -> state, range(5000000)))) spent about 63 ms per execution in query-plan optimization and now spends 0.15 ms. #113333 (Groene AI).
• Add opt-in Parquet serialization of UInt128/UInt256/Int128/Int256 as DECIMAL to enable row group and page pruning. #113347 (Ivan Babrou).
• Speed up query planning when uniq_v2 column statistics are used. The estimated number of distinct values is now cached instead of being recomputed on every request, which query planning issues many times per query. (issue #113038). #113357 (Groene AI).
• Do not prefetch unneeded substreams when reading a whole JSON path from advanced shared data. #113380 (Pavel Kruglov).
• Filter pushdown now works for Tuple subcolumns in Parquet and ORC files. A predicate such as WHERE tup.1 = 555555 over file, s3 or url now prunes row groups and row index strides using the tuple element’s own statistics instead of reading the whole file. #113383 (Groene AI).
• Replaced the per-bucket hash map inside the timeSeries*ToGrid aggregate functions with a flat sorted array of samples: sample ingestion becomes an O(1) append for in-order inputs (the overwhelmingly common case) and the per-bucket copy-and-sort at finalization is gone. #113681 (Nikita Mikhaylov).
• Improve PromQL query performance over the TimeSeries engine: the internal selector SQL now reads the samples-table primary-key columns directly (without no-op casts) and checks the timestamp range before the series-id set, which makes cold selector-heavy queries 30-44% faster. #113768 (Nikita Mikhaylov).
• Use typed id maps in the time-series tags collector. #113839 (Vitaly Baranov).
• Skip the redundant sortedness scan when writing MergeTree parts whose sorting keys are all constant. #113899 (Perfloop Agent).
• PromQL selectors that match all series of one metric now filter the samples table of a TimeSeries table with a continuous primary-key range on id during index analysis instead of a large id IN <set> condition, when the id layout is a two-component tuple with the canonical id generator. Removes the dominant single-threaded index-analysis cost of selector-heavy PromQL queries: up to −45% cold latency on dashboard and rule query shapes, −11% cold geomean over the full suite on a 62-billion-sample table. #114131 (Nikita Mikhaylov).
• Speed up vectorized TimeSeries tag transformations by replacing hash-based remapping with direct indexing when tag-group IDs occupy a compact range. #114244 (Minh Vu).
• The Prometheus HTTP API endpoints /api/v1/query and /api/v1/query_range now evaluate a PromQL subquery shared by several plan steps once, as the prometheusQuery and prometheusQueryRange table functions already did. Previously a request made with enable_analyzer = 0, or by a user whose profile sets it, rescanned the shared subquery once per referencing step. #114261 (Groene AI).
• Bounded the cost of estimating the selectivity of col IN (...) from column statistics, which could add hundreds of milliseconds to the planning of a single query. The estimator no longer runs the subquery behind col IN (subquery) to fill a set it only needs one selectivity number from — an unbuilt set is skipped instead. For a set larger than the new statistics_max_set_size_for_exact_selectivity_estimation setting (default 10000), the selectivity is now derived from the size of the set and its bounding range. #114389 (Nikita Taranov).
• Parsing an HTTP request that carries many query parameters or headers with the same name (for example thousands of repeated role= parameters) no longer takes quadratic time. #114410 (James).
• Fixed a regression in 26.7: for a MergeTree table ordered by toUnixTimestamp (or another integer conversion) of a DateTime64 column, a plain range filter on that column no longer used the primary key and read all granules of the matched parts. Additionally, a filter like toInt64(ts) >= c over a table ordered by the raw DateTime64 column now uses the primary key. #114413 (Alexey Milovidov).
• The heavy consistency check in ColumnArray is now performed only in debug builds. #114469 (Anton Popov).
• Do not merge-sort a distributed gather whose sort description is all-constant. #114626 (Groene AI).
• Speed up set building for IN (subquery) on partitioned MergeTree tables by keeping each partition’s rows within a single stream and deduplicating each stream independently, so the single set-filling transform — previously hashing every row serially — only sees unique rows. This applies when the partition expression is a deterministic function of the subquery’s output columns. The optimization is not applied when the largest partition holds more than twice the rows of the average partition; the new setting force_creating_set_partitions_independently (disabled by default) bypasses this check. Controlled by the new setting allow_creating_set_partitions_independently (enabled by default). #114645 (Nihal Z. Miaji).
• Parallelize the single-level to two-level conversion of per-thread aggregation states before merging. This fixes a 2-3x regression of short GROUP BY queries with heavy aggregate states (e.g. COUNT(DISTINCT ...), ClickBench Q10/Q11) on machines with very many cores, introduced in 26.7 development builds when the aggregation memory accounting was made accurate. #114691 (Alexey Milovidov).
• Speed up the timeSeries*ToGrid aggregate functions: per-sample bucket math now avoids Int128 division, and consecutive samples in the same bucket skip the hash-map lookup. #114889 (Nikita Mikhaylov).
• Appending to a system log queue no longer deep-copies the whole queue when it grows. #115030 (Groene AI).
• Process samples of the timeSeries*ToGrid aggregate functions in runs of consecutive samples falling into one grid bucket, speeding up the sample-ingestion stage of timeSeriesRateToGrid and related functions by 1.2-2x on sorted time series data. #115041 (Nikita Mikhaylov).
• Use compact presence masks for PromQL set operators. #115224 (Minh Vu).
• Faster evaluation of PromQL instant queries: the timeSeries*ToGrid aggregate functions use the vectorized sample-classification path for single-point grids, and timeSeriesIdToGroup fills its result column directly. #115267 (Nikita Mikhaylov).
• A JOIN whose ON section has several OR-ed equalities no longer materializes the right-hand key columns that the query does not select. #115465 (Nikita Taranov).
• Naming a DateTime or DateTime64 type no longer builds a UTC time zone lookup table that nothing reads, which makes constructing these types — and therefore starting up clickhouse local and clickhouse client — faster. #115488 (Konstantin Bogdanov).
• Attaching the system tables no longer copies each table’s metadata, including its full column description, twice just to set the table comment. #115490 (Konstantin Bogdanov).
• The client hostname is now resolved when it is first reported instead of on every clickhouse client and clickhouse local invocation. #115491 (Konstantin Bogdanov).
• Fixed quadratic time when the functions h3kRing, h3ToChildren, h3PolygonToCells and h3PolygonToCellsWithContainment build their result over many rows. A query over 196 000 rows spent 838 seconds copying memory in a CI stress run; it now takes about 5 seconds. #115773 (Groene AI).
• Speed up duplicate-series checks for all-zero PromQL condition vectors. #115822 (Minh Vu).
• Speeds up GROUP BY without aggregate functions over String, FixedString and LowCardinality keys: the query no longer rewrites an unused placeholder into the hash table for every row, only for the rows that add a new key. #115834 (Nikita Taranov).
• Speed up toUTCTimestamp/to_utc_timestamp and fromUTCTimestamp/from_utc_timestamp on DateTime64 arguments, restoring the throughput they had before the overflow fix in #109738. Results are unchanged. Closes #115802. #115838 (Groene AI).
• Reduced memory usage and query-plan optimization time for keyed aggregations over large constant-folded literals. The aggregation hash-table-statistics cache key is now hashed as it is serialized instead of being accumulated in memory first, so a plan carrying a large literal no longer holds a copy of it during optimization. For an 80 MB literal, peak memory drops by about 250 MB and optimization time by about 2.5x. The computed key is unchanged. #115847 (Groene AI).

​

Improvement

• Outgoing OpenAI-format HTTP requests are now tagged with the SQL name of the calling AI function for upstream endpoint identification. #110449 (George Larionov).
• Reduce cancellation latency for positive forms of LIMIT BY: KILL QUERY and Ctrl+C now interrupt within a single chunk instead of waiting for the whole step to finish. #106070 (Roman Vasin).
• When an INSERT fails because a materialized view’s TO target table rejects the write while the insert pipeline is being built, the error message now names the materialized view and its target table instead of only reporting the bare storage error. #107234 (Groene AI).
• With input_format_protobuf_oneof_presence, ClickHouse no longer requires the oneof presence enum to list exactly the tags declared in the Protobuf schema: extra tags without matching columns in the target table are now allowed, which makes schema changes easier. #109174 (Ilya Golshtein).
• The transposed distance functions over QBit (L2DistanceTransposed, cosineDistanceTransposed, dotProductTransposed, and their quantized variants) now accept a reference vector longer than the number of searched dimensions, ignoring the extra trailing elements. This allows reusing a full-size query vector for a reduced-dimension Matryoshka search without slicing it first. #109388 (Alexey Milovidov).
• The basic statistics NullCount sub-statistic is renamed to DefaultCount and now counts rows equal to the column type’s default value, not only NULL rows of Nullable columns. #109977 (Han Fei).
• Support concurrent writes to Iceberg tables on a local disk: LocalObjectStorage now implements conditional writes, so the compare-and-swap that publishes a new snapshot through version-hint.text works there and concurrent writers cannot lose each other’s updates. #112556 (Alexey Milovidov).
• ALTER TABLE ... MOVE/REPLACE PARTITION between two MergeTree tables with incompatible granularity now throws BAD_ARGUMENTS instead of LOGICAL_ERROR. #110535 (Groene AI).
• The Values format no longer records false parse errors in system.errors and system.error_log when streaming parsing successfully falls back to SQL expression parsing. #111141 (Pablo Marcos).
• Added the DuplicationDataHashComputations ProfileEvent, which counts column-wise data-hash computations performed while deduplicating INSERT blocks into *MergeTree tables. #111173 (Valery Petrov).
• Highlight the ./clickhouse and sudo ./clickhouse install commands in the universal installation script output. #111514 (Alexey Milovidov).
• Cartesian joins and analyzer-planned constant-predicate joins no longer fail because of an incompatible join_algorithm setting when the analyzer is used. #108289 (János Benjamin Antal).
• bitmaskToArray and bitmaskToList now support Int128, UInt128, Int256, and UInt256 arguments. bitPositionsToArray is also faster for big-integer arguments because it now scales with the number of set bits rather than the highest set-bit position. #110743 (Manuel).
• Fix a crash in JIT-compiled expressions on native from-source macOS/aarch64 builds by using the correct LLVM host triple for the host OS. #111591 (Raúl Marín).
• Fix JSON_EXISTS, JSON_VALUE, and JSON_QUERY on Dynamic input when tuple path arguments contain LowCardinality elements, avoiding LOGICAL_ERROR exceptions. Closes #110345. #109944 (Groene AI).
• Corrupt ORC files whose type tree declares more columns than the stripe footer now fail schema inference with CANNOT_EXTRACT_TABLE_STRUCTURE instead of aborting in debug/sanitizer builds. #110967 (Groene AI).
• Rotated system log tables, such as system.metric_log, can now be marked table_readonly even when their metadata exceeds max_query_size, so rotation no longer logs QUERY_IS_TOO_LARGE and continues normally. #111703 (Alexey Milovidov).
• In Play, added a button to close all tabs except the current one. #111835 (Mikhail Artemenko). Fixed keyboard focus after activating it with Enter or Space when the button hides itself. #111899 (Alexey Milovidov).
• In Play, holding Ctrl (Cmd on Mac) while hovering over a table cell whose value is an image URL now shows an image preview. #109347 (Alexey Milovidov).
• On Linux, ClickHouse now adds a system.warnings entry at startup when rseq is unavailable, because per-CPU profile counters then fall back to a slower sched_getcpu path. #109283 (Azat Khuzhin).
• COUNT(*) over the file table function no longer throws an exception when schema inference wraps columns in Nullable. Closes #102044. #102509 (Rory Shanks).
• Functions L1Normalize, L2Normalize, LinfNormalize, and LpNormalize (and their aliases) now work for Array arguments, not only Tuple arguments, consistently with L2Norm and the distance functions. #110052 (Alexey Milovidov).
• Removed the legacy Apache Arrow-based ORC reader. The ORC input format now always uses the native ClickHouse decoder, which was already the default. The input_format_orc_use_fast_decoder setting is obsolete and has no effect. #110074 (Alexey Milovidov).
• Made commonly used clickhouse-client arguments configurable from XML (<hints/>, <highlight/>, <echo/>, <echo_query_id/>, <echo_formatted/>, <print-profile-events/>), and preserved client-config values when the corresponding defaulted CLI flags are omitted, for settings such as history_max_entries, suggestion_limit, progress, progress-table, enable_progress_table_toggle, print-memory-to-stderr, chime-threshold-seconds, and profile-events-delay-ms. #110607 (Larry Snizek).
• Added the analyzer_compatibility_apply_final_to_all_joined_tables setting to restore the old behavior where FINAL on the left-most table of a JOIN also applies to the other joined tables. The setting is registered in the settings changes history, so compatibility with versions before 26.6 restores the old semantics automatically. #111589 (Nikita Fomichev).
• In Play, the + (new tab) button now renders at the correct height in Firefox. #109524 (Alexey Milovidov).
• A nested correlated EXISTS subquery that references a column from a scope beyond its immediate outer query now fails with NOT_IMPLEMENTED instead of the internal NOT_FOUND_COLUMN_IN_BLOCK exception. #110310 (Groene AI).
• Formatting an EXECUTE AS query after AST rewrites no longer throws a client-side exception, for example in the AST fuzzer. #111725 (Groene AI).
• Improve DESCRIBE query support for parameterized views, and support parameterized views in scalar expressions in the new analyzer. Closes #66307. Closes #69598. #68978 (Dmitry Novik).
• Malformed CREATE INDEX and CREATE HYPOTHETICAL INDEX expressions that cannot survive a format-parse-format round trip now report an exception instead of hitting an Inconsistent AST formatting logical error in debug and sanitizer builds. #109175 (Groene AI).
• Web UI: fix the user/password field staying red even when the credentials are correct and the server is reachable, and re-check credentials after editing the connection so a correct value turns green. #109414 (Alexey Milovidov).
• Added the text_index_max_processed_tokens_before_flush and text_index_max_memory_usage_before_flush settings to control when text index builders flush temporary segments. #111573 (Rory Shanks).
• Added the system_cache_extensions setting to control which file extensions are stored in the system cache when use_split_cache is enabled. #111575 (Kirill).
• Error messages on musl-based builds now show the human-readable strerror text instead of an empty or numeric value such as strerror: 0. #111620 (Konstantin Bogdanov).
• The effective named collections storage type is now available as named_collections_storage.type in system.server_settings and via getServerSetting('named_collections_storage_type'). #111806 (Pablo Marcos).
• Relaxes an overly strict permutation size check in ColumnReplicated::permute so that it follows the general IColumn::permute contract, which allows a permutation shorter than the column when a limit is set. #109897 (Groene AI).
• Improve planning of streaming queries by unifying the internal streaming read plan, which avoids double-optimizing direct text-index reads. #111821 (Mikhail Artemenko).
• Web UI: the gap between the Run one and Run all buttons now matches the left margin before the first button. #111988 (Alexey Milovidov).
• Memory reservation scheduling (experimental) no longer evicts an allocation when a pending memory release would free the room needed by an over-limit reservation increase: the eviction decision now waits for in-flight decreases to be applied, avoiding unnecessary query kills under memory pressure. #112299 (Sergei Trifonov).
• Iceberg v2 requires sequence_number in manifest files, but other engines such as BigQuery can write manifest files without it. ClickHouse can now read such tables. #112431 (Konstantin Vedernikov).
• Show warning when statistics for join reordering are missing. #107666 (Vladimir Cherkasov).
• Added MySQL-compatibility columns to INFORMATION_SCHEMA views: SCHEMATA.DEFAULT_COLLATION_NAME and DEFAULT_ENCRYPTION; TABLES.ENGINE and related MySQL columns; and COLUMNS.COLUMN_KEY, PRIVILEGES, GENERATION_EXPRESSION, SRS_ID. This lets MySQL-aware clients run their catalog introspection queries without hitting UNKNOWN_IDENTIFIER. #109351 (Alexey Milovidov).
• An unqualified reference to an INNER JOIN key that is equated in the ON condition (e.g. SELECT id FROM a INNER JOIN b ON a.id = b.id) is no longer reported as an ambiguous identifier, since both sides are guaranteed to carry the same value. #109366 (Alexey Milovidov).
• Return HTTP code 403 Forbidden (instead of 500 Internal Server Error) for ACCESS_DENIED exceptions over the HTTP interface. #111043 (Schum).
• Write the UNIQUE KEY index SST directly into part storage instead of staging it in a local temporary file, removing the dependency on a local temporary volume and avoiding an extra copy. #111189 (johnjing).
• UTMToGeo now accepts the MGRS latitude band letter returned by geoToUTM as its fourth argument (in addition to the integer hemisphere flag), so a geoToUTM result round-trips through UTMToGeo directly. #111521 (Alexey Milovidov).
• Fix a LOGICAL_ERROR (“Part level Min-Max index was constructed from unexpected columns set”) that could abort a background merge in debug/sanitizer builds after part_minmax_index_columns was lowered (for example from with_block_number_offset back to partition_key_only). #111511 (Groene AI).
• Fixes ProfileEvents['JoinResultRowCount'] under-reporting the result size of a JOIN that spills to disk. Rows emitted from delayed buckets were counted only in JoinDelayedJoinedTransformRowCount and were missing from the total. Query results were always correct; only the profile event was wrong. #112673 (Groene AI).
• SYSTEM DISABLE FAILPOINT now rejects a fail point name that does not exist, raising BAD_ARGUMENTS like SYSTEM ENABLE FAILPOINT already did. Previously a mistyped name reported success while leaving the intended fail point enabled, with no indication that nothing had been disabled. #112680 (Groene AI).
• Added a new column skipping_indices_types to the system.tables table. It contains the distinct types of data-skipping indices defined for each table. #106388 (Anton Popov).
• Added a sanity check that a distributed query always carries a known client version, throwing a logical error instead of silently forwarding a zero version to remote shards. Server-initiated queries (background flushes, streaming consumers, dictionary reloads, asynchronous insert flushes) now report the server’s own version as the initiator version. #109408 (Alexey Milovidov).
• Reduced the verbosity of the per-part primary-key binary-search diagnostics in MergeTree by lowering them from trace logging to test logging. #110324 (Alexey Milovidov).
• The hive partitioning sample path for object storage tables (e.g. S3) is resolved on the first use of the table instead of CREATE/ATTACH, so an unreachable endpoint no longer blocks table creation and server startup. #111842 (Nikolay Degterinsky).
• Fixed ReadBufferFromPocoSocketBase::setReceiveTimeout applying its argument as seconds rather than microseconds, and fixed builds configured with -DENABLE_LIBFIU=OFF, which did not compile. In builds without libfiu, SYSTEM ENABLE FAILPOINT and related commands now throw an exception instead of silently doing nothing. #112767 (Alexey Milovidov).
• Fix clickhouse-local under-reporting rows_read and the other progress counters in the statistics of the JSON/XML output and in the progress bar: a progress increment that arrived while the progress packet was being assembled was discarded. #112958 (Alexey Milovidov).
• Fixed the error reported when an asynchronous INSERT into a Distributed table needs a queue directory whose name exceeds the 255-byte filesystem limit (reachable with use_compact_format_in_distributed_parts_names = 0). It was a LOGICAL_ERROR for a shard with internal_replication and an unattributed Code: 1001. std::exception otherwise; both now report ARGUMENT_OUT_OF_BOUND naming the table, the cluster and the limit. #113083 (Groene AI).
• Reject out-of-range execution_time quota limits with BAD_ARGUMENTS instead of relying on undefined behavior when scaling them to internal nanoseconds. #113178 (Alexey Milovidov).
• Automatic parallel replicas now support JOIN queries. #106073 (Nikita Taranov).
• Support refresh-token authentication for the OneLake catalog. #110413 (alesapin).
• The write_snapshot_version Keeper setting is now hot-reloadable. #112150 (alesapin).
• Parallelize SYSTEM DROP FILESYSTEM CACHE. #112532 (Kseniia Sumarokova).
• Unsupported nesting of a SQL SECURITY DEFINER or SQL SECURITY NONE view over s3Cluster, urlCluster, or fileCluster is now rejected with a normal query error instead of hitting a LOGICAL_ERROR in debug and sanitizer builds. #113371 (Groene AI).
• Added the setting analyzer_compatibility_multiple_joins_qualify_column_names (default false). When enabled and the FROM clause of a query contains two or more JOINs, result column names produced by the analyzer mimic the old analyzer’s multiple-joins rewrite: columns expanded from * are named <alias-or-table>.<column>, and an unaliased column reference in the SELECT list keeps its name exactly as written. This makes outer queries that reference such qualified names, like SELECT ll.Date FROM (SELECT * FROM t AS ll JOIN t1 ON ... JOIN t2 ON ...), work as they did with the old analyzer. It also fixes the analyzer losing qualified result column names for columns expanded from * when group_by_use_nulls is combined with ROLLUP, CUBE, or GROUPING SETS, which produced duplicate result column names and broke outer references to those columns. #110746 (Dmitry Novik).
• Renamed allow_experimental_delta_kernel_rs to allow_delta_kernel_rs. The old name is kept as an alias. #113476 (Kseniia Sumarokova).
• When the remote/remoteSecure table functions or the Remote/RemoteSecure table engines are given a missing named collection together with a key = value override, ClickHouse now reports the missing named collection instead of reparsing the call positionally and reporting an unrelated error. #113510 (Groene AI).
• Shell startup with clickhouse-bootstrap completion no longer dumps environment variables when the completion is loaded from the user completion hook. #113694 (Azat Khuzhin).
• Added the sainte_lague value for shared_merge_tree_merge_coordinator_distribution_algorithm and made it the default instead of water_filling. #113442 (Mikhail Artemenko).
• Added the filesystem_cache_wait_for_concurrent_download_timeout_milliseconds setting to bound how long a read waits for another query’s filesystem-cache download of a single chunk (about 1 MiB), not a whole file segment. #113322 (Kseniia Sumarokova).
• Added the Keeper setting min_time_between_fsyncs_ms (default 5 ms), which batches changelog appends that arrive shortly after the previous flush instead of starting another flush immediately. #113749 (Michael Kolupaev).
• Kafka tables with Keeper-stored offsets (kafka_keeper_path) now throw ABORTED instead of LOGICAL_ERROR when the table becomes inactive after a Keeper session loss, for example during a direct read or materialized-view streaming. #113913 (Alexey Milovidov).
• Fixed inconsistent AST formatting of viewIfPermitted: the table function form no longer formats an operator such as not incorrectly in its ELSE branch, and the expression form viewIfPermitted(...) no longer gets a spurious ELSE. #113652 (Alexey Milovidov).
• Fixed a spurious Failed to drop temporary table after refresh. Table ... is left behind and requires manual cleanup. error logged when a refreshable materialized view refresh is cancelled (for example by server shutdown) while database_atomic_wait_for_drop_and_detach_synchronously is enabled. The temporary table was in fact dropped, so no manual cleanup was needed. #113957 (Groene AI).
• The server now throws UNKNOWN_ELEMENT_IN_CONFIG on unknown configuration options in the server config, helping catch typos and misconfigured options early. The check can be disabled with <skip_check_for_incorrect_settings>. #100332 (Alexey Milovidov).
• In the Web UI, the browser history and the URL are no longer updated on every keypress; they are recorded when a query is successfully run or on a tab change. #113596 (Alexey Milovidov).
• ClickHouse now reports a stack overflow instead of dying silently: the fatal signal handlers run on the alternative signal stack, so SIGSEGV, SIGABRT, SIGILL, SIGBUS, SIGSYS, SIGFPE, and SIGTRAP produce a stack trace even when the faulting thread has exhausted its stack. #113927 (Groene AI).
• ALTER TABLE ... MODIFY COLUMN <col> Tuple(...) on a named Tuple is now metadata-only when only adding subfields, matching the speed of top-level ADD COLUMN. Gated behind SET allow_metadata_only_named_tuple_alter = 1. #107305 (Amos Bird).
• Add input_format_json_max_object_size setting to limit JSON object size on parsing. Closes #106704. #107669 (Pavel Kruglov).
• Avoid logging sasl.kerberos.kinit.cmd configuration parameter is ignored. for Kafka engine configurations that do not use Kerberos keytab authentication. #108235 (Ivan Shelestov).
• Add system.s3(azure)_queue_metadata. #108522 (Kseniia Sumarokova).
• Log tolerated connection failures to remote MySQL/PostgreSQL databases as warnings instead of errors. #109472 (Shaohua Wang).
• When neither port nor secure is specified, clickhouse-client tries both the default port 9000 and the secure port 9440 concurrently and uses the one that answers first, so clickhouse-client --host play.clickhouse.com --user play connects over TLS without --secure, even though the plain port of that server is silently dropped rather than refused. If the port that answered turns out to be unusable — a secure port with an untrusted certificate, for example — the client falls back to the other one, because the protocol was not requested explicitly. #110130 (Alexey Milovidov).
• Clicking the cloud logo in the Playground now opens the ClickHouse website in a new tab, so you don’t lose progress while queries are running. #110422 (William Hatcher).
• Add STREAM BOUNDED modifier, which read only the first snapshot of a streaming query, then finish instead of subscribing for updates. #110653 (Smita Kulkarni).
• Add information of internal state of joins to EXPLAIN ANALYZE. #110892 (Kirill Kopnev).
• When an RWLockImpl::getLock() re-entrancy assertion fires (RWLock is already locked in exclusive mode / Cannot acquire exclusive lock while RWLock is already locked), the error message now names the other query_ids currently owning the lock and how many locks the requesting query_id already holds, to make such lock-ordering errors diagnosable. #110971 (Groene AI).
• ClickHouse Keeper now logs the “Client has not sent any data” event at Information level instead of Warning, matching the server’s TCP handler. This avoids spurious warnings from TCP health-check probes (for example Kubernetes tcpSocket liveness probes) that open and immediately close a connection to the Keeper client port. #111217 (Zeynel).
• MaterializedPostgreSQL now preserves unchanged PostgreSQL TOAST values during updates instead of replacing them with default values. #111552 (OrpheusAgent).
• keeper: Increase default write_snapshot_version to 8 and make it hot-reloadable. #111715 (Michael Kolupaev).
• Add STREAM UNORDERED modifier: skip the per-snapshot commit-order sort. #111794 (Smita Kulkarni).
• Added a new MergeTree setting text_index_version that controls the on-disk format version of text indexes: v0_initial, v1_with_codec, or v2_with_positions. During a rolling upgrade or before a downgrade, set it to an older version so that newer servers keep writing text index parts in a format older servers can still read; the compatibility setting automatically adjusts it. #111803 (Anton Popov).
• AI text functions (aiGenerate, aiClassify, aiExtract, aiTranslate) now reject truncated or otherwise incomplete provider responses (e.g. when the model hits the max_tokens limit) instead of silently returning partial output. Behavior follows the ai_function_throw_on_error setting. #111830 (George Larionov).
• Improve reduced-precision L2DistanceTransposedQuantized, cosineDistanceTransposedQuantized, and dotProductTransposedQuantized by reconstructing p < 8 QBit(Int8) codes with Gaussian conditional-mean prefix centroids. Full precision (p = 8) remains bit-exact; reduced-precision approximate distances may change, and recall/latency gains are workload-dependent. #111867 (Sergey Kuznetsov).
• Fixed the query condition cache not being populated for granules eliminated by any but the last conjunct of a WHERE clause with multiple AND conditions. #112083 (Shankar Iyer).
• Added a new setting input_format_json_max_string_column_growth_step that caps the power-of-two growth of internal String buffers when building a JSON column, reducing peak memory when inserting large JSON documents. Disabled by default. #112159 (Pavel Kruglov).
• Reduced peak memory when writing JSON columns with bucketed shared data serialization (map_with_buckets and advanced) by splitting the shared data into buckets one at a time on write instead of materializing all buckets simultaneously. This is mostly beneficial for Array(JSON) columns: for scalar JSON the shared data is split per granule (bounded by index_granularity_bytes), so the saving is small, whereas a single Array(JSON) granule can hold many nested rows and thus a large amount of shared data, where the one-bucket-at-a-time split significantly lowers peak memory. #112172 (Pavel Kruglov).
• The documentation of a setting in system.documentation — shown by the built-in /docs page and by the help command — now includes the history of the changes of its default value: the version in which the setting was introduced and every later change of its default, with the previous value, the new value and the reason for the change. #112177 (Alexey Milovidov).
• EXPLAIN ANALYZE now works for streaming queries. #112445 (Kirill Kopnev).
• The modifiers of a column declaration - COMMENT, CODEC, STATISTICS, TTL, COLLATE, PRIMARY KEY and per-column SETTINGS - can now be written in any order in CREATE TABLE, and each of them at most once. Previously only one fixed order was accepted, and, for example, x UInt64 CODEC(ZSTD) COMMENT 'text' was a syntax error. In ALTER TABLE ... ADD COLUMN / MODIFY COLUMN, the supported modifiers - COMMENT, CODEC, STATISTICS, TTL and per-column SETTINGS - can also be written in any order; per-column SETTINGS in ADD COLUMN and a declared STATISTICS in ADD COLUMN / MODIFY COLUMN are now applied instead of being silently dropped, and COLLATE and PRIMARY KEY in these ALTER commands now throw an exception instead of being silently ignored. #112788 (Alexey Milovidov).
• The pread_threadpool read method needs the preadv2 system call with the RWF_NOWAIT flag to read the data that is already in the page cache without handing the read off to a thread pool. It is now checked at start time whether that system call can be used, and if it cannot - the Linux kernel is older than 5.11, or a seccomp profile of a container runtime rejects the system call - the default value of local_filesystem_read_method is switched to pread, and the reason is reported in the server log. Previously, every read paid for a thread pool hand-off on such systems, and a seccomp profile that answers EPERM made queries fail with CANNOT_READ_FROM_FILE_DESCRIPTOR. #112945 (Alexey Milovidov).
• Report malformed Parquet thrift metadata as INCORRECT_DATA. #113311 (Groene AI).
• Fixed GCD codec handling of signed fixed-width values with negative numbers. Previously it computed the divisor from signed values directly, which could miss the common divisor and significantly reduce compression efficiency for signed integer and decimal columns. It now computes the divisor from magnitudes for signed types while preserving existing behavior for unsigned types and decompression. #113798 (Kirill Shcherbatov).
• Add settings enable_alp_codec, enable_sz3_codec, enable_zxc_codec, and enable_quantized_codec to enable each experimental compression codec individually. #113824 (Raufs Dunamalijevs).
• Allow ALTER TABLE ... MODIFY COLUMN of a column whose subcolumns are used in the primary or partition key, as long as the subcolumns used in the key keep an on-disk-compatible type. Previously any such ALTER was forbidden. #113862 (Pavel Kruglov).
• Add Filesystem cache dashboard to system.dashboards. #113884 (Kseniia Sumarokova).
• Put the per-buffer-refill cache logs behind filesystem_cache_verbose_logging. #113885 (Groene AI).
• When a column’s declared type and its data diverge during a MergeTree read (mixed type provenance, e.g. after an ALTER TABLE ... MODIFY COLUMN whose mutation has not finished), the server now reports a clear exception naming the column and both structures, instead of an unchecked cast: debug and sanitizer builds previously aborted with a bare Bad cast from type A to B naming no column, and release builds walked mismatched memory silently. #114087 (Alexey Milovidov).
• A window function with PARTITION BY now works under make_distributed_plan. When the plan shape allows, the window runs in parallel across buckets that each receive complete partitions. #114111 (Vighnesh Pathrikar).
• Web UI: the hourglass loading indicator in the databases panel is animated again. #114187 (Alexey Milovidov).
• 64-bit hash function is used now for external nullable fixed-width aggregation methods (avoids collision disaster for very high cardinality aggregation). #114210 (Nikita Taranov).
• Distributed and Buffer tables that omit their column list now always infer the structure under the creating user’s access rights. #114211 (Pedro Ferreira).
• x IN (subquery) over a LowCardinality column now returns LowCardinality(UInt8) — the same type as x IN (literal list). Previously the two forms returned different types, and distributed plans with such an IN needed a special case in the plan type check. #114229 (Alexander Gololobov).
• Enabled ie_join by default: the default value of join_algorithm is now direct,parallel_hash,hash,ie_join. A JOIN whose ON section has only inequality conditions (two comparisons <, <=, >, >= between expressions of the joined tables) is now executed with the sort-based IEJoin algorithm instead of a CROSS JOIN with a filter, and LEFT/RIGHT/FULL/SEMI/ANTI joins with such conditions are supported. Since ie_join is last in the list, it is used only when the other algorithms do not apply. #114327 (Vladimir Cherkasov).
• Fixed charts in the Web UI showing the data of a mirrored point in the hover balloon for a result ordered by x descending. Date, Date32, DateTime and DateTime64 columns can now be plotted; previously only integer Unix timestamps worked. Dates on the axes and in the balloons are rendered in ISO 8601, and quantities with the K/M/G/… suffixes, the same way the advanced dashboard does. #114349 (Alexey Milovidov).
• Support parsing integers that exceed the 64-bit range in the JSON data type, JSONExtract, and isValidJSON. Such values are read into wide integer types such as Int128/UInt256 instead of causing the whole document to be rejected. #114379 (Pavel Kruglov).
• QueryRunner tables now start worker threads on demand and release them once idle, instead of occupying threads for the table’s whole lifetime. #114522 (Miсhael Stetsyuk).
• Added observability for the stacks of fibers, which are used for asynchronous communication with remote replicas: profile events FiberStackAllocs, FiberStackAllocBytes, FiberStackAllocNanoseconds, FiberStackFreeNanoseconds, and metrics FiberStacks, FiberStackBytes. #114541 (Alexey Milovidov).
• PromQL: support @ start() and @ end() timestamp modifiers. #114558 (Minh Vu).
• Propagate the OpenTelemetry trace context into the fibers used for asynchronous connection establishment and remote query execution (hedged requests, async_socket_for_remote, async_query_sending_for_remote). Spans of remote queries in a distributed query are now correctly parented under the initiator’s Connection::sendQuery span instead of being attached directly to the client-supplied trace context, and each asynchronous task execution produces its own span. #114813 (Diego Gomes Tomé).
• With this change, while using the WHATIF indices, the projection-served baseline reports per-candidate not_applicable instead of failing the statement. #114846 (Yarik Briukhovetskyi).
• DROP DATABASE now drops tables in reverse topological order of both loading and referential dependencies (previously only loading dependencies), so a crash in the middle of DROP DATABASE cannot leave a table whose dependencies were already dropped. #114952 (Alexey Milovidov).
• In the Web UI dashboards, the global error screen is replaced by a note above the dashboards, so a failing chart no longer hides the whole page. #115019 (Mikhail Artemenko).
• Spill the adaptive aggregator’s learning-phase tables under memory pressure. #115038 (Groene AI).
• clickhouse-local now answers CORS preflight requests with the same permissive headers as clickhouse-server, so the HTTP interface opened by SYSTEM START LISTEN HTTP can be used from a browser out of the box, including the Web UI opened from a file:// URL. #115045 (Alexey Milovidov).
• Support PostgreSQL-style regular expression match operators: ~ (an alias for the match function), ~* (case-insensitive match), !~ and !~* (negations). The \d, \dt, \dv commands of psql now work when connected to ClickHouse over the PostgreSQL compatibility protocol, and a failed query no longer terminates the connection. #115066 (Alexey Milovidov).
• clickhouse-client no longer prints the Connecting to ... message twice when it asks for a password interactively. #115081 (Alexey Milovidov).
• EXPLAIN WHATIF now says why the empirical estimate was skipped, in a new empirical_reason line shown when empirical_status is unsupported. #115140 (Yarik Briukhovetskyi).
• When no join algorithm enabled by the join_algorithm setting can execute a JOIN, the error message now names the algorithms that were tried and the strictness and kind of the JOIN that failed, instead of reporting only that none of them worked. #115226 (Alexey Elkin).
• Vector search queries can now use the vector similarity index when the reference vector is an integer array literal, e.g. ORDER BY L2Distance(vec, [1, 2]). Previously, such queries silently fell back to a brute-force scan. #115255 (Robert Schulze).
• Azure client-side network failures (connection reset/refused, DNS and TLS errors) are now retried as transport errors instead of surfacing as a synthetic 500, matching the S3 client. #115269 (Arsen Muk).
• The documentation of all SQL statements (e.g. name, syntax, description, examples) can now be retrieved from system table system.statements. #115341 (Robert Schulze).
• BACKUP no longer leaves a lock file behind in its destination when the attempt that created it fails before writing anything — whether it fails while claiming the destination or while opening an archive. No later backup could match such a lock, so every retry to the same destination failed until the file was deleted by hand. Backups to S3 and Azure destinations also create the lock exclusively, and a lock that belongs to another backup is reported as a concurrent backup rather than as whatever error the storage returned. #115387 (Julia Kartseva).
• Support bearer-token authentication in BuilderRWBufferFromHTTP (attempt 2). #115400 (Sergei Trifonov).
• Add some dimensional metrics for S3Queue. #115422 (Bharat Nallan).
• Added keyword as an alias for text index tokenizer array for improved compatibilty with OpenSearch, Elasticsearch, SolR, and Lucene. #115507 (Robert Schulze).
• Allow to create own s3 tables. #115652 (Konstantin Vedernikov).
• Do not log per-thread untracked memory on non-server wide MEMORY_LIMIT_EXCEEDED errors. #115658 (Azat Khuzhin).
• The Prometheus remote-write protocol now supports asynchronous inserts: data from multiple concurrent remote-write requests is batched before forming parts, following the async_insert setting (enabled by default; add async_insert=0 to the handler URL to keep the synchronous behavior). A request is acknowledged only after the data is flushed to all inner tables of the target TimeSeries table. #115688 (Nikita Mikhaylov).
• Makes all timeSeries* aggregate functions handle duplicate timestamps by one rule (the greatest value wins, NaN loses to any other value), fixing the order-dependent results of timeSeriesInstantRateToGrid, timeSeriesInstantDeltaToGrid, timeSeriesLastTwoSamples and timeSeriesResampleToGridWithStaleness when some values at a shared timestamp are NaN. The rule is now documented and tested. #115920 (Vitaly Baranov).

​

Bug Fix (user-visible misbehavior in an official stable release)

• Fix column-name corruption and BAD_ARGUMENTS exceptions with inject_random_order_for_select_without_order_by: multi-column SELECT queries no longer fail, and output column names for CREATE TABLE AS SELECT, CREATE VIEW AS SELECT, and named formats such as JSONEachRow are preserved instead of being replaced with internal __subquery_column_<UUID> aliases. Closes #102812. Closes #101107. #105896 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK in the native Parquet V3 reader when a PREWHERE has conjuncts that share a common intermediate expression. #107059 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK for queries that use the _part_starting_offset or _part_offset virtual columns in WHERE together with lazy materialization. #108287 (Vladimir Cherkasov).
• Fix Not found column or subcolumn c.null in block when running WHERE col IS NULL or IS NOT NULL on a schema-evolved Iceberg column that was added after older data files were written, with optimize_functions_to_subcolumns = 1 (the default). #109515 (Groene AI).
• Fix false ICEBERG_SPECIFICATION_VIOLATION errors when reading an Iceberg table whose decimal or other parameterized primitive type is serialized with different whitespace across metadata files, such as decimal(20,0) in the table metadata and decimal(20, 0) in the manifest. #109676 (Groene AI).
• Fix deferred PREWHERE with optimize_move_to_prewhere = 1 when PREWHERE is applied after FINAL: WHERE conditions are no longer moved ahead of a deferred row policy, which could change query results. #109703 (Yarik Briukhovetskyi).
• Fix BAD_GET when comparing an Array(String) column with an array literal such as arr = ['x'] and the column has a text, tokenbf_v1, or ngrambf_v1 skip index. Such comparisons now fall back to a full scan instead of failing the query. #110057 (Groene AI).
• Fix wrong results for has, mapContainsKey, and mapContainsValue with an empty needle when a text index is present. #110246 (Robert Schulze).
• Fix Host is empty in S3 URI when a scalar subquery is used as the URL argument of the s3 table function, or as an argument of any other table function, in CREATE TABLE ... AS SELECT queries. #110254 (Sema Checherinda).
• Fix ON CLUSTER privilege checks for SYSTEM STOP/START CLEANUP and SYSTEM STOP/START VIRTUAL PARTS UPDATE: they no longer require SYSTEM PULLING REPLICATION LOG and now use SYSTEM CLEANUP or SYSTEM VIRTUAL PARTS UPDATE. #110289 (Groene AI).
• Fix a read-only object-storage replica, with read_only = true, not discovering new parts via refresh_parts_interval, and table_disk = true failing on such a disk with table_disk is not supported for non-ObjectStorage disks. #110460 (Julia Kartseva).
• Fix writes through data-lake table functions such as INSERT INTO FUNCTION icebergS3(...) producing Parquet files whose string columns lost the String type annotation, making the written data unreadable for external readers such as Spark. #110465 (Shaohua Wang).
• Fix schema inference failing with Map cannot have a key of type LowCardinality(Nullable(String)) when reading ORC files whose map keys are dictionary-encoded, including files written by Spark and files written by ClickHouse itself with output_format_orc_dictionary_key_size_threshold enabled. #110492 (Shaohua Wang).
• Fix wrong IS NULL / IS NOT NULL / count() results with optimize_functions_to_subcolumns on a column made Nullable by a metadata-only ALTER MODIFY COLUMN T to Nullable(T). For parts written before the conversion, the .null subcolumn is now derived from the physically present parent column instead of the storage-type default. #110584 (Groene AI).
• Fix wrong query results from the optimize_and_compare_chain optimization when transitive conditions mix types with different comparison semantics, such as Enum with String, Decimal with Float64, or FixedString values of different widths. #110621 (Yarik Briukhovetskyi).
• Fix wrong results for some read-in-order queries when the optimizer widens the sort-key prefix. #110725 (Groene AI).
• Fix CREATE statements with a COMMENT clause to raise a syntax error when the comment string literal is missing, instead of silently accepting a misparsed query. For example, CREATE VIEW v COMMENT AS SELECT 1 no longer creates a view without a comment. #111159 (Shaohua Wang).
• Fix wrong results from SummingMergeTree and CoalescingMergeTree with FINAL when optimize_move_to_prewhere_if_final = 1 and the WHERE clause is an implicit-boolean condition on a sorting-key column, such as WHERE k. #111342 (Groene AI).
• Fix spurious Directory '...' was created concurrently with remote path '...' exceptions for metadata-only operations on object-storage-backed databases and tables, such as DROP DATABASE. #111353 (Mikhail Artemenko).
• Data-lake REST catalogs now report stale credentials with the proper exception instead of silently treating them as a missing table or returning UNKNOWN_TABLE. #111379 (alesapin).
• Fix SQL SECURITY DEFINER and SQL SECURITY NONE not being honored for parameterized views when the analyzer is disabled (enable_analyzer = 0). Previously, selecting from such a view as a user without privileges on the underlying table failed with ACCESS_DENIED, even though the same view works with the analyzer enabled. #111458 (Groene AI).
• Reject a Nullable(Nothing) column, or any other type not allowed in tables, at CREATE TABLE time when the schema is inferred from the storage, for example in a columnless remote, Remote, or Distributed table over a view that does SELECT NULL, instead of persisting metadata that fails to load on the next server restart and blocks loading of unrelated tables. #111487 (Groene AI).
• Fix std::bad_variant_access and incorrect Iceberg schema evolution for nested Array(Map(...)) columns whose inner Struct is altered. #111489 (Konstantin Vedernikov).
• Fix ALTER MODIFY COLUMN failing when converting a Nullable column to a different type with a DEFAULT, for example from Nullable(UInt8) to String, LowCardinality(String), or Array(UInt8). #102156 (DQ).
• Fix BAD_ARGUMENTS (Dictionary not found) from dictGet after DETACH DICTIONARY ... PERMANENTLY is rejected with HAVE_DEPENDENT_OBJECTS; the dictionary now remains usable when the detach is rejected. #105259 (Groene AI).
• ClickHouse-written data files for Iceberg tables now embed Iceberg field IDs in Avro and ORC, ORC String columns are written as string instead of binary, and optional complex ORC fields (Array, Map, Tuple) now preserve the correct iceberg.required metadata. This fixes silent NULL results after column renames and restores compatibility with spec-compliant readers such as Spark and iceberg-java. #109994 (Groene AI). #111775 (Groene AI).
• Fix reading files on plain_rewritable disks after an existing path is rewritten with content of a different size: stale in-memory metadata could otherwise trigger UNEXPECTED_END_OF_FILE. #110304 (Konstantin Bogdanov).
• Fix drift of the per-user memory tracker caused by system log records such as query_log and query_views_log, which could otherwise trigger a false max_memory_usage_for_user limit for users with continuous workloads. #110708 (Raúl Marín).
• Fix a rare race where a comment-only or settings-only ALTER on a ReplicatedMergeTree table could silently drop a column added by a concurrent ALTER ... ADD COLUMN, leaving a replica with outdated table metadata. #111029 (Shaohua Wang).
• Fix incorrect primary-key pruning on tables with a descending sorting-key suffix, such as ORDER BY (g, r DESC), which could drop matching rows. #111059 (Nihal Z. Miaji).
• Fix a peer-certificate memory leak on every TLS handshake when certificate verification is enabled. #111425 (Konstantin Bogdanov).
• Fix mutations with a query parameter as the partition in ALTER TABLE ... UPDATE/DELETE ... IN PARTITION {param:Type}: the serialized partition value can now be parsed back correctly, so the ALTER no longer writes mutation entries that break table loading. #111518 (Michael Kolupaev).
• Fix a macOS/aarch64 crash when a stack trace is captured on a coroutine or fiber stack, for example by the memory profiler. #111656 (Raúl Marín).
• Fix server crash when reading Protobuf data with input_format_allow_errors_num > 0 and a valid message precedes a bad but skippable message in the same block. #107739 (Andrey Tsarevskiy | Андрей Царевский ).
• Fix NOT_FOUND_COLUMN_IN_BLOCK when a query over a Distributed table uses _shard_num or shardNum that becomes Nullable, for example under FULL JOIN with join_use_nulls = 1, together with ORDER BY and the analyzer. #109798 (Groene AI).
• Fix wrong results and UNKNOWN_ELEMENT_OF_ENUM exceptions in LEFT/RIGHT/FULL/ASOF JOIN with join_algorithm = 'full_sorting_merge', where non-joined rows could be filled with 0 instead of the column type default. #111197 (Groene AI).
• Fix a LOGICAL_ERROR exception when a higher-order function’s lambda body is a constant-condition if or multiIf that mixes a Bool literal branch with a UInt8 comparison branch, for example arrayFilter(p -> multiIf(false, true, p = 'ALL'), ['ALL']). #111245 (Groene AI).
• A {_partition_id} placeholder in the path of S3, AzureBlobStorage, URL, and similar file-like engines without an explicit partition_strategy now implies the wildcard strategy again, restoring compatibility for pre-26.6 DDL that started failing with BAD_ARGUMENTS. #111279 (Nikita Fomichev).
• Fix NUMBER_OF_COLUMNS_DOESNT_MATCH for GROUP BY queries over a Merge table wrapping a Distributed table when the grouping key is a function of a column and an aggregate reads the same column. #111334 (Groene AI).
• GRANT role TO role on the same role is now rejected with BAD_ARGUMENTS instead of silently creating a self-referential entry in system.role_grants. #103315 (zxuhan7).
• Fix writes to Iceberg tables when the metadata file has no refs section. #106896 (Aleksandr Musorin).
• Fixed system.session_log becoming unreadable after an Arrow Flight login: the interface column is an Enum8, but the ArrowFlight value was missing from its enumeration, so a login over the Arrow Flight protocol wrote the raw value 10 and any SELECT reading such a row threw the Unexpected value 10 in enum exception. #110758 (Alexey Milovidov).
• Fixed a signed integer overflow in toStartOfInterval with subsecond intervals: for DateTime64 values within one interval of the lower bound of Int64, the rounded result silently wrapped to a positive garbage value; now the function throws a DECIMAL_OVERFLOW exception. #111370 (Alexey Milovidov).
• Fix readWKTPoint and readWKT returning uninitialized (scalar) or previous-row (vectorized) coordinates for a dimension-tagged empty point such as POINT M EMPTY. Such inputs are now rejected with CANNOT_PARSE_TEXT, consistent with POINT EMPTY. #111517 (Groene AI).
• Fix Unexpected number of columns in result sample block exceptions for RIGHT and FULL JOIN when the left side carries a duplicated same-named column, for example 1 AS b, 1 AS b, that survives through UNION ALL. #111554 (Groene AI).
• Fixed CANNOT_CONVERT_TYPE and LOGICAL_ERROR exceptions in distributed queries using the -Tuple combinator with a RESPECT NULLS or IGNORE NULLS modifier under nested combinators, for example anyRespectNullsStateTuple(...) IGNORE NULLS. #111570 (Alexey Milovidov).
• Fixed SQL syntax highlighting in Play on iPhone, where iOS Safari’s automatic text inflation could misalign the highlight overlay. #111683 (Alexey Milovidov).
• Fixed UNKNOWN_QUERY_PARAMETER when a parameterized view is called with a subquery-valued argument whose type is Array, Tuple, or LowCardinality, for example SELECT * FROM v(param = (SELECT groupArray(x) FROM t)). #111790 (Groene AI).
• Fix TYPE_MISMATCH (Key type for complex key ... does not match) when a JOIN onto a dictionary uses a Nullable, LowCardinality, or LowCardinality(Nullable) join key while the dictionary key is not wrapped. The direct-join dictionary lookup now normalizes the key to the dictionary’s declared key type, as dictGet and dictHas already do, and a NULL key never matches. #111857 (Groene AI).
• With fsync_part_directory = 1, renaming a part directory from its temporary to its final name is now itself made durable by fsyncing the parent directory, and for moves across parents the source parent too. Previously only the moved directory was fsynced, so a power loss right after a part was committed could leave the part missing on disk and lose rows despite the setting being enabled. #111861 (Groene AI).
• Fixed hangs in randChiSquared, randStudentT, randFisherF, and randBinomial with extreme parameters. Such queries could not be interrupted by max_execution_time or KILL QUERY. These parameters are now rejected even when max_rand_distribution_parameter or max_rand_distribution_trials is set to 0, and max_rand_distribution_trials can no longer be raised above its default of 10^9 for randBinomial. #111909 (Alexey Milovidov).
• Fixed wrong results when the analyzer rewrites comparison chains such as x = c1 OR x = c2 OR ..., x != c1 AND x != c2 AND ..., or has(const_array, x) to IN and NOT IN. It also fixes Illegal type ... of argument of function in exceptions for expressions with Dynamic, Array(Dynamic), or JSON structure. #111924 (Groene AI).
• Fixed wrong results and LOGICAL_ERROR when selecting a Nullable(Tuple(...)) column together with one of its null-carrying element subcolumns (Variant, Dynamic, or LowCardinality) from a MergeTree table. #111949 (Groene AI).
• Reject native-protocol handshakes that use USER_INTERSERVER_MARKER for an unknown cluster, or for a cluster configured without a <secret> element. Previously such handshakes could enter interserver mode before authentication and read information about local tables. #99675 (Dan Anderson).
• Fix JOIN against EmbeddedRocksDB, Redis, and KeeperMap tables that could silently return zero rows when DirectKeyValueJoin is used with String or Nullable(String) keys. #105253 (Vladimir Cherkasov).
• Fix coordinated refreshable materialized views in a Replicated database getting stuck after re-attach on a Keeper without the MULTI_READ feature flag; the view now stops cleanly with status Disabled instead of logging Unexpected exception in refresh scheduling forever. #108890 (Groene AI).
• Fix the ZooKeeper client sending an overflowed int32 length prefix for requests larger than INT32_MAX bytes. Oversized requests are now rejected on the client side. The limit is configurable via max_request_size under <zookeeper> and is otherwise learned from Keeper. Closes #109165. #109190 (Arsen Muk).
• Fix S3Queue/AzureQueue ordered mode with persistent processing nodes: bucket locks are now refreshed during streaming, so TTL cleanup controlled by persistent_processing_node_ttl_seconds does not remove locks of a live processor. If lock ownership is nevertheless lost, it is detected and reported, and streaming recovers with a fresh file iterator. #110292 (Kseniia Sumarokova).
• Fix wrong data and possible out-of-bounds reads when reading String columns with string_serialization_version = 'with_size_stream' after skipping leading rows, and when reading sub-objects of JSON columns stored with object_shared_data_serialization_version = 'map'. #110471 (Alexey Milovidov).
• Fix CREATE TABLE and ALTER TABLE with the analyzer accepting a DEFAULT or MATERIALIZED expression that references a virtual column such as _table or _database. These expressions are now rejected at definition time instead of failing later on insert with NOT_FOUND_COLUMN_IN_BLOCK. #111776 (Groene AI).
• Fix a race in executable UDFs, the executable table engine, and executable dictionaries where an fcntl on an already-closed file descriptor could corrupt an unrelated descriptor of a concurrent query and make queries hang indefinitely. #111779 (Raúl Marín).
• Creating deprecated minmax statistics now emits a warning instead of throwing an exception. #111785 (Han Fei).
• Fix Iceberg tables written by ClickHouse being unreadable by external readers such as PyIceberg and Spark. Manifest-list and manifest Avro schemas now include the required Iceberg field-id properties. #111786 (Groene AI).
• Regenerate materialized-into-part block_number and block_offset columns during part attach/detach/move/clone operations, so they no longer become stale after part manipulation. #111801 (Mikhail Artemenko).
• Fix the query_plan_split_filter optimization leaving an internal __split_filter column in one branch of INTERSECT or UNION, which could cause a Block structure mismatch in IntersectOrExceptStep stream: different number of columns exception. #111930 (Groene AI).
• Fix AMBIGUOUS_COLUMN_NAME for queries that repeat a JOIN ON condition in WHERE with join_use_nulls = 1; such queries now return their result instead of failing. #112007 (Groene AI).
• Fix DLF authentication in the Paimon REST catalog: every request after the first no longer fails with HTTP 401, and table-existence checks now detect HTTP 404 correctly. #112015 (JIaQi Tang).
• Fix LOGICAL_ERROR during concurrent Kafka table cleanup in Keeper; creation or removal no longer fails if the table’s Keeper path disappears after session expiry. #112078 (Groene AI).
• Fix merges and mutations when the part-level min-max index is enabled for _block_number and _block_offset columns. #109281 (Mikhail Artemenko).
• Fix failed INSERT into DeltaLakeLocal or deltaLake tables leaving an unfinalized data file after writing had already started. #109328 (Groene AI).
• Fix credential leaks in SHOW CREATE, system.query_log, server logs, and EXPLAIN output for s3/s3Cluster, the S3-backed engines, BACKUP ... TO S3, and secret arguments of encrypt, decrypt, and HMAC functions built by expressions or SQL UDFs. #109768 (Raúl Marín).
• Fix reading large Parquet files from uncompressed STORED zip entries via s3 and file. #110495 (Eva Wuuu).
• Fix EXPLAIN ANALYZE hanging on queries with a streaming FROM ... STREAM read hidden in a subquery, CTE, or view. Such queries are now rejected with NOT_IMPLEMENTED, like top-level streaming reads already are. #110935 (Groene AI).
• Reject vectors whose squared magnitude overflows to infinity when using a vector_similarity index with i8 quantization. Previously, such vectors produced silent garbage. #111085 (Groene AI).
• Fsync the text index files of a merged or materialized part when min_rows_to_fsync_after_merge, min_compressed_bytes_to_fsync_after_merge, or fsync_part_directory are set. Previously, these files were the only ones in the part left unsynced, so a power loss right after a merge could leave a committed but broken part and cause data loss. #111335 (Groene AI).
• Fix path validation against user_files. #111442 (Alexander Tokmakov).
• Fix SYSTEM REFRESH on a stopped RabbitMQ table sometimes failing to consume the queued backlog. The refresh operation now waits for the AMQP delivery loop to start before it spends its one out-of-order streaming cycle. #111480 (Groene AI).
• Restrict local object access to the user_files path. #111483 (Konstantin Vedernikov).
• Fix a segfault due to out-of-bounds memory access when deserializing a malformed aggregate function state containing a String value. Such states are now validated and rejected. #111606 (Miсhael Stetsyuk).
• Fix integer overflow while parsing system.zookeeper_info. #111629 (Kseniia Sumarokova).
• Fix NOT_FOUND_COLUMN_IN_BLOCK when a FINAL query filters on a sorting-key column that is not in the SELECT list and the filter is moved to PREWHERE, for example SELECT s FROM t FINAL WHERE k GROUP BY s with optimize_move_to_prewhere_if_final = 1. #111721 (Groene AI).
• Fix partitioned INSERT into DeltaLake with allow_experimental_delta_lake_writes = 1 corrupting partition values that contain path characters and rejecting NULL partition values. Partition values are now percent-encoded into a single path segment, and NULL or empty-string values are written as __HIVE_DEFAULT_PARTITION__ with a null partitionValues entry. #111793 (Groene AI).
• Fix an exception and silently skipped data when reading a dynamic subcolumn through a Buffer table. #111948 (Alexey Milovidov).
• Fix clickhouse-client losing the current database after USE {db:Identifier}. If the client had to reconnect later, it silently switched the session to the default database. #111982 (Alexey Milovidov).
• Fix a LOGICAL_ERROR exception (Type mismatch when building statistics for column) during MergeTree mutations after a metadata-only ALTER MODIFY COLUMN changes a column type while the column carries statistics. Columns actually rewritten by the mutation now have their statistics recomputed, and untouched columns keep their existing statistics. #112009 (Groene AI).
• Fix Paimon table reads that could fail while the mutable snapshot/LATEST hint was being replaced concurrently. Invalid or stale hints are now ignored, and the snapshot is loaded from the catalog instead. #112010 (JIaQi Tang).
• Fix reading GeoParquet, Arrow, and ArrowStream files whose WKT-encoded geometry column uses a non-uppercase type keyword such as point(1 2) or an empty container such as LINESTRING EMPTY or POLYGON(). This also fixes reading back files that store the output of wkt for an empty geometry. #112020 (Groene AI).
• Fix query-level settings overrides not being applied to Azure native copy. #112037 (Smita Kulkarni).
• Fix query_masking_rules rewriting the type of an EPHEMERAL column, which could make CREATE TABLE store the wrong type or fail. #112048 (Alexey Milovidov).
• Table function parameters no longer accept aggregate states, which previously triggered a logical error because such parameters are unsupported. Closes #112085. #112087 (Pedro Ferreira).
• Fix wrong results for a JOIN whose ON expression contains a constant conjunct, for example ON l.id = r.id AND CAST(NULL AS Nullable(UInt8)). #112155 (Vladimir Cherkasov).
• Fix wrong results for hasAllTokens and other multi-token text-index searches on builds that do not use simdcomp, such as aarch64, when text_index_posting_list_apply_mode = 'lazy'. #112178 (Groene AI).
• Fix wrong results when a parameterized view is called with an argument that is not a parameter = value assignment. Such calls are now rejected with UNKNOWN_QUERY_PARAMETER on both execution paths. #112194 (Groene AI).
• Fix parameterized views whose SELECT body uses INTERSECT, EXCEPT, or a UNION chain that mixes UNION DISTINCT with UNION ALL being misclassified as ordinary views, which made their metadata unloadable. #112211 (Groene AI).
• Fix schema inference for the TSV family inferring String for decimal values written with a leading zero, such as 0.0 or 0.5, and consuming such a first row as a header. Also fix TSKV inferring a numeric type for an escape-encoded value such as x=1\x2E5. Closes #112105. #112226 (Groene AI).
• Fix system.s3_queue_settings reporting incorrect values for bucketing_mode, partitioning_mode, partition_regex, and partition_component. #112300 (Bharat Nallan).
• Fix wrong results, wrongly ordered ORDER BY output, and BAD_TYPE_OF_FIELD exceptions when the primary key is an Array or Tuple and the query applies plus, minus, multiply, divide, or intDiv to it. #112339 (Groene AI).
• Fix remote execution of queries containing PASTE JOIN, for example through a Distributed table or a *Cluster table function. ClickHouse no longer formats the query as the invalid syntax ALL PASTE JOIN. #112404 (Groene AI).
• Fix a LOGICAL_ERROR (Invalid number of rows in Chunk) in JoiningTransform for a LEFT JOIN with unique right keys, a mixed ON condition, and a small max_joined_block_size_rows. #106928 (Groene AI).
• Fixed the _tags virtual column for the S3Queue table engine: it was declared but never populated, so SELECT _tags always returned an empty map. It now returns the object tags, like the S3 engine and the s3 table function. #108676 (Groene AI).
• Fix data loss of a column with no default expression when a merge runs concurrently with ALTER TABLE... RENAME COLUMN, or when the column’s only values come from a lightweight UPDATE. The column could be dropped from the merged part, so all of its values read back as NULL. #109356 (Groene AI).
• Fix skip indices being dropped from a data part (or CHECK TABLE failing with UNEXPECTED_FILE_IN_DATA_PART) after a full-part-rewrite mutation, such as ALTER TABLE... DROP COLUMN of a MATERIALIZED column, on a Wide part. #109616 (Groene AI).
• Fixed two bugs on the DateTime64 path of toUTCTimestamp/fromUTCTimestamp (and their to_utc_timestamp/from_utc_timestamp aliases) and toTime64. First, a signed integer overflow near the Int64 boundary (now computed in Int128 and clamped to the representable range). Second, a wrong result for negative fractional values near a timezone-offset boundary: the seconds split truncated toward zero, so timezoneOffset() inspected the next second and could pick the wrong side of a DST/offset change. #109738 (Groene AI).
• Fix a logical error (“Cannot determine row level filter; 0 columns deleted, 0 columns added”) when reading from a Merge table whose child table has a row policy whose filter is a bare existing column (for example CREATE ROW POLICY... USING flag). #109843 (Groene AI).
• Fixed two errors raised when a query with GROUP BY GROUPING SETS (or a scalar or mutation subquery containing one) runs with force_aggregation_in_order = 1: Trying to get name of not a column: ExpressionList under the old analyzer (enable_analyzer = 0), and Memory bound merging of aggregated results is not supported for grouping sets. for a distributed query with enable_memory_bound_merging_of_aggregation_results = 1 under the analyzer. #109862 (Groene AI).
• Fix parsing of numeric 1/0 as Bool values inside container types (e.g. Array(Bool), Tuple(Bool,...)). Previously [1,0] failed with CANNOT_READ_ARRAY_FROM_TEXT while [true,false] worked. #109976 (Groene AI).
• Fixed a spurious ILLEGAL_COLUMN error from ALTER TABLE... ADD COLUMN IF NOT EXISTS when the column already exists at apply time. This happened when the same column was added twice with IF NOT EXISTS in one statement, or when a concurrent ALTER added the column between the prepare and apply phases. #110080 (Groene AI).
• Fixed a wrong result where length (and the QBit dimension) on a FixedString column returned 0 instead of the fixed size N under parallel replicas when the query had a selective filter. #110427 (Groene AI).
• Fixed flipCoordinates losing the Geometry type: for a Geometry argument the result is now typed Geometry again instead of the underlying Variant(...), so it can be passed directly to functions like areaCartesian. #110694 (Groene AI).
• Fixed optimize_injective_functions_in_group_by changing query results for GROUP BY GROUPING SETS and GROUP BY... WITH TOTALS. The setting is documented as result-preserving, but eliminating an injective function of a grouping key produced a wrong value (or dropped a row) for rows where that key is absent from the aggregated set (a GROUPING SETS non-member set) or for the WITH TOTALS row. The wrong value could coincide with a genuine key value, making a super-aggregate row indistinguishable from a real group in the same result set. #110721 (Groene AI).
• Numeric 1/0 in JSON Bool parsing now honors the allow_special_bool_values setting. Previously JSON input such as {"v":1} for Variant(Bool, UInt32) was read as Bool even with allow_special_bool_values_inside_variant = 0, because the JSON Bool parsers ignored the setting and Bool has a higher Variant deserialize priority than integer types. #110835 (Groene AI).
• Fix a logical error block.rows() == getRows() (an out-of-bounds read and broken insert deduplication in release builds) when an INSERT flows through a dependent materialized view whose target is an Alias and whose inner query changes the number of rows, with a deduplicating table reachable behind the alias hop. #111103 (Alexey Milovidov).
• Fixed a logical error Block structure mismatch (in debug and sanitizer builds) and an Illegal types of arguments error for set operations over compatible aggregate-state columns (e.g. quantileState and quantilesState(0.9)) nested inside container columns such as Tuple, Array, Map, Nullable, or Variant. #111191 (Alexey Milovidov).
• Fixed DISTINCT with an ordered LIMIT/OFFSET returning wrong rows when the DISTINCT early-stop limit hint could not bound the head of the result: a negative LIMIT (returned the head instead of the tail), a fractional LIMIT/OFFSET (the fraction is only resolved after a full read), and a bare OFFSET with no LIMIT (the tail after the offset was dropped). For example SELECT DISTINCT intDiv(x, 100) FROM t ORDER BY intDiv(x, 100) LIMIT -1 over a multi-part table returned the smallest distinct value instead of the largest. #111326 (Groene AI).
• Fixed wrong results when a query has a set skip index and a predicate of the shape (x AND <null-value>) OR y, where the NULL is produced by a function (e.g. toInt64OrNull('x'), nullIf(1, 1), CAST(NULL AS Nullable(Int64))). Such a value was wrapped by __bitWrapperFunc and propagated NULL instead of the “unknown” mask, so the set index could prune granules that actually match and the query returned fewer rows than it should. #111604 (Groene AI).
• Fix DELETE/UPDATE on an Iceberg table silently removing the wrong rows (or throwing NOT_FOUND_COLUMN_IN_BLOCK) when the table contains data files written both before and after ALTER TABLE... ADD COLUMN and optimize_move_to_prewhere is enabled. #111693 (Groene AI).
• Fixed a LOGICAL_ERROR (“Invalid number of rows in Chunk”) — and, under hardened builds, a related out-of-bounds abort — for ORDER BY ... WITH FILL ... INTERPOLATE over a Distributed table with two or more shards, over clusterAllReplicas, or over custom-key parallel replicas, including on empty results. WITH FILL is now applied only on the node producing the final result, which also fixes duplication of the generated fill rows across shards/replicas. #111919 (Yakov Olkhovskiy).
• Keep randBinomial usable with the degenerate probabilities 0 and 1 for any number of trials. #112021 (Alexey Milovidov).
• Fixes wrong results caused by a vector-search query poisoning the query condition cache. A SELECT... WHERE <condition> ORDER BY <distance function> LIMIT n query over a table with a vector_similarity index could record granules as not matching <condition>, so a later ordinary SELECT... WHERE <condition> on the same table silently returned fewer rows. Vector-search reads no longer write to the query condition cache. #112086 (Groene AI).
• Fixes a bug where a CREATE VIEW or CREATE FUNCTION whose definition calls substr, mid or byteSlice in a form the substring grammar cannot re-parse is accepted, but the metadata ClickHouse writes for it is not valid SQL. Reading that definition afterwards fails with SYNTAX_ERROR, and at server startup it aborts metadata loading, so the server cannot start at all. Such definitions now round-trip, and existing metadata files of the form substring(x, a, b, c) become readable again, so an affected server starts up without manual intervention. #112195 (Groene AI).
• Fixes a TimeSeries table whose local metadata diverged from Keeper making replica recovery in a Replicated database hang forever, so the replica never served the database again and stayed broken across restarts. The inner tables of such a table are now dropped while the recovery metadata transaction is still available, instead of being deferred to a background task that could not execute the DROP at all. #112218 (Groene AI).
• Fixed reading a subcolumn of a Nullable(Tuple(...)) column whose element is a non-nullable LowCardinality(T) when the same query also reads the parent subcolumn. On a Compact part the parent returned wrong values, or the query failed with Unexpected return type from assumeNotNull, or the server was killed by a string_view hardening assertion, because the extracted subcolumn was deserialized into a LowCardinality(Nullable(T)) buffer whose substreams were then shared with the parent’s read. #112232 (Groene AI).
• Fixed a permanent server hang on macOS caused by pthread_rwlock losing wakeups when the sampling query profiler is enabled (Apple bug FB24027930): pthread_rwlock is now replaced with a signal-robust implementation on Darwin. #112267 (Raúl Marín).
• Fix max_execution_time and KILL QUERY being ignored while a query evaluates geohashesInBox, arrayFold, sleep, base58Decode, h3PolygonToCells, or h3PolygonToCellsWithContainment. Such queries could keep running for minutes after cancellation. This now also covers evaluations inside expressions stored in table metadata, such as sorting keys, skip indexes, and PARTITION BY keys. #112273 (Groene AI). #113456 (Groene AI).
• Fixes a logical error Cannot add step Expression to QueryPlan because it has incompatible header with root step JoinLazyColumnsStep for a query with FINAL, a PREWHERE filter, a small LIMIT and make_distributed_plan = 1, when the selected columns are listed in an order that differs from the table’s column order. Lazy materialization replaced a plan node without preserving that node’s output header, which made the query plan inconsistent. #112303 (Groene AI).
• Fixes writing Iceberg tables whose schema contains a tuple nested inside a tuple. The published table metadata replaced the inner tuple’s field names with positional names 1, 2,…, so with the default Parquet format the INSERT failed with INCORRECT_DATA and with Avro the column could not be read back. #112480 (Groene AI).
• Fix ATTEMPT_TO_READ_AFTER_EOF error when merging parts with a text index if one of the merged parts was empty, for example after a mutation that deleted all rows of the part. #112490 (Anton Popov).
• Fixes wrong results and a logical error when an argument that must be constant, such as the scale of toDecimal32 or the timezone of toDateTime, comes from a scalar subquery and the old analyzer is used. Such an argument was analyzed as if its value were zero or empty, so the declared result type disagreed with the computed value, and CREATE VIEW and CREATE MATERIALIZED VIEW persisted a wrong column type. #112492 (Groene AI).
• Fixes a LOGICAL_ERROR (“Part… contains column… that is absent in table…”) when a ReplicatedMergeTree replica mutates a part it fetched from another replica before applying its own pending ALTER... ADD COLUMN. Such a mutation is now postponed until the metadata change is applied. #112510 (Groene AI).
• Fixed incorrect query results when using GROUP BY with ORDER BY and arrayJoin in the projection. #101775 (Yash ).
• Fix a server crash (out-of-bounds access) in S3Queue/AzureQueue with enable_hash_ring_filtering = 1 when a batch contained a non-processable file and the Keeper request to set the batch as processing failed at the same time. #108977 (Groene AI).
• Fix a logical error Stream <col>.variant_discr... is not found (server abort in debug and sanitizer builds) when ALTER TABLE... MODIFY COLUMN changes a column into a type with dynamic subcolumns (for example Variant into Dynamic) on a table with Wide parts, and the part is later read or merged. #110204 (Groene AI).
• Fix a signed-integer overflow (UBSan) when converting an extreme DateTime64 value (e.g. INT64_MIN) to Time/Time64 in a timezone with a negative offset. #110451 (Groene AI).
• Fixed a MergeTree table with a _part_offset or commit-order projection becoming permanently unattachable after allow_part_offset_column_in_projections or allow_commit_order_projection was disabled and the table was detached or the server restarted. These two settings are CREATE-time gates, so they are no longer re-checked on ATTACH. #110570 (Groene AI).
• With analyzer_compatibility_join_using_top_level_identifier = 1, an identifier in JOIN... USING can now be resolved from an alias defined on a subexpression inside the SELECT list (for example, SELECT uniqExact(lower(x) AS id) FROM t1 JOIN t2 USING (id)), matching the old analyzer behavior. Previously only top-level projection aliases were considered, and such queries failed with an UNKNOWN_IDENTIFIER exception even with the setting enabled. The error hint suggesting the setting is now also produced when the matching alias is nested. #110739 (Dmitry Novik).
• Fixes an issue where unauthenticated requests could cause ClickHouse to resolve hostnames supplied in forwarded client-address headers. ClickHouse now accepts only numeric IP addresses from these headers. Invalid values are ignored for forwarded-address quota attribution without DNS resolution. When auth_use_forwarded_address is enabled, invalid forwarded addresses are rejected during authentication. Rejections are logged at debug level. #111060 (Dmitriy Borisenko).
• Fix NOT_FOUND_COLUMN_IN_BLOCK when a row policy and additional_table_filters reference the same column that is not otherwise selected by the query. #111099 (Groene AI).
• Fix UNKNOWN_IDENTIFIER error for columns qualified by CTE name (cte_name.column) in queries stored in views when the analyzer is enabled. #111386 (Dmitry Novik).
• Fix a LOGICAL_ERROR (Got read request from replica N for unknown stream...) that could abort the server when reading with parallel replicas and the query is fully answered by a projection optimization (exact-count, minmax-count, or a stored aggregate/normal projection selecting no ranges) on the initiator. #111689 (Groene AI).
• Fixed a signed integer overflow in the three-argument overload of toStartOfInterval: for an origin argument near the lower bound of Int64 the function returned a wrapped-around value instead of reporting an error. #112045 (Alexey Milovidov).
• Fixes a LOGICAL_ERROR exception (Block structure mismatch in joined block stream) on a multi-table join whose two sub-joins both produce no output columns, for example SELECT count() FROM t1, t2, t3, t4 WHERE (t1.b = t2.b) AND (t3.a = t4.a) with query_plan_optimize_join_order_limit set to 0 or 1. #112205 (Groene AI).
• Fixed the values table function rendering a non-literal Bool constant (e.g. CAST(true, 'Nullable(Bool)')) inserted into a String/textual column as 1/0 instead of true/false. #112308 (Yakov Olkhovskiy).
• Fix a LOGICAL_ERROR (partitions_count > 0) when partitioning Iceberg data chunks emptied by position deletes, e.g. during compaction of a partitioned Iceberg table with fully deleted data files. #112432 (Pedro Ferreira).
• Fixed schema inference proposing a numeric type for text values the value parser cannot read back. Inference accepted a dangling exponent (1e+), a missing mantissa (.) or a repeated sign (++inf), which readFloatTextPrecise then rejected with CANNOT_PARSE_NUMBER when reading the inferred schema; on a Dynamic column this was an error at INSERT time. Inference now validates the number with the parser the reader actually uses. This also removes a case where the inferred type depended on max_read_buffer_size, and lets valid leading-zero exponent forms such as 0e5 infer Float64 in TSV as they already do in CSV. Follows #112226 and closes the exponent half of #112105. #112453 (Groene AI).
• Report the correct position for errors in PromQL queries. #112494 (Minh Vu).
• Fixes a LOGICAL_ERROR (Cannot add transform Resize to Pipes because it has no outputs) when a query aggregating over a cluster runs with serialize_query_plan = 1 and plan-based parallel replicas, caused by the merge step resizing the pipeline to zero streams. #112506 (Groene AI).
• Fixes an unsigned underflow in the table name length check that allowed creating a table which could never be dropped. When the escaped database name reached 214 bytes, the limit computed for the metadata_dropped filename wrapped around, so CREATE TABLE was accepted and the subsequent DROP TABLE or DROP DATABASE failed with File name too long. Such a name is now rejected with ARGUMENT_OUT_OF_BOUND, the same error already returned just below that boundary. #112527 (Groene AI).
• Fix the Context has expired exception when getClientHTTPHeader is used inside a subquery. #112533 (Alexey Milovidov).
• Fixed s3_allow_server_credentials_in_user_queries incorrectly rejecting a query-supplied extra_credentials(role_arn = '...'). STS assume-role with an explicit role is now always allowed: it does not expose the server’s own credentials to the query, and it is the documented way to grant access to a private S3 bucket. #112603 (Raúl Marín).
• Fixes the read counters (read_rows, read_bytes, ProfileEvents['SelectedRows'], ProfileEvents['SelectedBytes']) and the READ_ROWS and READ_BYTES quotas being over-charged for queries whose ORDER BY spills to disk. Rows re-read from the external sort temporary files were counted again as source reads, so a spilling query over-reported what it read and could be rejected with QUOTA_EXCEEDED even though it stayed within its quota. #112645 (Groene AI).
• Fixes an Iceberg table whose write format is Avro serializing a field declared "required": false whose type is a list, map or struct as required, with no ["null", T] union, so that the data file’s own schema disagreed with the table metadata and other Iceberg readers saw a required field. Such a field is now written as the ["null", T] union the Iceberg spec uses for an optional one. #112648 (Groene AI).
• Fixes DeltaLake tables returning too few rows when the WHERE predicate contains a sub-expression the delta-kernel predicate translator cannot handle, positioned under a NOT. Such a sub-expression is now reported to delta-kernel as an explicit unknown predicate instead of as the end of the child iterator, which used to silently truncate the enclosing conjunction. #112652 (Groene AI).
• Fixes heap corruption, which could crash the server at shutdown, caused by the const method AccessRights::getFilters modifying an access-rights tree shared by concurrent queries. #112692 (Groene AI).
• Fixes a logical error in query formatting when an ARRAY JOIN clause carries no expressions. Such a clause is now rejected while parsing with SYNTAX_ERROR, consistent with an empty USING, instead of being accepted and formatted into text that cannot be parsed back. #112699 (Groene AI).
• Fix a query hang in local execution of distributed query plans (distributed_plan_execute_locally = 1): the in-memory exchange reader could block the only pipeline thread before the sources producing its data were started. #111762 (Alexander Gololobov).
• Fix Iceberg fields declared with "required": false whose type is list, map, or struct being written as REQUIRED in the Parquet footer. Such fields are now written as OPTIONAL, matching the published Iceberg schema and other readers. #112669 (Groene AI).
• Fixed a heap-use-after-free in the pipeline executor that could crash the server when a query removes processors at runtime (for example FINAL with lazy reads or external sort), found by the AST fuzzer. #111017 (Groene AI).
• Fix a use-after-free (segmentation fault) when a query with a shared storage snapshot (setting enable_shared_storage_snapshot_in_query) strips data parts from the snapshot — e.g. a streaming subquery reading the same MergeTree table that another part of the query reads concurrently. #111192 (Alexey Milovidov).
• Reject unsupported uses of AggregateFunction columns in TTL expressions at CREATE TABLE time (e.g. TTL toDateTime(state)) instead of failing later during TTL execution with ILLEGAL_TYPE_OF_ARGUMENT. This also covers states carried inside Variant alternatives and Dynamic values. Valid state-aware consumers such as finalizeAggregation are still accepted. #107366 (Ria Khatoniar).
• Fix a LOGICAL_ERROR (CTE '...' does not have query tree, but was not planned yet) when a MATERIALIZED CTE is referenced from more than one join-tree position (for example two joined subqueries, or a joined subquery plus a scalar subquery in WHERE) in a query over a Distributed table. The query now executes instead of raising the exception (it aborted the server in debug/sanitizer builds). #108924 (Groene AI).
• Preserve original key order in bucketed Map serialization to fix comparison operations that depend on it. #109178 (Pavel Kruglov).
• Fixed made_current_at in system.iceberg_history becoming 1970-01-01 for a retained snapshot after ALTER TABLE... EXECUTE expire_snapshots on an Iceberg table. Also fixed system.iceberg_history returning no rows for a table whose metadata has no snapshot-log. #111781 (Groene AI).
• Fixes a FixedString text try-parse that leaves partially appended bytes in the column when parsing fails, which can cause a Sizes of nested column and null map of Nullable column are not equal logical error during serialization of Variant/Array/Nullable columns, or silently shifted result bytes. #111908 (Groene AI).
• Fixes replaceOne, replaceAll, replaceRegexpOne and replaceRegexpAll ignoring max_execution_time and KILL QUERY. A single call now checks for cancellation while it works, so a long-running replacement over a large value or a large number of rows stops instead of holding a server thread until it finishes. Such a call has no partial result to return, so under timeout_overflow_mode = 'break' it raises the deadline as an error: always when constant-folded, and in the pipeline whenever this checkpoint is reached before the executor’s between-block check, which still stops the query silently. #112483 (Groene AI).
• Fixes a logical error and a wrong dependency record when an ordinary VIEW body references a table or dictionary through a query parameter, as in CREATE VIEW v AS SELECT x FROM {db:Identifier}.t. Such a view no longer records a referential dependency on a same-named table of the current database, so unrelated tables can be dropped again. #112704 (Groene AI).
• Fixes a wrongly typed read of a Nested element subcolumn whose column was dropped and re-added. Selecting such a subcolumn together with its own parent column returned the whole element in the subcolumn’s slot while the block still declared the element type, which produced arbitrary values in release builds and a logical error in debug and sanitizer builds. #112769 (Groene AI).
• Fixed Cannot read all data of type FixedString when reading the product_quantization_codebook subcolumn of a column with a Quantized('product',...) codec. The per-part codebook is written after the data of all granules, so it is not delimited by marks, and a read whose mark range ended before the part’s final mark could read it short. #112818 (Alexey Milovidov).
• Fixes a segfault and silent data corruption when an INSERT into the File engine or the file table function appends to a non-empty file in a format that does not support appending, such as Avro. Writing through a file descriptor or a partitioned path bypassed the existing check, so the format prefix was suppressed and a second header was written after the existing bytes, leaving the file unreadable. Such an INSERT is now rejected with CANNOT_APPEND_TO_FILE, as it already is for a plain path. #112839 (Groene AI).
• Fix connectionId raising Context has expired when it is used inside a subquery by capturing the client’s connection id when the function is built. Closes #112533. #112870 (ClickGap AI Bot).
• Fixed parsing of integers that do not fit into the target type in Poco::NumberParser, which is used to read JSON numbers. Such a number was silently parsed as a wrong value (for example, 18446744073709551617 became 1) instead of being rejected, and a UInt64 number above the Int64 maximum did not survive an AST JSON round trip. #112904 (Alexey Milovidov).
• Fix IS DISTINCT FROM for Array and Map values compared with NULL. Closes #103042. #103162 (yanglongwei).
• Prevent exhausted retries for leader-executed replicated distributed DDL from permanently blocking subsequent ON CLUSTER DDL queries. #108505 (Ahaan Limaye).
• Fix a LOGICAL_ERROR (“Trying to extract chunk from ChunkBuffer before all inputs are finished”) when a set operation (INTERSECT / UNION ALL / EXCEPT) combines branches that use correlated subqueries and correlated_subqueries_default_join_kind = 'left'. #108554 (Groene AI).
• Conversion of out-of-range floating-point values (BFloat16, Float32, Float64) and of UInt64 values above Int64::max() to DateTime and Time now reliably saturates to the range boundaries on all architectures (previously the result of an out-of-range conversion was architecture-dependent and could wrap around on x86-64), and non-finite values (NaN, inf) now throw an exception instead of producing an arbitrary result. The same fix is applied to the float-to-Date conversion path of toDate and to the conversions of UInt64 and of wide integers ((U)Int128, (U)Int256) to Date and Date32, which could return a pre-epoch day instead of saturating. Conversion of a number to Time now saturates to the range of the type for every numeric source type (previously values from narrow unsigned or wide integer sources were stored unclamped and clamped only when printed), and the accurate cast to Time no longer rejects the negative values that Time supports, nor silently saturates the values that do not fit into it. The accurate cast of a number to Date32 (accurateCast, accurateCastOrNull) no longer silently saturates an out-of-range or non-finite value, but throws or returns NULL as it already did for Date and DateTime, and the OrDefault flavours return the default value of Date32 for such an input. The accurate cast of a non-integral floating-point number to Date, Date32, DateTime or Time is now rejected instead of being silently truncated. Also fixed the toDateTime32 function documentation, which mistakenly showed toDateTime64 examples and described the DateTime64 value range. #110459 (Alexey Milovidov).
• ALTER TABLE... DETACH PARTITION/PART now fsyncs the clone it creates under detached/ (the part subtree plus the directory chain up to the disk root) when the table’s fsync_part_directory setting is enabled, for both Full (directory) and Packed part storage. Previously the clone was never fsynced while the empty covering part that removes the rows from the active set was, so a power loss right after the acknowledgement could destroy the detached data on btrfs (the un-synced clone directory entries roll back, leaving detached/ empty and nothing to ATTACH). Gated on the existing fsync_part_directory setting; default behavior is unchanged. #111426 (Groene AI).
• Fix CREATE OR REPLACE leaving the replaced object behind under an internal name when the caller lacks the drop privilege on it. #111466 (Nikolay Degterinsky).
• Fixed a crash when running a STREAM read (enable_streaming_queries = 1) over a table with more than one partition and a row policy defined on it. In debug and sanitizer builds it failed with the Filter column... not found in DAG outputs logical error; in release builds it could segfault. #111572 (Groene AI).
• Fix count() on Iceberg tables whose snapshot summary is inaccurate or corrupted: the trivial count optimization now derives exact row counts from the manifest files, and falls back to a real scan when row-level deletes make metadata counts inexact. Also fix Iceberg OPTIMIZE TABLE writing incorrect manifest-list counts and lineage, and make totalBytes metadata robust against corrupted summaries. #111617 (Melvyn Peignon).
• Fix a DEFAULT column missing in a part being filled with the type default instead of the DEFAULT expression when the PREWHERE condition was split into multiple read steps. Closes #111757. #111795 (Yarik Briukhovetskyi).
• Fixed Expected UInt64 column for __grouping_set, got BLOB and ColumnBLOB should be converted to a regular column before usage logical errors on a distributed query executed on a shard with enable_parallel_blocks_marshalling (enabled by default), for example with GROUP BY GROUPING SETS over a cluster that has a local replica, or when reading a Merge table that spans a local and a distributed table. BlocksMarshallingStep is no longer added to plans whose blocks are consumed in the same process. #111997 (Groene AI).
• Operations now fail when corrupt MergeTree statistics are loaded instead of ignoring the error. #112201 (Sergey Kuznetsov).
• Fixes a logical error when PREWHERE is used on a Merge table over another Merge table or a materialized view whose column type differs. Also fixes a row policy being silently ignored on tables that read from remote servers (a Distributed table, or a table wrapping one), which could expose rows the policy was meant to hide; such queries are now rejected with an ILLEGAL_PREWHERE error. Define these policies on the underlying local tables on each remote server instead; note that such a policy is still not applied to reads shipped with serialize_query_plan = 1, which is a separate, pre-existing bug. #112326 (Pedro Ferreira).
• Fixes connection_pool_max_wait_ms so that its documented and default value 0 means an infinite timeout. A query that found the connection pool full used to retry in a tight loop, consuming CPU and logging No free connections in pool. Waiting 0 ms. on every iteration. It now waits until a connection is returned to the pool. Closes #112053. #112380 (Groene AI).
• Fixes silent loss of acknowledged Iceberg writes on HDFS: an Iceberg commit publishes each new metadata file with a conditional write, but HDFSObjectStorage dropped the condition and performed a plain overwrite, so concurrent writers were all acknowledged while one writer’s snapshot became unreachable. HDFS now refuses a conditional write instead, so storage-managed Iceberg writes on HDFS, CREATE TABLE... ENGINE = IcebergHDFS(...) with an explicit column list included, now fail. Reads are unaffected. #112437 (Groene AI).
• Fix an out-of-bounds read (a crash in release builds) in insert deduplication when an INSERT flows through a materialized view whose inner query changes the number of rows and the deduplicating target table is partitioned. DeduplicationInfo::filterToPartition attributed each deduplication token to partitions by walking the token’s source-row range over a partition selector sized to the (smaller) view-output block, reading past its bounds. It now keeps every token in every partition for any materialized-view target — where the source-row-to-partition mapping no longer exists after the row count changed — instead of the out-of-bounds walk (and instead of the NOT_IMPLEMENTED that a later guard raised for the same insert). #112649 (Sema Checherinda).
• Fixes a logical error (which aborts in debug and sanitizer builds) when finalizeAggregation is applied to a column whose aggregate states come from different functions that share a state representation but finalize to different types — for example quantileState and quantilesState(0.9) brought together by a UNION or produced via arrayReduce. Such queries now raise a normal error instead of crashing. #112662 (Zain Ul Abidin).
• Fix DELETE on KeeperMap tables deleting only the first max_block_size matched rows while reporting the mutation as successfully completed. Also fix DELETE with keeper_map_strict_mode = 1 not being atomic: it could fall back to unversioned removals and delete rows that were concurrently updated, and it applied the matched rows block by block so a conflict could leave the delete partially applied. #112777 (Alexey Milovidov).
• Fix ALTER on MergeTree tables being applied while an incomplete RENAME COLUMN alter-mutation is still in progress. That could update the storage metadata incompatibly and block later mutations completely. #112783 (Mikhail Artemenko).
• Fixed server startup and ATTACH failing with WITH RECURSIVE is not supported with the old analyzer for a view with a recursive CTE when enable_analyzer = 0 is the server default. #112784 (Nikolay Degterinsky).
• Fixes a logical error when a Join table engine is used on the right side of a JOIN whose ON section contains a condition referencing columns of both tables, for example ON (t.key = j.key) AND (t.a < j.a). Such a query is now rejected with a clear exception instead of raising required columns:... but not found any in left table or std::bad_variant_access. #112815 (Groene AI).
• Fixes wrong results for a JOIN whose ON clause carries a mixed condition, that is a cross-side non-equi residual such as ON (t1.key = t2.key) AND (t1.a * 10 < t2.a). Only the hash family evaluates such a condition, but full_sorting_merge, partial_merge and the direct key-value join claimed they could run these joins and then silently ignored the residual, returning extra rows. They now decline, so a hash algorithm is used instead. #112831 (Groene AI).
• Fix ReadBufferFromEncryptedFile: Wrong file position logical error when reading a Compact part from an encrypted disk with direct I/O, which made merges get stuck in system.replication_queue. #112943 (Alexey Milovidov).
• Fix a quadratic slowdown when parsing a PromQL query with a long tail of unrecognized characters (e.g. a FixedString value padded with NUL bytes): the parse of a megabyte-sized invalid input used to keep a thread busy for tens of minutes and was uncancellable, because it happens during query analysis. Now the parser stops at the first error. #112944 (Alexey Milovidov).
• Reject PromQL durations whose units are repeated or not ordered from longest to shortest. #112954 (Minh Vu).
• Fix a hang when writing Parquet with output_format_parquet_parallel_encoding enabled (the default) and max_threads greater than 1. If the encoder failed to schedule an additional thread, its live-thread counter underflowed and the write could stop making progress permanently instead of finishing or reporting an error. #112959 (Groene AI). #113170 (Alexey Milovidov).
• Fixed geohashesInBox spending an unbounded amount of time on a box of zero area, such as geohashesInBox(0., 0., 180., 0., 12). Such a query returned the correct single geohash but burned roughly 5e8 empty loop iterations per row and ignored both max_execution_time and KILL QUERY. #112976 (Groene AI).
• Fixed undefined behavior (null pointer passed to memcpy) in replaceRegexpOne/replaceRegexpAll when the replacement string substitutes a capturing group that did not participate in the match, e.g. replaceRegexpAll('abc', '(a)|(b)', '\\2'). #113012 (Alexey Milovidov).
• Reject a SETTINGS change marked as written without a value when it carries a value other than true, and never elide the value of such a change when formatting a query. Previously a crafted AST JSON payload could execute a Bool setting with false while system.query_log and formatQueryFromJSON showed the valueless form. #113025 (Alexey Milovidov).
• Fixed undefined behavior and a silent scheduling error when a background task is scheduled with a very large delay, for example by a refreshable materialized view with a huge REFRESH AFTER period. Delays are now bounded to the largest representable value instead of overflowing. #113041 (Groene AI).
• Fixed an out-of-bounds write when reading a Parquet DECIMAL column whose physical type is wider than the type its declared precision maps to, for example DECIMAL(9, 2) stored as physical INT64. Such files are valid Parquet and other writers produce them, but the reader sized the destination column from the declared precision while the decoder wrote the physical width, corrupting memory. Reading such a file now also raises DECIMAL_OVERFLOW when a value does not fit the declared precision, instead of returning corrupted data, and reads losslessly with a type hint at least as wide as the physical type. #113046 (Groene AI).
• Fixed intDiv on a non-constant column with the constant divisor -1 silently wrapping the division of the minimal signed number (e.g. intDiv(-9223372036854775808, -1)) instead of throwing the ILLEGAL_DIVISION exception like the other execution paths do. The silent wrap could also produce wrong ORDER BY/DISTINCT results, because the read-in-order optimization correctly treats intDiv by a negative constant as monotonic, and the wrapped value violated the resulting sort order. #113048 (Alexey Milovidov).
• Fix an incompatibility that text indexes created in ClickHouse 26.3 with unicode_word tokenizer cannot be loaded in later versions. #113061 (Robert Schulze).
• Fixed a Pipeline stuck logical error in queries that scatter data by partition, such as a window function with PARTITION BY or a join with join_algorithm = 'parallel_full_sorting_merge', when one shard’s downstream finished while a block was only partially distributed. #113190 (Groene AI).
• Fixed wrong results when a hash join has a single LowCardinality key of a type wider than 8 bytes (UInt128, Int128, UInt256, Int256, and their Nullable variants). Such a join silently returned rows whose key values are not equal. #113230 (Groene AI).
• Fix ALTER TABLE ... ADD PROJECTION over a table-level ALIAS column leaving the table unusable, with subsequent INSERT failing with UNKNOWN_IDENTIFIER (Missing columns), when the session had optimize_respect_aliases = 0. #109091 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK and LOGICAL_ERROR exceptions in a direct JOIN over a MergeTree right table, and related aborts in debug and sanitizer builds. #109932 (Groene AI).
• Fix comparisons that previously threw ILLEGAL_TYPE_OF_ARGUMENT when the operands have no least common supertype. Arrays such as Array(Int64) vs Array(UInt64) and Array(Int256) vs Array(UInt256) can now be compared lexicographically with =, !=, <, <=, >, >=, IS DISTINCT FROM, and IS NOT DISTINCT FROM. isDistinctFrom and isNotDistinctFrom now also compare supported non-array numeric and decimal pairs by value instead of throwing, and equality comparison for arrays nested inside Tuple(Nullable(...)) is also fixed. Closes #33897. #110245 (Diego Gomes Tomé).
• Fix a LOGICAL_ERROR (Bad cast from ColumnString to ColumnLowCardinality) during primary-key index analysis when a LowCardinality key column is wrapped in a nested CAST chain that re-introduces LowCardinality, e.g. WHERE CAST(CAST(s, 'LowCardinality(String)'), 'String') < '5'. In debug and sanitizer builds this aborted the server; in release it failed the query. #111050 (Groene AI).
• Fix wrong results for a FULL JOIN on a Join engine table with USING when the left key is Nullable and the storage key is not: an unmatched left row with a NULL key returned 0 instead of NULL. #111371 (Groene AI).
• Fix stale and deleted row versions being visible when a MaterializedPostgreSQL database is read through a Merge table. Such reads now go through the MaterializedPostgreSQL wrapper, with the same FINAL and _sign = 1 filtering as direct reads. #112382 (Alexey Milovidov).
• Fix Bad cast from ColumnNullable to ColumnString when a mutation that rewrites the whole part runs on a part whose column type is older than the one in the table metadata, for example after ALTER TABLE ... MODIFY COLUMN followed by ALTER TABLE ... MATERIALIZE PROJECTION. #112501 (Alexey Milovidov).
• Fix row-based input formats ignoring KILL QUERY and max_execution_time for minutes while parsing a block. They now check for cancellation every 8192 rows, so a killed INSERT (including a background async-insert flush) stops promptly instead of running to the end of the block. #112646 (Groene AI).
• Fix RENAME TABLE checking the maximum table name length against the source database instead of the destination one. Renaming a table into a database with a long name could produce a table that DROP TABLE could not remove, and renaming a table out of such a database was wrongly rejected. #113019 (Groene AI).
• Fix UNKNOWN_IDENTIFIER when selecting from a VIEW whose declared column types differ from the types its inner query produces, for example SELECT sum(length(arr)) FROM v where v declares arr Array(UInt8) over a String column. With the default optimize_functions_to_subcolumns = 1, the optimizer rewrote length(arr) into a read of the arr.size0 subcolumn and forwarded that name into the inner query, which has no such subcolumn. #113057 (Groene AI).
• Fixed an exception (in release builds, an error that prevents the server from starting) when a mutation predicate uses the merge table function without an explicit database name. The database of the mutated table is now used, the same way as for table names. #113185 (Alexey Milovidov).
• Fix wrong results and a LOGICAL_ERROR when reading a subcolumn of a dropped-and-re-added Nested member together with a present member of the same group, on a MergeTree table with share_nested_offsets = 1. The present members returned values from a later block whenever the read spanned more than one block. #113225 (Groene AI).
• Fix a hang of ALTER, OPTIMIZE, TRUNCATE, and partition-manipulation statements with alter_sync = 2 on ReplicatedMergeTree when the table starts shutting down while the statement waits for another replica, for example because of a concurrent DROP DATABASE ... SYNC. Such a statement now fails with UNFINISHED and also honors max_execution_time and KILL QUERY. #113228 (Groene AI).
• Do not write ANSI escape sequences to INTO OUTFILE when stdout is a TTY. #113252 (Azat Khuzhin).
• Fix a rare use-after-free in the async insert queue when different entries have identical deadlines. #113363 (Miсhael Stetsyuk).
• Fix silently empty tables when more than one ENGINE = Backup database is created from the same backup, or when such a database is recreated under a different name. The storage policy that backs a Backup database is now identified by a name that also covers the local database name, so each database gets the disk built for its own data path. #113378 (Groene AI).
• Fix rows never being deleted when a MergeTree table has two or more TTL ... DELETE WHERE rules whose time expression is identical and which differ only in the WHERE condition. After a TTL delete merge, the part’s rows_where_ttl_info was reset to zero, so no further TTL merge was ever scheduled for it and the expired rows were retained indefinitely with no error or log message. Closes #113116. #113384 (Groene AI).
• A query over a Merge table (or the merge table function) that matches many tables now reacts to KILL QUERY and max_execution_time while the per-table query plans are being built, instead of only after the planning of all children has finished. With timeout_overflow_mode = 'break', it now also stops building query plans for the remaining matched tables and returns the partial result without an error, as that mode documents. #113415 (Alexey Milovidov). #113642 (Groene AI).
• Fix clickhouse-client exiting with code 139 instead of 0 after cancelling a query with Ctrl+C, or after an INSERT that reads data from stdin. #113431 (Groene AI).
• Allow the preprocessor and postprocessor expressions of a text index to reference ALIAS columns. #110014 (Jimmy Aguilar Mena).
• Fixed several cases where parallel replicas returned duplicated rows: distributed_aggregation_memory_efficient was ignored by plan-based parallel replicas, so the buckets of the memory-efficient merge arrived out of order and were merged twice; and a shard was read from all of its replicas although the data was not split between them, when parallel replicas were not applicable to the query or when parallel_replicas_mode requested custom-key filtering without parallel_replicas_custom_key. #112942 (Nikita Taranov).
• Fixed CREATE TABLE ... AS SELECT becoming unkillable when its populating SELECT failed and database_atomic_wait_for_drop_and_detach_synchronously was enabled: the internal cleanup DROP of the temporary table waited for the background drop queue and ignored both KILL QUERY and max_execution_time. #113504 (Groene AI).
• Fixed ALTER TABLE ... EXECUTE remove_orphan_files deleting a live metadata/version-hint.text file of an Iceberg table. Its reachable-set entry was built by string concatenation assuming a trailing slash on a value that never has one, so the real hint file was never recognized as reachable and was removed once older than the older_than threshold. With iceberg_use_version_hint = 1, the table then failed to read at all. #113548 (Groene AI).
• Fixed a parameterized view losing its database when a query referencing it without qualification was sent to a shard, so the shard resolved the name against its own default database. Depending on what that database contained, the query either failed with UNKNOWN_FUNCTION or silently returned rows from a different view. Affects IN, JOIN, and remote through a Distributed table. #113550 (Groene AI).
• Fixed a LOGICAL_ERROR when a query filters system.databases on the name column with a subquery inside indexHint, for example SELECT count() FROM system.databases WHERE indexHint(name IN (SELECT 'foo')). #113556 (Groene AI).
• Fixed an endless loop in background merges and OPTIMIZE when merge_max_block_size_bytes was set below the byteSize() of an empty column. An empty LowCardinality column still carries its dictionary, so the byte limit could be reached before a single row was merged; the merge then produced empty blocks forever, burning 100% of a CPU core and never releasing its background-pool slot. #113585 (Groene AI).
• Fixed a query over a Variant column ignoring max_execution_time and KILL QUERY. Calling a function with several Variant arguments resolved the function once per alternative of every argument, and neither of the two loops doing that work checked for cancellation, so such a query could not be stopped until it finished. #113612 (Groene AI).
• Fixed Unknown compression method for the auto and none compression hints when written in any letter case other than all-lowercase, for example file('data.csv.gz', 'CSV', 'x String', 'AUTO'). Codec names such as GZIP were already accepted case-insensitively; the two special hints now behave the same way. #113631 (Groene AI).
• Fix constant folding under the new analyzer changing the type of a constant Geometry value. Geometry is a named Variant whose alternatives can share the same storage layout (LineString/Ring, MultiLineString/Polygon, and empties), so two folded constants that differed only in their alternative could collapse into one, e.g. MULTIPOLYGON EMPTY became LINESTRING(). #110709 (Groene AI).
• Fix an exception (Cannot calculate columns sizes when columns or checksums are not initialized) during a merge of a single-column MergeTree table whose only column has a fully expiring column TTL, with min_bytes_for_wide_part = 1. #111169 (Groene AI).
• Failing to start a background schedule pool (for example the iceberg pool under global thread-pool exhaustion) no longer aborts the server: the error is now a recoverable exception. Reading system.background_schedule_pool also no longer creates those pools as a side effect. #111209 (Groene AI).
• Fix an out-of-bounds read in the native protocol on corrupted index data. #112331 (Pavel Kruglov).
• Fix a deadlock that permanently hangs a Filesystem or HDFS database when two queries resolve the same table name at the same time. Every later query against that database blocks forever and cannot be interrupted, so the server needs a restart to recover. #112495 (Groene AI).
• Fix max_execution_time and KILL QUERY being silently ineffective for a distributed query that is establishing connections to replicas. Such a query now stops at the next connection attempt instead of working through every remaining replica and retry. #112689 (Groene AI).
• Fix DROP TABLE and SYSTEM STOP VIEW hanging when a refreshable materialized view is blocked while planning its refresh query. #113188 (Nikolay Degterinsky).
• Fixed CREATE TABLE with ENGINE = Distributed(...) or ENGINE = Buffer(...) and no column list revealing the structure of a referenced table the creating user is not allowed to see. For a local CREATE, the structure is now inferred under the user’s own context, so SHOW COLUMNS on the referenced table is required. A CREATE replayed from the DDL queue (ON CLUSTER, or inside a Replicated database) is not covered. #113220 (Groene AI). #113372 (Groene AI).
• Fixed reading the internal _temporary_and_external_tables database through the merge table function and the Merge table engine, which allowed one session to read the temporary tables of other sessions and other users. A database regexp now skips that database, and naming it explicitly is denied, the same way as direct access to it is. #113224 (Alexey Milovidov).
• Fixed countSubstrings, countSubstringsCaseInsensitive, countSubstringsCaseInsensitiveUTF8, countMatches, and countMatchesCaseInsensitive ignoring KILL QUERY and max_execution_time: a call over many rows, many matches, or one large value could run for minutes after the query was cancelled. #113369 (Groene AI). #113886 (Groene AI).
• Fixed a mutation of a Compact part producing a different result depending on whether the part’s serialization info was still in memory or had been reloaded from serialization.json, when the table uses serialization_info_version = 'basic'. On ReplicatedMergeTree, this made two replicas holding a byte-identical source part write different mutated parts, and the mutation failed with CHECKSUM_DOESNT_MATCH. Closes #113500. #113588 (Groene AI).
• Fixed made_current_at in system.iceberg_history reporting the earliest time a snapshot became current instead of the latest one, when the same snapshot-id appears more than once in the Iceberg snapshot-log (as a rollback leaves it). #113598 (Groene AI).
• Fix wrong values for virtual columns such as _table or _database selected from the right side of a JOIN against a key-value source such as EmbeddedRocksDB, KeeperMap, Redis, or a dictionary. The query could return a data-column value instead, or fail with a LOGICAL_ERROR exception. #113698 (Groene AI).
• Fix EXPLAIN ANALYZE over a Merge table with several children, a bare-column PREWHERE, and a separate WHERE failing with Required output position N is out of range for pass-through inputs. Closes #113621. #113725 (Groene AI).
• Fixed clickhouse-disks move of a directory failing with DIRECTORY_ALREADY_EXISTS on disks with metadata_type = plain_rewritable (for example s3_plain_rewritable), which also left an empty destination directory behind. #113727 (Groene AI).
• Fix Merge table reads failing with No set is registered for key ... when a child declares an ALIAS column containing IN and the children disagree about that column’s default. #113757 (Groene AI).
• Fixed UNKNOWN_IDENTIFIER with enable_analyzer = 0 when a WHERE refers to a column produced by untuple inside a subquery, for example SELECT t.keys FROM (SELECT untuple(arrayJoin(m)) AS t FROM ...) WHERE t.values > 0. The enable_optimize_predicate_expression rewrite pushed such a predicate into the subquery as a HAVING, where the name does not yet exist. #113760 (Groene AI).
• Fix ALTER TABLE ... MODIFY COLUMN failing with Cannot specify codec for column type ALIAS when one statement both turns an ALIAS column into a physical one (DEFAULT or MATERIALIZED) and specifies a CODEC. Closes #113693. #113853 (Groene AI).
• Fix ORDER BY ALL and GROUP BY ALL over a joined Merge table under the old analyzer (allow_experimental_analyzer = 0). The query no longer fails with an internal exception, and debug and sanitizer builds no longer abort. #109139 (Groene AI).
• Fix a sporadic Cannot parse string ... as UInt64 ... While executing ValuesBlockInputFormat exception that could occur when inserting values whose expressions contain a numeric cast, such as initializeAggregation('sumState', 0::UInt64). #110637 (Groene AI).
• Fixed the periodic local disk check reporting failures for a healthy disk. When a server runs an embedded Keeper alongside a local disk, two check threads could probe the same file, transiently log FILE_DOESNT_EXIST, CANNOT_READ_ALL_DATA, or CANNOT_UNLINK, and temporarily count the disk in the BrokenDisks metric. #110773 (Groene AI).
• Fixed wrong results and a logical error Cannot serialize FutureSetFromSubquery with no query plan for queries over Merge tables and the merge table function when plan-based parallel replicas were enabled. #112849 (Alexey Milovidov).
• A table in the URL database engine no longer discloses whether a local file exists to a user without the read source grant. Both EXISTS TABLE, which requires only the SHOW TABLES privilege, and table-name resolution now wait for the grant check before probing the filesystem. #113029 (Alexey Milovidov).
• Fixed a TTL expression becoming unusable when constant folding pruned the columns it refers to, for example TTL d + INTERVAL 1 DAY DELETE WHERE isNull(x) over a non-Nullable x, which made every INSERT into the table fail with Missing columns. Also fixed the server failing to start when a table had a TTL expression over a Variant column created with variant_throw_on_type_mismatch = 0. #113042 (Alexey Milovidov).
• Fixed cancellation of postgresql table function and PostgreSQL table engine reads. Cancelling while the read was still starting up could race on the transaction pointer and be lost, leaving the query running until PostgreSQL finished it. #113150 (Groene AI).
• UNDROP TABLE can now be interrupted by KILL QUERY while it waits for running queries to release the dropped table. #113263 (Alexey Milovidov).
• Fix a memory leak of the whole XML parser when a Poco XML handler throws mid-parse, for example on Poco::XML::NamePool overflow while parsing a document with many unique element names. #113348 (Alexey Milovidov).
• Fixed writes to the HDFS disk: they no longer fail with Parent directory doesn't exist when the generated object key contains nested directory prefixes, because the missing parent directories are now created. #113396 (Alexey Milovidov).
• Fix wrong results and, in builds with assertions enabled, an aborted assertion for a JOIN onto a dictionary whose ON clause has a non-equi condition over both tables, such as ON (t.key = d.key) AND (t.a * 10 < d.a), when join_use_nulls is enabled. #113534 (Alexey Milovidov).
• Fix a query hanging for up to receive_timeout seconds when it finishes very quickly on the initiator. #113551 (Nikita Taranov).
• Fix UPDATE and lightweight DELETE reading the wrong tables when the database of the updated table differs from the session database. Unqualified table names and the merge table function are now resolved against the database of the updated table, so the statement no longer silently updates or deletes rows selected from a different table. #113748 (Groene AI).
• Fix a spurious TOO_MANY_PARTS rejection of a parallel INSERT: when max_insert_threads parallelizes a plain INSERT, the query no longer counts parts committed by its own sibling sinks. Closes #114015. #114016 (Alexey Milovidov).
• Fixed NO_SUCH_COLUMN_IN_TABLE when a query with parallel replicas selects an ALIAS column of a table shipped as a GLOBAL JOIN temporary table, MULTIPLE_EXPRESSIONS_FOR_ALIAS when both sides of a distributed JOIN declare an ALIAS column with the same name, and UNKNOWN_IDENTIFIER when a Distributed table declares an ALIAS column over another ALIAS column, including as a JOIN USING key. #107700 (Yakov Olkhovskiy).
• Fixed reading a CSV Tuple whose leading element is NULL. With input_format_csv_deserialize_separate_columns_into_tuple and input_format_null_as_default enabled, a leading \N in a row such as \N,1 was taken for the whole column being NULL, so the row was replaced by a default and the remaining element cells caused a parse error. The leading NULL is now the value of the first element, with input_format_null_as_default applying to that element. A row that gives a single \N field for a whole bare Tuple is now short by the remaining elements and is rejected; set input_format_csv_deserialize_separate_columns_into_tuple = 0 to read such a field as the whole column again. #109744 (Groene AI).
• Fix a rare Logical error: 'No more packets are available.' for a valid distributed query with a LIMIT. In debug and sanitizer builds it also aborted the server. #110334 (Groene AI).
• Fixed a logical error (!part.empty()) when a Distributed table defined with an empty remote database name is used in a JOIN with distributed_product_mode = 'local'. #110677 (Groene AI).
• Fix resolution of fully qualified column names such as db.table.column in the analyzer for tables that have the same name as their database. Closes #82084. #110976 (Alexey Milovidov).
• Fixed hash joins losing their post-build optimizations (conversion to a fixed hash table, the shared runtime filter, right-table reranging) whenever an automatic spill-to-disk threshold was configured, which has been the default since 26.5. #111972 (Alexey Milovidov).
• Fixed a bug where ALTER TABLE ... MODIFY COLUMN that changes a JSON type hint (with allow_experimental_json_lazy_type_hints enabled) on a column whose subcolumn is used in the primary/sorting key, partition key, or an explicit skip index could corrupt the on-disk key or index. Ordinary reads masked it via the cached structure, but after it was reloaded (for example on server restart) queries failed with CANNOT_READ_ALL_DATA or returned wrong results. Such ALTERs are now rejected whenever they change the on-disk type of a referenced JSON subcolumn. Type-hint changes on columns outside those structures, or changes that leave the on-disk type of the referenced subcolumns unchanged (for example adding an unrelated typed path), remain metadata-only. #112302 (Pavel Kruglov).
• Fix wrong results when query_plan_join_swap_table swaps an ANY JOIN while join_any_take_last_row = 1: ClickHouse now keeps the original join side order so ANY JOIN still returns the last matching row. Closes #111645. #112458 (Vladimir Cherkasov).
• Fixed listObjects on a prefixed-endpoint AzureBlobStorage disk returning wrong blob names on every page after the first, because pagination bypassed the client wrapper that strips the endpoint prefix. Pagination now re-enters the wrapper via the continuation token so every page’s blob names are correct. #112872 (Arsen Muk).
• Fixed unbounded memory allocation when reading a Parquet file whose Thrift metadata declares a container or string length larger than the metadata itself. Such a length is now rejected instead of being allocated for. #113212 (Groene AI).
• Fixed Cannot insert element into Set (NOT_IMPLEMENTED) when read_in_order_use_virtual_row is enabled and a MergeTree read-in-order query combines WITH FILL ... INTERPOLATE with an IN or has filter. #113264 (Groene AI).
• Fix optimize_read_in_order being silently skipped for some ORDER BY queries, including queries that use a virtual row or toString/CAST(..., 'Nullable(String)') over a String sorting key. Also fix wrong results for filters or ORDER BY on toString of Time, Time64, or DateTime values in time zones with daylight saving time. #113291 (Vladimir Cherkasov).
• Reject invalid parallel_replicas_custom_key expressions such as tuple() or (x, y) with ILLEGAL_TYPE_OF_COLUMN_FOR_FILTER instead of reading past the end of the custom-key type list. #113359 (Nikita Taranov).
• Fix a server restart failure after the server is killed during DROP TABLE of a table that uses a named collection: a later DROP NAMED COLLECTION could succeed, leaving table metadata that references a missing named collection and making the next startup fail with NAMED_COLLECTION_DOESNT_EXIST. #113524 (Alexey Milovidov).
• Fixes two cases where an AWS Glue DataLakeCatalog database reported a table as absent instead of reporting an error. SHOW TABLES and system.tables no longer fail with DATALAKE_DATABASE_ERROR (“Exception calling GetTables … not found”) when a Glue database is absent while its tables are listed, for example because an unrelated database in the same AWS account was dropped concurrently; such a database now contributes no tables instead of aborting the listing. EXISTS TABLE no longer returns 0 for a table that exists when the underlying GetTable call fails for a reason other than the table being absent, such as an authentication, DNS, or 5xx failure. Every other Glue error still propagates in both paths. #113591 (Groene AI).
• Fixed a Keeper durability bug where a crash while persisting the Raft state file (storing term and voted_for) could leave the node with no valid state and no backup. On restart the node could reset to term = 0 and grant a vote it had already granted in that term, which could cause two leaders and loss of committed data. The previous state is now kept until the new state file is fully written and synced. #113654 (Alexey Bakharew).
• Fix a data race when a MergeTree table is destroyed while its background statistics-refresh, parts-refresh, or outdated/unexpected-parts-loading tasks are still running. #113722 (Alexey Milovidov).
• Fix ORDER BY ALL in a WINDOW VIEW, and UNKNOWN_IDENTIFIER errors for INTERPOLATE, WINDOW, and LIMIT BY in a query over a Merge or Distributed table with a join, with the old analyzer. #113759 (Alexey Milovidov).
• Fix a memory leak in the experimental SZ3 compression codec: when compression of poorly compressible or corrupted data failed with an exception, temporary scratch buffers were left allocated. #113916 (Alexey Milovidov).
• Fix quadratic complexity of parsing the address argument of the remote, mysql, and similar table functions: a very long address (for example a megabyte-sized zero-padded FixedString) made the query hang during analysis, where it could not be cancelled. #113918 (Alexey Milovidov).
• Fixed OPTIMIZE TABLE on an Iceberg table failing with Unsupported snapshot's operation type OVERWRITE when the table’s history contained a row delete that removed more than one row per position-delete file. Closes #113745. #113964 (Groene AI).
• Do not replicate ALTER TABLE ATTACH PARTITION FROM for replicated db. #94835 (MikhailBurdukov).
• Invalid Azure upload-size settings (for example azure_min_upload_part_size = 0 or azure_max_blocks_in_multipart_upload = 0) are now rejected early with a clear error instead of triggering an internal exception (second_size > 0) deep in the write path. This applies to both query settings and Azure disk configuration. #103394 (Statxc).
• Fixed incorrect SQL literal escaping in StorageSQLite and sqlite() table function when pushing WHERE predicates to SQLite: single quotes and control characters (\n, \r, \t, \) were escaped with backslashes, which SQLite does not interpret, causing syntax errors or wrong query results. Also fixed the escaping of string literals pushed down to PostgreSQL: strings nested inside IN lists kept ClickHouse escaping, and control characters were sent as backslash sequences that PostgreSQL reads back as different bytes. #104217 (Shaohua Wang).
• Fix NOT_FOUND_COLUMN_IN_BLOCK exception in INNER JOIN ... ON arrayJoin(...) = ... queries when query_plan_convert_join_to_in is enabled and the SELECT references the source array column. #104809 (Shaohua Wang).
• Fix IN and NOT IN expressions with non-constant right-hand side operands referencing columns from the current row, and align new analyzer tuple right-hand side handling with existing ClickHouse IN semantics. Under the old analyzer, a bare source column as the right-hand side (x IN (arr)) still resolves as a table name and stays out of scope. #104993 (Yue Ni).
• Fixed the error Cannot add column ...: column with this name already exists on INSERT SELECT from a table function such as file, s3 or input, when the same source column is selected more than once. #105982 (etienne).
• Reject max_replication_lag_to_enqueue = 0 on Replicated databases. The value made the post-recovery unsynced check trivially true and aborted the server with LOGICAL_ERROR in debug and sanitizer builds. 0 is now rejected at parse time with BAD_ARGUMENTS from every source (CREATE DATABASE ... SETTINGS, <database_replicated> server-config block, ATTACH replay, and upgrade-time replay of existing metadata). The smallest valid value is 1. Closes #98823. #106006 (Groene AI).
• Fix Logical error: 'Pipeline stuck' in queries that use enable_sharding_aggregator = 1 together with a UNION ALL and max_streams_for_union_step smaller than the pipeline width. BufferedShardByHashTransform now finishes empty-queue output ports as soon as its input is exhausted, and keeps pulling input when a demanded empty port has nothing drainable even if a sibling queue hit the back-pressure cap. #106251 (Groene AI).
• Fixed a signed integer overflow when a wait-timeout setting, such as interactive_delay, was set to a huge value: the wait could time out immediately instead of waiting, and the server could abort under the Undefined Behavior Sanitizer. #106961 (Groene AI).
• Exporting Time/Time64 values outside a valid time-of-day (negative or >= 24h) to the Arrow format is now rejected with a clear error instead of writing invalid Arrow time32/time64 data. #107179 (Shaohua Wang).
• Fix wrong results for GROUP BY on a Distributed table over duplicate ALIAS columns that expand to the same expression (e.g. two columns both defined as toString(x)) under two-level / memory-efficient distributed aggregation. The duplicate keys were collapsed into one on the shards but merged by the full key set on the initiator, so under two-level aggregation equal groups from different shards landed in different buckets and were not merged, returning a group once per shard with a split count(). #108855 (Groene AI).
• Fixed a spurious Inconsistent AST formatting error for ALTER TABLE ... MOVE PART ... TO SHARD '<path>'. The query formatter dropped the SHARD keyword, so the reformatted query could not be parsed back, raising a LOGICAL_ERROR (a handled exception in release builds, an abort in debug and sanitizer builds). #109060 (Groene AI).
• Rejects non-monotonic Array offsets on every read path that decodes them as absolute values, instead of only in the Native format. A corrupt set skip index granule previously passed unchecked and was then read out of bounds. #109212 (Groene AI).
• JSONExtract now honours cast_string_to_date_time_mode when converting string JSON values to DateTime/DateTime64, consistently with CAST. Closes #109126. #109252 (Utkal Singh).
• Fix possible logical error “Unexpected substream … for column …” during bump of compatibility setting. #109496 (Pavel Kruglov).
• set missing coordination component in system.detached_parts size calculation. #109553 (Den Kalantaevskii).
• Fixed the block-wait timeout of streaming inserts (input_format_max_block_wait_ms) potentially never expiring while the query profiler is active: the file-descriptor poll restarted with the full timeout after every profiler signal, so a stalled input source could delay the partial-block flush indefinitely instead of flushing when the wait limit is reached. #109602 (Shaohua Wang).
• Constant folding during query planning no longer executes a function whose argument type differs from the type the function was resolved for during analysis. Such a mismatch used to raise a LOGICAL_ERROR while the output header was being computed, which is an assertion failure in debug and sanitizer builds. Folding is now skipped instead, and a genuinely invalid query still reports a recoverable ILLEGAL_TYPE_OF_ARGUMENT error from the normal execution path. #109747 (Groene AI).
• Fixes a mutation that references the _sample_factor virtual column failing mid-execution with Unexpected const virtual column: _sample_factor and leaving the mutation stuck. A mutation now reads it as 1, the same value a query without SAMPLE returns. #109813 (Groene AI).
• Fix three issues with use_constant_folding_in_index_analysis: a logical error Invalid partition key size that could abort a SELECT using a normal projection when the table sets part_minmax_index_columns = 'with_block_number_offset'; wrong results (silently dropped rows) for a filter on a modulo partition key such as PARTITION BY id % 200; and a crash (data race) when a text index with a sparseGrams tokenizer is queried with a LIKE predicate over many partitions with max_threads > 1. #109896 (Groene AI).
• Fix a data race (and rare crash) when a parsing/analysis error is reported while reading a file/object storage table. A parse error raised during lazy KeyCondition initialization was shared as a single exception object across reader threads, and two threads could concurrently append (in file/uri ...) to its message, racing on the exception message string. #109911 (Groene AI).
• Fix propagation of query settings in accurateCastOrDefault, and preserve source NULLs encoded by Dynamic and Variant. Closes #109943. #109946 (Pavel Kruglov). #114912 (Alexey Milovidov).
• Fixed incorrect results from subBitmap, bitmapSubsetInRange and bitmapSubsetLimit for bitmaps still held in the small representation, including UInt64 and Int64 element values above 2^32, which were truncated to 32 bits. Comparisons in bitmap functions over signed element types now consistently use the unsigned value of the element type, so in an Int8 bitmap the element -1 is compared as 255 instead of as the sign-extended 4294967295; this also fixes bitmapMin, bitmapMax, bitmapContains and bitmapTransform on bitmaps that have grown past the small representation. Queries that passed sign-extended thresholds have to be adjusted: over an Int8 bitmap, bitmapSubsetInRange(bm, 4294967168, 4294967296) becomes bitmapSubsetInRange(bm, 128, 256). Also fixed groupNumericIndexedVector returning different results for Int8 and Int16 index columns than for wider index types, and numericIndexedVectorGetValue returning 0 for negative indexes. Corrected the bitmap function documentation, including the subset functions that were described as using 1-based indexing and the signed bitmapBuild / bitmapToArray support. #110072 (Rami Darwiche).
• Fix optimize_aggregation_in_order ignoring query cancellation. AggregatingInOrderTransform now checks for cancellation while aggregating a chunk, so a query stopped by KILL QUERY or by max_execution_time (in the default timeout_overflow_mode = 'throw') stops promptly instead of running the whole chunk to completion. #110104 (Groene AI).
• Fixed a race between KILL TRANSACTION and ALTER TABLE ... UPDATE/DELETE executed inside a transaction: a rollback landing between the mutation’s registration in the transaction and its registration in the table left an orphaned mutation entry, which raised a logical error (Cannot find transaction ... that has started mutation ...) in background jobs and blocked all subsequent mutations of the affected parts until server restart. #110226 (Shaohua Wang).
• Fix a LOGICAL_ERROR (Left and right columns have same names) in the join order optimizer that could be triggered by a comma join over a multi-table view when query_plan_merge_expression_into_join is enabled. #110227 (Groene AI).
• Join-order cardinality estimation now composes column statistics over the parts surviving partition/PK pruning instead of all active parts. Previously a query pruned to a small partition was planned against table-wide statistics (observed 2500x row overestimation flipping the hash-join build side), and disabling statistics paradoxically produced a better plan. Also fixed: lazy FINAL was silently disabled for MergeTree relations under a JOIN because the join-order optimizer’s memoized index-analysis result was mistaken for an applied projection. #110283 (Sergey Kuznetsov).
• Fixed wrong results from correlated scalar subqueries when the outer query has duplicate values for the correlated column and the subquery is decorrelated via the CROSS JOIN path. Previously aggregates such as count() and sum() were multiplied by the duplicate count. This affects the experimental allow_experimental_correlated_subqueries feature. #110311 (Groene AI).
• Fix a logical error (Function node with name '...' is not resolved as ordinary function) that could occur with optimize_or_like_chain enabled when an aggregate or window function appeared on the left-hand side of a LIKE/ILIKE/match inside an OR chain. #110641 (Groene AI).
• Fix a TTL ... DELETE WHERE <cond> never removing rows that started matching <cond> because of an ALTER TABLE ... UPDATE on a column referenced in the WHERE. The mutation now recalculates the row TTL, so the newly matching expired rows are deleted instead of being silently retained. #110648 (Groene AI).
• Fix an accounting leak in memory tracking: when an allocation was rejected by a parent memory tracker (for example, the server-wide memory limit), the thread, query, and user level trackers kept the rejected amount. The error accumulated over time, inflating MemoryTracking and per-query/per-user memory usage, and could lead to spurious MEMORY_LIMIT_EXCEEDED errors. #110784 (Seva Potapov).
• Fixed non-canonical WKB serialization of an empty polygon: wkb(POLYGON EMPTY) now emits numRings = 0 instead of a spurious single zero-point ring, so a WKB round-trip of a standard empty polygon is identity-preserving. #110796 (Groene AI).
• Fixes a case when using CTE column aliases with INTERSECT or EXCEPT queries, e.g. WITH t (a, b) AS (SELECT 1, 2 INTERSECT SELECT 1, 2) SELECT a, b FROM t, the query was failing with UNKNOWN_IDENTIFIER error. Closes #104770. #110885 (Yarik Briukhovetskyi).
• Fixed a std::length_error (LOGICAL_ERROR) when reading a MergeTree table with a pathological max_streams_for_merge_tree_reading value (near the maximum of UInt64) together with allow_asynchronous_read_from_io_pool_for_merge_tree = 1. The setting is now clamped to the same ceiling as max_threads. #110936 (Groene AI).
• Fixes a crash (null pointer dereference during query analysis) when getSubcolumn is called on a JSON argument with an internal type-hint subcolumn name, that is one starting with a backtick type marker. #110940 (Groene AI).
• Fixed Nested type LineString cannot be inside Nullable type (ILLEGAL_TYPE_OF_ARGUMENT) when reading a nullable MySQL spatial column (LINESTRING, POLYGON, MULTILINESTRING, MULTIPOLYGON, MULTIPOINT, or the generic GEOMETRY) through the mysql table function or the MySQL table/database engines. Such columns now fall back to Nullable(String) holding the value as MySQL returns it (a 4-byte SRID prefix followed by the WKB payload). POINT keeps mapping to Nullable(Point). #110943 (Groene AI).
• Fixes Native serialization of SimpleAggregateFunction wrappers containing versioned aggregate states for older peers. #110997 (Groene AI).
• Fixed a LOGICAL_ERROR: ReadBuffer is canceled. Can't read from it. that could occur in a grace_hash join when reading spilled blocks failed mid-read while multiple threads shared the reader. #111056 (Groene AI).
• Each SSH key now counts individually toward the max_authentication_methods_per_user limit, so the limit applies consistently regardless of syntax. Previously, multiple keys in a single ssh_key method counted as one. #111181 (Yecine Megdiche).
• Fix a LOGICAL_ERROR (Reading from materialized CTE '...' before its materialization completed - DelayedPortsProcessor gate is missing in the query plan) when a reused materialized CTE (enable_materialized_cte) is read on a branch whose WITH TOTALS or extremes output is discarded, including outer aggregation, INSERT ... SELECT, set operations and EXPLAIN ANALYZE. #111194 (Groene AI).
• Fixed map[key] returning the value of a later duplicate key instead of the first one. When a Map column contained duplicate keys, the result of map[key] could depend on preceding rows in the same block and on the optimize_functions_to_subcolumns setting, so a SELECT and a WHERE over the same row could disagree. #111246 (Groene AI).
• Fix a logical error !rhs_literal->getValue().isNull() in the logical-expression optimizer when an AND chain compares a column against a NULL-valued constant of a non-Nullable type (for example a NULL-valued Variant) with use_variant_default_implementation_for_comparisons = 0. #111292 (Groene AI).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK / THERE_IS_NO_COLUMN errors (and, on some releases, silently wrong results) for a RIGHT JOIN when parallel_replicas_min_number_of_rows_per_replica is set and a left-table column is projected. #111332 (Groene AI).
• RESTORE now fsyncs the restored part files when the destination table has fsync_after_insert enabled, so a restored part is as durable against power loss as an inserted one. Previously restored files were only written and closed (never fsynced), so a power loss right after RESTORE returned RESTORED could leave the parts torn and the table empty. #111378 (Groene AI).
• Fix a rare Keeper data-loss window on local snapshot disks: the snapshot write path now fsyncs the snapshot directory after removing the tmp_ marker, so a power loss can no longer drop the marker unlink while keeping the snapshot content. Previously such a crash made Keeper treat a good, fully-synced snapshot as incomplete and delete it on restart, losing acknowledged writes even with force_sync=true. Snapshot disks backed by object storage are not covered. #111397 (Groene AI).
• Fix an out-of-bounds read (release) / logical error (debug) in the lazy materialization optimization, and a logical error in Pipe::unitePipes for Merge tables, when the direct-read-from-text-index optimization leaves an unused index virtual column in the read header of a MergeTree table with multiple text indexes. #111467 (Groene AI).
• Fixed the _time virtual column of the url/urlCluster table functions always being NULL on musl-based builds, caused by non-portable parsing of the HTTP Last-Modified header. #111619 (Konstantin Bogdanov).
• Fixed a LOGICAL_ERROR (Cannot create part with type Compact and storage type Full because table does not support polymorphic parts) raised when a MergeTree table loads a pre-existing Compact part while its current settings disable polymorphic parts (non-adaptive granularity). This could happen, for example, when a read-only table shares an on-disk path with another table and reloads its parts via SYSTEM RESTART DISK or on server startup. In debug and sanitizer builds the exception aborts the server. #111626 (Groene AI).
• Fixed CREATE TABLE ... ENGINE = KeeperMap failing with Coordination error: No node, path .../metadata/drop_lock_version when another table on the same Keeper path is dropped concurrently. The create now treats the vanished metadata subtree as a completed cleanup and retries, as it already does for the same situation one statement later. #111966 (Groene AI).
• Fix a race between a ReplicatedMergeTree mutation and the removal of old temporary directories that could publish a mutated part with some of its files missing, making reads of that part and all its later mutations fail. #111992 (Alexey Milovidov).
• Fix quantileDeterministic, quantilesDeterministic and medianDeterministic returning different results for the same data on every run when the query uses parallel replicas, a distributed table, or external aggregation, or when the states are stored in a table (including inside a Dynamic value). States written before this version under the unversioned type spelling keep the previous behavior, since their layout cannot change. #112052 (Alexey Milovidov).
• Fixes serialize_string_in_memory_with_zero_byte being dropped from the serialized query plan. With serialize_query_plan = 1, a cluster running with the setting disabled could diverge on the in-memory string representation between the initiator and the node executing the deserialized plan, because the setting was read on deserialization but never written on serialization and therefore always resolved to its default on the executing node. #112131 (Groene AI).
• Fixed a stack frame being displayed as ? instead of its function name in fatal log messages and in the trace_full column of system.crash_log. It affected frames of ordinary functions that have code attributed to a libc++ __functional header, which is common in release builds. #112152 (Alexey Milovidov).
• Fixes silently missing rows when a query filters on intDiv(constant, key) or divide(constant, key). Primary-key and partition analysis reports the wrong direction of monotonicity for a constant divided by a key column, so matching granules are pruned away and the query returns fewer rows with no error. Also fixes intDiv with an unsigned constant dividend whose high bit is set, stops claiming monotonicity when a Decimal operand makes the division compute in the decimal’s native width, and stops key analysis from raising Cannot compare DB::IPv4 with long for a valid query that divides an IPv4/IPv6 constant by a key column. #112156 (Groene AI).
• Fixed DROP TABLE ... SYNC hanging forever when a query over system.replicas had been killed before its status requests were processed. #112164 (Nikolay Degterinsky).
• Fixed a server abort (SIGABRT) when a BUFFERED_V1 WebAssembly UDF returns an empty result buffer; the query now fails with a catchable WASM_ERROR and the server stays up. #112182 (Nikolai Ovchinnikov).
• Fixed a use-after-free when a JOIN with join_algorithm = 'direct' onto a dictionary was given a join key that is stored with sparse serialization. getColumnVectorData returned a reference to a temporary column, so the dictionary lookup read freed memory: release builds could return wrong results and debug or sanitizer builds aborted. #112327 (Groene AI).
• Fixes cases where row policy was not used for MergeTree Projections during query execution. #112329 (Yarik Briukhovetskyi).
• Fixed data corruption when inserting into an Iceberg table reached through a MATERIALIZED VIEW. The data file was written against the target’s CREATE-time column types instead of its live Iceberg schema, so a DateTime value was stored as a bare Avro int and read back as microseconds, and with DateTime64(0) or DateTime64(9) the file could not be read back at all. As a consequence, INSERT INTO <materialized view> against a target whose Iceberg schema is wider than the view’s declared columns now raises TYPE_MISMATCH instead of writing a corrupt row. #112359 (Groene AI).
• Keeper no longer refuses a Set request with empty data when it is over max_memory_usage_soft_limit. Such a request cannot increase the amount of stored data, and refusing it prevented clients from re-establishing their session, which could keep tables read-only for the whole duration of a Keeper memory event. #112386 (Shaohua Wang).
• Fixed a Keeper follower dropping out of quorum for a long time, causing mass Session expired errors, after it successfully installed a large snapshot: the leader discarded the follower’s acknowledgement whenever applying the snapshot took longer than the snapshot-install timeout, and then walked that follower’s log index backwards one entry per round trip. Also added the Keeper setting nuraft_snapshot_sync_ctx_timeout_ms to control that timeout, which previously could only be changed through raft_limits_response_limit and therefore not without also slowing down dead-peer detection. #112405 (Shaohua Wang).
• Fix data races in profile events and memory tracker. #112466 (Miсhael Stetsyuk).
• Security related bugfix. #112479 (Kseniia Sumarokova).
• Fixes reading a stale secondary (skip) index after an ALTER TABLE ... MODIFY COLUMN type change whose mutation never ran, for example because KILL MUTATION removed it: the granules on disk were written with the old type but decoded with the new one, raising LOGICAL_ERROR, requesting multi-exabyte allocations, or silently returning a wrong result. Also fixes a wrong result from an index over an expression whose meaning changes while every stored byte stays identical, so no mutation is created at all: a MODIFY COLUMN altering only a DateTime timezone, or only a custom type name such as UInt8 to Bool. Such an index is now skipped for the affected part. Closes #112213. #112484 (Groene AI).
• Fixed a crash when reading a Parquet file with inconsistent bloom filter metadata. Such files could also silently return fewer rows than they should. #112498 (Shaohua Wang).
• Fix AsynchronousBoundedReadBuffer’s readBigAt data race. Closes #109678. #112573 (Kseniia Sumarokova).
• Fixes a column alias list such as (SELECT 1) AS t(x) or WITH t(x) AS (SELECT 1) not being visible inside the view table function, where a reference to t.x failed with UNKNOWN_IDENTIFIER although the same construct works outside view. #112690 (Groene AI).
• Fixed the logical error SortingAggregatedTransform already got bucket with number N that could occur in multi-layer distributed GROUP BY queries using two-level aggregation. #112794 (Nikita Taranov).
• Fixes a NATS table with nats_stream set silently consuming nothing after the NATS server is restarted. The JetStream subscription is now re-established automatically instead of requiring DETACH TABLE and ATTACH TABLE. #112828 (Groene AI).
• Fixed a transient Azure Blob Storage error on the read or merge path — an RBAC permission-propagation 403 Forbidden, a credential authentication failure, or a connect timeout surfaced as 408 — being misclassified as non-retryable and reported as POTENTIALLY_BROKEN_DATA_PART (Code 740), raising a false critical alert for a healthy part. These transient failures are now treated as retryable, and a genuinely persistent one surfaces as a plain Azure error instead of a phantom broken data part. #112871 (Arsen Muk).
• Fixed Azure batch object deletion recording no system.blob_storage_log Delete events when the batch request itself failed, leaving the whole batch unlogged. A Delete event is now recorded for each object on a batch-level failure. #112873 (Arsen Muk).
• Fix the per-part minmax index over _block_number and _block_offset being lost after a table reload when the part was produced by a mutation that does not rewrite the whole part. #112878 (Alexey Milovidov).
• Fixes a logical error and wrong results when a non equi JOIN ON condition that is evaluated during the join contains arrayJoin. A hash, parallel_hash or grace_hash join now rejects such a condition with INVALID_JOIN_ON_EXPRESSION. Where the expansion depends on one side only, move it into an ARRAY JOIN in a subquery before the join; a condition whose arrayJoin argument reads columns from both sides has to be restructured. #112889 (Groene AI).
• Fix wrong results of ORDER BY when reading from a Merge table over Distributed tables: rows could be returned out of order, and rows belonging in the result could be missing when LIMIT was used. #112931 (Alexey Milovidov).
• Fixed a Bool constant on the right-hand side of IN being converted to a String left-hand side as '1'/'0' instead of the canonical 'true'/'false' (now consistent with CAST and the values table function). #113051 (Yakov Olkhovskiy).
• Fixes views column in query_log for inserts to an Alias table whose target triggers materialized views, which was always empty. #113076 (Enric Calabuig).
• Fixed LOGICAL_ERROR when reading a table with a set skip index defined on an expression and the query repeats that expression with a Nullable operand, for example WHERE intDiv(id, (SELECT 1000)) = 3 over INDEX b_set intDiv(id, 1000) TYPE set(100). A scalar subquery is enough to trigger it, since it is typed Nullable. #113244 (George Larionov).
• Fix a logical error (exception) Chunk info was not set for chunk in MergingAggregatedTransform when the setting inject_random_order_for_select_without_order_by is enabled and an aggregation query reads from a Merge table containing a Distributed child: the random ORDER BY rand() wrapper is no longer injected into queries planned only up to an intermediate stage. #113266 (Alexey Milovidov).
• Fixed a query plan optimization stall when a WHERE clause contains a large string constant with many dots and the table carries a bloom_filter, tokenbf_v1, ngrambf_v1 or text skip index. Matching a filter column name against JSONAllPaths(...) index columns enumerated every dot split of the name, which made skip-index condition building quadratic in the constant’s length. Closes #113003. #113289 (Groene AI).
• Fixed ATTACH of a Kafka table when kafka_num_consumers exceeds the limit derived from the number of CPU cores. #113390 (Nikolay Degterinsky).
• Fixed a permanent DROP TABLE hang and memory leak after a query using a materialized CTE failed. The table stayed stuck in the drop queue until the server was restarted. #113397 (Shaohua Wang).
• Fix a logical error IDENTIFIER is not a table expression when a remote-family table function with a view(...) argument is used on the left side of a JOIN ... USING whose key is a SELECT list alias, with analyzer_compatibility_join_using_top_level_identifier = 1. #113400 (Groene AI).
• Fixed reading Paimon tables partitioned by a TIMESTAMP column with a precision higher than milliseconds, which failed with scale 6 is not supported, only support scale <= 3. Closes #112768. #113401 (JIaQi Tang).
• Fixes a segfault and a LOGICAL_ERROR (No host found for exchange stream final_result__0_0) when a subquery sets distributed_plan_execute_locally in its own SETTINGS clause under make_distributed_plan. The distributed plan and the code executing it read the setting from two different contexts and could disagree, so the initiator built a self-contradictory pipeline. #113422 (Groene AI).
• Fixed reading Paimon tables that contain a nullable ARRAY or MAP column. Such a table could not be read at all, because the schema mapper wrapped the composite type in Nullable, which ClickHouse forbids, so both DESC and SELECT failed with Nested type Array(Nullable(Int32)) cannot be inside Nullable type. A nullable composite column is now mapped to a non-Nullable composite type and a NULL value is read as an empty one. Closes #113337. #113450 (Groene AI).
• Fix part and granule pruning for a JOIN ON condition involving constant columns of the other side. #113484 (Vladimir Cherkasov).
• Stopped writing a part whose Map layout contradicts its own metadata, for a Map value inside a Dynamic column. Tables differing only in map_serialization_version shared one cached serialization object, so a part could get the basic Map layout while its metadata declared with_buckets, and reading it later failed with LOGICAL_ERROR: Stream ...buckets_info ... is not found. Parts an earlier version already wrote this way stay unreadable and have to be dropped. #113514 (Groene AI).
• Fixed the inferred column type depending on the order of previous queries for a source whose schema is cached. Settings that change an inferred type were missing from the schema inference cache key, so the first query’s value decided the type every later query saw: input_format_try_infer_exponent_floats, max_parser_depth, input_format_json_infer_array_of_dynamic_from_array_of_different_types, and for Parquet also input_format_parquet_local_time_as_utc, input_format_parquet_allow_geoparquet_parser, input_format_parquet_skip_columns_with_unsupported_types_in_schema_inference and schema_inference_make_json_columns_nullable. Also registers the missing cache key getter for Form and makes the Template getter use its own row format’s escaping rule. #113533 (Groene AI).
• Fix a theoretical data race in profile events when using the trace_profile_events_list setting to write stack traces of certain profile events to system.trace_log. #113553 (Miсhael Stetsyuk).
• Fixed RENAME DATABASE accepting a target name long enough to make the database’s tables impossible to drop. The table name length check now runs regardless of check_table_dependencies, and it also covers detached tables, which were never checked and so were affected even at default settings. Closes #101747. #113557 (Groene AI).
• Fixed the LOGICAL_ERROR Reading from materialized CTE '...' before its materialization completed - DelayedPortsProcessor gate is missing in the query plan raised when a Merge table with more than one child reads a materialized CTE that the outer query references. #113558 (Alexey Milovidov).
• Fixed a row policy containing a scalar subquery being evaluated only once and then reused forever after the table had been read through a Merge table. The parsed policy condition is cached and shared by all queries, and reading it through Merge rewrote the cached expression in place, which also caused a data race between concurrent queries using the same policy. #113563 (Alexey Milovidov).
• Fixed Unsupported JOIN keys of type keys256 in StorageJoin when reading a Join-engine table whose key is composite (several columns) or a single 16/32-byte value such as UUID, Int256 or FixedString(4). Such a table could be created, written and joined against, but never read back with SELECT. Closes #74362 and #77394. #113578 (Groene AI).
• In the Web UI, opening a link with run=1 no longer makes other tabs’ URLs auto-run their queries on the next page reload. #113595 (Alexey Milovidov).
• Fix a logical error comparing arrays whose element type is Nothing. #113609 (Groene AI).
• Backups now report the underlying lock-file access error instead of incorrectly reporting a concurrent backup when the destination storage cannot read the .lock file. #113637 (Kirill Shokhin).
• Fix a logical error Sending a distributed query with unknown (zero) client version when a structure-less Distributed table over a remote shard is attached or loaded from metadata written by an older server version. #113675 (Alexey Milovidov).
• Treat a NUL byte as an ordinary literal in match() patterns. #113690 (Vitaly Baranov).
• Fix theilsU over a window frame returning an arbitrary value instead of 0 when the first argument is constant within the frame. #113691 (Alexey Milovidov).
• Fix Logical error: Duplicate announcement received for replica number N when automatic parallel replicas (automatic_parallel_replicas_mode) adopts a plan whose single-node variant was sharded by query_plan_join_shard_by_pk_ranges. #113730 (Alexey Milovidov).
• Fix CANNOT_CONVERT_TYPE and an exception in GroupingAggregatedTransform when reading a Merge table with one child being a Distributed table under custom-key parallel replicas (parallel_replicas_mode = 'custom_key_sampling' or 'custom_key_range'). Closes #113741. #113742 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR when a SELECT used the STREAM modifier on a named table together with parallel replicas in the read-tasks mode. With a materialized CTE used as an IN set in PREWHERE, the server raised Reading from materialized CTE ... DelayedPortsProcessor gate is missing in the query plan (a server abort in debug and sanitizer builds). A table carrying STREAM is now rejected when parallel replicas are requested: at enable_parallel_replicas = 2 the query fails with SUPPORT_IS_DISABLED, at 1 it runs without them, exactly as already happens for FINAL. #113754 (Groene AI).
• Fixed a server crash when joining with a constant ON expression against a right side that contains an ARRAY JOIN. Reading a lazily replicated right-side column used a freed ColumnReplicated. Closes #113859. #113901 (Groene AI).
• Fix undefined behavior and wrong results in avg over Date/DateTime/DateTime64/Time/Time64: the average is now computed exactly in integer space, fixing both the Int64-boundary overflow (UB, wrong result on x86) and Float64 precision loss above 2^53 (visible at nanosecond scale). #113912 (Alexey Milovidov).
• Fix wrong results in JOIN with join_algorithm = 'partial_merge' when the right side carries totals: the right side could be left unmerged, so LEFT joins substituted default values for real matches and INNER/RIGHT joins returned no rows at all. In the RIGHT/FULL non-joined path the same state dereferenced a null pointer, which crashed the server. #113936 (Groene AI).
• Fixed a defect in the MySQL integrations where overriding a TLS credential of a named collection with an empty ssl_ca_pem/ssl_cert_pem/ssl_key_pem value was accepted when the collection stored the credential in the contents form, silently dropping the configured CA or client certificate instead of rejecting the override. #113947 (Alexey Milovidov).
• Fixes a std::out_of_range exception (Logical error: 'std::exception. Code: 1001, type: std::out_of_range', which also aborts the server in debug and sanitizer builds) in the Parquet v3 reader when a column needed only for filter evaluation was scheduled for decoding after its output slot had already been dropped. #113956 (Groene AI).
• Fix a logical error (Unexpected return type) when a filter over a view whose column type differs from the underlying storage (e.g. information_schema.tables, where engine is Nullable(String), over system.tables, where it is String) is pushed down to the storage. #113984 (Alexey Milovidov).
• Fix ClickHouse Keeper refusing to start with CORRUPTED_DATA after a crash during cross-segment Raft log truncation (writeAt), which could leave an acknowledged log entry stranded behind stale changelog segments. #114003 (Nishant Agarwal).
• A table or database whose name is substituted into a ReplicatedMergeTree ZooKeeper path through the {database} or {table} macro is now rejected if the name would not be a single safe path component. Previously a name containing / was spliced in unescaped, so the table registered its znodes inside another table’s keeper subtree and permanently broke ALTER and OPTIMIZE there. A . or .. component, a control character and macro syntax are rejected too. Existing tables keep loading and attaching. #114006 (Groene AI).
• Fixes a memory leak with optimize_aggregation_in_order = 1 when the table sorting key is a strict prefix of the GROUP BY key and an aggregate function whose state owns heap memory is used, such as quantileDD. Server memory grows with every such query until restart. #114010 (Groene AI).
• Fixes a logical error when a distributed query plan contains a window function producing an aggregate function state, for example theilsUState(a, b) OVER (ORDER BY a) with make_distributed_plan = 1. Affects the cramersV, cramersVBiasCorrected, theilsU and contingency families. #114020 (Groene AI).
• Fix wrong results for SELECT DISTINCT with a LIMIT that has to read its whole input: LIMIT n WITH TIES dropped tying rows, exact_rows_before_limit = 1 reported a truncated rows_before_limit_at_least, and GROUP BY ... WITH TOTALS computed the totals row from a prefix of the stream. #114025 (Groene AI).
• Fixed isProbablePrime on UInt128/UInt256 ignoring max_execution_time and KILL QUERY. A block of wide values ran to completion inside a single function call with no cancellation point, so a query could run for minutes past its deadline. #114032 (Groene AI).
• Fixes a missing system.query_views_log row for a materialized view that fails while its dependencies are being collected with materialized_views_ignore_errors enabled. In debug and sanitizer builds this also aborts the server with std::out_of_range instead of logging the view. #114064 (Groene AI).
• Fixed Code: 46. DB::Exception: Unknown function exists. (UNKNOWN_FUNCTION) thrown when a PREWHERE clause contains an IN (subquery) predicate and rewrite_in_to_join = 1 (or make_distributed_plan = 1, which force-enables it) is set. The same query spelled with WHERE worked correctly. #114067 (RohithPariki).
• Fixed ORDER BY ... LIMIT returning fewer rows than requested, possibly none, when a row policy was the only filter of the query and the sort column had a minmax skip index. The top-K optimization narrowed the read before the row policy was applied. #114073 (Alexey Milovidov).
• Fixes hasAny, hasAll, has, indexOf, mapContainsKey, mapContainsValue and mapContains returning too few rows, or failing with TOO_LARGE_STRING_SIZE, when a bloom_filter index is queried with a FixedString constant. The index hashed the padded form of the constant while the function compares the unpadded one, so a matching granule was skipped. #114089 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR “Not-ready Set is passed as the second argument for function globalNullIn” and the loss of the query’s real error message, which could happen when statistics part pruning speculatively executed an IN/GLOBAL IN subquery during index analysis and that subquery failed. #114121 (Groene AI).
• Respect the allow_calculating_subcolumns_sizes_for_merge_tree_reading setting in the PREWHERE optimization. #114146 (Pavel Kruglov).
• Reject corrupted/malicious dictionary where lcp exceeds the previous token or lcp + data_size overflows, before the buffer write. #114155 (Elmi Ahmadov).
• Fix min and max on DateTime64 columns returning wrong results once the aggregate is JIT compiled and the data contains timestamps before 1970. is_signed was not specialised for DateTime64 and Time64, which derive from Decimal64, so the generated code compared tick counts as unsigned integers and a pre-1970 timestamp won max and lost min. #114168 (Groene AI).
• Fixed an exception when reading a MergeTree table by primary-key range layers when a layer produces an empty pipe. #114177 (Alexey Milovidov).
• Fixed the error code reported for some malformed Replicated-serialized columns received over the Native protocol: these are now rejected with INCORRECT_DATA instead of LOGICAL_ERROR (which aborted the server in debug and sanitizer builds). #114186 (Pavel Kruglov).
• Fixed ATTACH PARTITION FROM, REPLACE PARTITION, MOVE PARTITION TO TABLE and adding a ReplicatedMergeTree replica failing with Tables have different ..., METADATA_MISMATCH or INCOMPATIBLE_COLUMNS for tables whose definitions were written with redundant parentheses, such as PARTITION BY (a), ORDER BY (b), INDEX ix (b * c) TYPE minmax, PROJECTION p (SELECT (b) ...), CONSTRAINT c CHECK (a > 0), TTL (d + INTERVAL 10 YEAR) or DEFAULT (a + 1). #114188 (Alexey Milovidov).
• Fixed ALP, FPC and GCD omitting their type-derived parameter from the codec hash. In a Compact part the writer shares one compressed stream per distinct codec hash, so two columns of different value width under the same codec collided and shared one codec object carrying the wrong width. This rejected valid inserts with CANNOT_COMPRESS when the widths disagreed, and otherwise compressed one of the columns worse than it should have. #114189 (Groene AI).
• Fix std::out_of_range: map::at: key not found when a MergeTree key expression contains a mortonEncode or hilbertEncode call that is not itself a key column, for example ORDER BY mortonEncode(hilbertEncode(x, y), x, y) or PARTITION BY mortonEncode(x, y) % 4. #114233 (Groene AI).
• Clear plain LIMIT/OFFSET in the window-view backfill source query. #114247 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR in arrayAutocorrelation when the argument is a non-empty array whose element type is Nothing, for example SELECT arrayAutocorrelation([arrayMax([])]). Such an argument is now rejected with ILLEGAL_COLUMN. The empty array literal [] keeps returning []. #114249 (Groene AI).
• Fix a Keeper server never joining the cluster when nuraft_use_bg_thread_for_snapshot_io is enabled: the log batch sent to a joining server was silently dropped, so the membership change never completed. #114275 (Shaohua Wang).
• Fix XDG_CACHE_HOME being read from XDG_STATE_HOME. #114278 (Alexey Milovidov).
• Register Iceberg namespace in the catalog before writing table files (needed for SeaweedFS). #114285 (Azat Khuzhin).
• Fix inconsistent results when a filter on a non-key column of the right table was pushed down below an ANY INNER JOIN. #114313 (Kirill Shokhin).
• Fix bash completion with bash-completion 2.12+ (_comp_get_words). #114339 (Azat Khuzhin).
• Fixes a crash when reading a data lake table whose schema declares a column with an empty name. Malformed Iceberg metadata is rejected with ICEBERG_SPECIFICATION_VIOLATION; a schema supplied by a catalog, Delta Lake or Paimon is rejected with AMBIGUOUS_COLUMN_NAME. The check is unconditional, so an Iceberg table that merely retains an unused historical schema with an empty field name also becomes unreadable instead of aborting once that schema is read. #114394 (Groene AI).
• Fixes a bug where a client connecting to ClickHouse Keeper during a Raft leader change could be held for the whole session_timeout_ms (30 seconds by default) before its connection was rejected, instead of being rejected as soon as the in-flight request was dropped. The connecting client can now reconnect to another replica sooner. #114401 (Groene AI).
• BACKUP, RESTORE and ENGINE = Backup now authorize the backup location against the user’s SOURCES grants, the same way s3(), file() and azureBlobStorage() do. Previously a user holding BACKUP but not SOURCES could write a backup to, and read one from, any location the server could reach. Writing a backup now requires the WRITE direction on the destination’s source (WRITE ON S3, WRITE ON AZURE, WRITE ON FILE) and reading one requires READ; Disk(...) destinations stay restricted by backups.allowed_disk only. #114405 (Groene AI).
• Fixes CREATE OR REPLACE leaking an internal _tmp_replace_* table, and CREATE OR REPLACE VIEW failing with NOT_IMPLEMENTED, when ignore_drop_queries_probability is set. The DROPs it issues internally are steps of one user statement, so DROP fault injection no longer applies to them. #114420 (Groene AI).
• Fixed a LOGICAL_ERROR (Bad cast from type DB::ASTLiteral to DB::ASTPartition) and unparseable formatted SQL when an ALTER TABLE ... MOVE command deserialized from a JSON AST (dialect = clickhouse_json) paired the PART form with TO TABLE, or the PARTITION form with TO SHARD. Such payloads are now rejected with BAD_ARGUMENTS. #114429 (Groene AI).
• Do not remove columns when it’s referenced in both PREWHERE and WHERE filter conditions even if it’s rewritten into a text index virtual column. #114460 (Elmi Ahmadov).
• Fixed a Not-ready Set is passed as the second argument for function 'in' error when a mutation whose predicate contains IN (subquery) is cancelled, for example by KILL MUTATION or DETACH DATABASE, while the subquery’s set is still being built. Related: #107619. #114463 (Groene AI).
• ALTER TABLE ... DROP COLUMN of a column that is used in the sorting, primary or partition key now fails with ALTER_OF_COLUMN_IS_FORBIDDEN and an explanation, the same as ALTER TABLE ... CLEAR COLUMN does. Previously it failed with a confusing UNKNOWN_IDENTIFIER: Missing columns error coming from the recalculation of the key expressions. The same applies to dropping or clearing a whole Nested group by its common prefix when a column of the group is used in a key; previously that could silently destroy the group’s data or leave a mutation that never finishes. Dropping a column that is used only in a TTL expression now also reports which TTL expression it breaks. #114468 (Maksim Dergousov).
• Fixed reading a Tuple column whose first element is NULL in the CustomSeparated, Regexp and Template formats with the CSV escaping rule. A bare Tuple occupies one field per element there, so the leading \N field is that element and not the whole column. Previously it was consumed as the whole column, shifting the remaining element fields onto the following columns, so CustomSeparated could not read back its own output. Closes #114402. #114471 (Groene AI).
• Fixes arrayExists(x -> x = needle, arr) returning a wrong result, or failing with TOO_LARGE_STRING_SIZE, when the array element and the needle are different string types (for example a String needle against an Array(FixedString(N)) element). The optimize_rewrite_array_exists_to_has optimization, enabled by default, rewrote such a call to has, which does not compare zero-padded the way = does. #114496 (Groene AI).
• Reading an Iceberg table whose metadata describes a schema evolution that the Iceberg specification forbids no longer aborts the server. Four spec violations in IcebergSchemaProcessor were reported as LOGICAL_ERROR, which is treated as a failed assertion, or reached a fatal assertion unchecked; they now raise ICEBERG_SPECIFICATION_VIOLATION. #114518 (Groene AI).
• Fixed the Regexp input format silently discarding the unparsed rest of a matched field, which stored a truncated or fabricated value instead of reporting an error: v=-1 into UInt64 read 0, and v=2020-01-01junk into Date read 1975-07-14 under the JSON rule. Malformed fields are now rejected under the Escaped, CSV and JSON rules, matching the formats those rules are documented as behaving like. Under Raw, a matched field containing a tab is now read as a whole instead of being truncated at the tab. #114520 (Groene AI).
• Fix LOGICAL_ERROR when filtering an Iceberg column that ALTER TABLE ... MODIFY COLUMN made Nullable. Closes #85029. #114521 (Groene AI).
• CREATE TABLE and ALTER TABLE now reject lossy codec such as SZ3 on a column used in the sorting key. #114531 (Groene AI).
• Fixed a LOGICAL_ERROR when reading a Delta Lake table whose metaData.partitionColumns names a column that metaData.schemaString does not declare. Such metadata is now rejected with BAD_ARGUMENTS on the delta-kernel reader and INCORRECT_DATA on the legacy reader, whenever the table or its schema is read from a snapshot, not only when the query carries a predicate. #114538 (Groene AI).
• Reject a Variant whose ORC branches read back as one type. #114540 (Groene AI).
• Fix text index evaluation of LIKE/ILIKE operator built on Map or JSON containers. #114544 (Elmi Ahmadov).
• Support quoted label names in PromQL grouping modifiers.. #114545 (Minh Vu).
• Fixes incorrect, often sign-flipped, Time64 literals and IN-list constants produced when rescaling a lower-scale Decimal64 overflowed Int64. Such a conversion now reports DECIMAL_OVERFLOW, matching the DateTime64 branch and explicit CAST. #114546 (Groene AI).
• Accept trailing commas in PromQL grouping label lists.. #114548 (Minh Vu).
• Support quoted metric and label names in PromQL selectors. #114551 (Minh Vu).
• A query could be cancelled up to 1 ms before its max_execution_time had passed, failing with a self-contradictory Timeout exceeded: elapsed 999.672 ms, maximum: 1000 ms. #114559 (Alexey Milovidov).
• Fixed reading a table while a mutation is in progress: a MATERIALIZED column could return a stale value, and selecting one together with an updated column could fail with an error. #114561 (Shaohua Wang).
• Fixes clickhouse-local accepting a CREATE TABLE whose name is too long for the configured --default_database, which produced a table that could not be dropped. DatabaseOverlay now checks the table name length like the on-disk database it writes to. #114567 (Groene AI).
• Fixed an RBAC bypass in the Alias table engine that allowed users without privileges on the target table to reveal its schema, row count, size, and existence. #114596 (Kai Zhu).
• Fixed hash-table statistics — and with them aggregation hash-table preallocation — being silently and permanently disabled for the whole server when a query using the lazy FINAL optimization (query_plan_optimize_lazy_final) was the first aggregation to run after server startup. #114597 (Alexey Milovidov).
• Fixed a column that was renamed, dropped and added again in one ALTER returning the dropped column’s values instead of its own default while the mutation was still pending. #114601 (Shaohua Wang).
• Fixed an infinite, uncancellable loop in functions hop and windowID when the span of an interval argument in seconds is a multiple of 2^32 (for example, toIntervalDay(2147483648)): the wrapped subtraction dodged the time-overflow check, and with constant arguments the loop ran at analysis time, where the query could not even be killed. The same interval could also spin a background thread of a WINDOW VIEW forever; such a window view is now rejected at creation. #114607 (Alexey Milovidov).
• Fixes a bug where an access entity carrying a Map-valued setting, such as a settings profile with http_response_headers or additional_table_filters, is stored in a form that ClickHouse cannot read back, leaving the entity permanently unloadable after a restart. #114620 (Groene AI).
• Fixed a refreshable materialized view leaking its rotated-out target table when ignore_drop_queries_probability is enabled. The DROP a refresh issues to clean up the previous target is a step of the refresh, not a DROP the user asked for, so the fault injection no longer applies to it. #114622 (Groene AI).
• Fixes has, indexOf, countEqual, mapContainsKey, mapContainsValue and Map subscript returning “not found” for a constant LowCardinality needle equal to the element type’s default value, such as an empty String or a zero number. #114624 (Groene AI).
• Fixed logical error Multi-block postings must be compressed in queries over tables with a text index in the lazy posting-list apply mode, when the query was canceled during the read. #114636 (Anton Popov).
• Fixed formatting of a subquery argument of the view and viewIfPermitted table functions when the enclosing query has a trailing SETTINGS clause. Such a query was formatted as view((SELECT ...)), which cannot be parsed back, so the query failed the internal format-parse-format check and raised Inconsistent AST formatting. #114658 (Groene AI).
• Fixed heap memory corruption when reading Parquet through an input format that owns its read buffer, for example a dictionary with SOURCE(FILE(... format 'Parquet')). Background prefetch and decode tasks could still read and write through the buffer after the pipeline released it, which could abort the server. #114668 (Groene AI).
• Fix reading a subcolumn (.size0, .null, a tuple element, String.size) through a Merge table when the underlying table declares the parent column as ALIAS. Such a read returned the type default, and when the parent was selected in the same query it could return the value of an unrelated column. #114673 (Groene AI).
• Fix ANY RIGHT JOIN and SEMI RIGHT JOIN with several OR-ed conditions in the ON section returning some rows of the right table twice and losing the matches of others. A written ANY LEFT JOIN was affected too, because the planner may swap the tables and execute it as ANY RIGHT JOIN. #114676 (Vladimir Cherkasov).
• Fixed wrong results of the trivial GROUP BY ... LIMIT optimization (setting optimize_trivial_group_by_limit_query) for queries with DISTINCT, QUALIFY, window functions, or arrayJoin in the projection: these consume or filter the groups after the aggregation, so capping the aggregation at LIMIT + OFFSET keys could return too few rows or wrong values. The optimization no longer applies to such queries. #114695 (Alexey Milovidov).
• Fixed rows missing from the result of a SELECT filtered by pointInPolygon over a MergeTree primary key when the polygon argument is an invalid constant literal and validate_polygons = 0. Primary key analysis derived granule pruning from a polygon it never validated, and dropped granules holding matching rows. #114710 (Groene AI).
• Fixes a server crash that happens when aggregate states of the -OrNull or -OrDefault combinators are merged with group_by_overflow_mode = 'any' and a non-zero max_rows_to_group_by, for example when a query reads through an aggregate projection. #114740 (Groene AI).
• Fixed a LOGICAL_ERROR (“Unexpected number of kept output positions after removing unused columns from ReadFromMergeTree”) raised by a SELECT ... FINAL whose PREWHERE uses a column that FINAL needs for merging (ver, is_deleted, sign) and which does not select that column. #114746 (Groene AI).
• Fix a logical error (Block structure mismatch) when reading a Merge table with make_distributed_plan enabled over children with internal set operations. #114753 (Alexey Milovidov).
• Reject an absurdly large kafka_num_consumers with a clear error instead of failing an allocation inside the Kafka table engine. Previously, with kafka_disable_num_consumers_limit enabled, such a value produced a std::length_error exception. #114764 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR: Context has expired exception when a stored expression containing accurateCastOrDefault or a to<Type>OrDefault function is evaluated after the query that created it has finished, for example a MATERIALIZED column default on INSERT or a mutation expression. #114769 (Groene AI).
• Fixes a LOGICAL_ERROR (Cannot find input column ... on its position in inputs of expression actions DAG) when a query has a JOIN above a table whose text-indexed column is used by both a PREWHERE and a WHERE text-search predicate. #114771 (Groene AI).
• Fix lost wakeup of restarting thread on startup from the attach thread. #114802 (Azat Khuzhin).
• OPTIMIZE TABLE on an object-storage table whose engine does not implement compaction reported success while doing nothing. It now raises an error, like every other engine that does not support the statement. This affects DeltaLake, Paimon, Hudi and plain object-storage engines (S3, GCS, COSN, OSS, AzureBlobStorage, HDFS, and object-storage-backed URL); Iceberg, which does implement compaction, is unchanged. Closes #114765. #114805 (Groene AI).
• timeSeriesRange and timeSeriesFromGrid threw a false DECIMAL_OVERFLOW for timestamps before 1970 because the start/end timestamps were read as unsigned values. Also, for DateTime/UInt32 timestamps the step was silently truncated to its lower 32 bits. Now all the calculations are done in Int64. #114815 (Vitaly Baranov).
• Fixes a server abort (Logical error: Part patch-... doesn't exist) when a DROP TABLE runs concurrently with a lightweight UPDATE on a ReplicatedMergeTree table. The update now holds the table’s shared lock until its patch part is committed, so the DROP waits instead of removing it. #114826 (Groene AI).
• Fix few bugs in async inserts (possible crash in case of some queries failed with MEMORY_LIMIT_EXCEEDED and never worked async_insert_queue_flush_on_shutdown). #114839 (Azat Khuzhin).
• Fixed LOGICAL_ERROR: Unexpected return type from tuple. Expected Tuple(..., UInt8). Got Tuple(..., LowCardinality(UInt8)) when a LowCardinality column compared with IN (subquery) is read through tuple(...) across a subquery boundary and enable_analyzer = 0. With the old analyzer, x IN (subquery) over a LowCardinality column now returns LowCardinality(UInt8), the same type as x IN (literal list) and the same as with enable_analyzer = 1. #114856 (Nikita Fomichev).
• The AI-function quota settings (ai_function_max_api_calls_per_query, ai_function_max_input_tokens_per_query, ai_function_max_output_tokens_per_query) were enforced per block rather than per query, letting a query exceed the configured limit in proportion to the number of blocks and threads processed. They now bound the whole query per server: the API-call limit is an exact cap (atomic reservation), and the token limits are best-effort. A distributed query enforces the limits independently on each participating server / fragment. #114885 (George Larionov).
• Fixes a server abort that happens when the global thread pool is momentarily saturated while tables are still loading. AsyncLoader now keeps running if it cannot spawn an extra loader worker and some worker will still drain the queued jobs, instead of calling std::terminate. #114897 (Groene AI).
• Fixed undefined behavior and incorrect boundary results in quantile and quantiles for DateTime64 and wide Decimal values. #114919 (Alexey Milovidov).
• Fix server abort on a corrupt ORC stripe footer. #114924 (Groene AI).
• Fix possible crash due to ThreadGroup use-after-free (for async INSERTs/MVs/EXPLAIN ANALYZE). #114940 (Azat Khuzhin).
• Fixed a hang when dropping a TimeSeries table whose name sorts lexicographically below its inner tables’ names, for example DROP TABLE `-ts` or dropping a materialized view declared with ENGINE = TimeSeries. The drop self-deadlocked on the DDL guard and could not be cancelled with KILL QUERY. #114953 (Groene AI).
• Fix BAD_DATA_PART_NAME (“Trying to get part name in new format for old format version”) on a ReplicatedMergeTree created with the deprecated positional syntax under allow_deprecated_syntax_for_merge_tree. A mutation could be reported as done on a replica that had not applied it, and the table could stay readonly after a restart. #114980 (Groene AI).
• Fix the logical error Not-ready Set is passed as the second argument for function 'in' for a query with an IN subquery reading from a remote cluster: the query now fails with the actual error from the subquery, such as a connection failure. #114983 (Alexey Milovidov).
• Fix logical error for IN with an identical non-constant tuple on both sides in the old analyzer. #114988 (Groene AI).
• Fix reading a tuple element through a positional Tuple structure hint. #115000 (Groene AI).
• Do not alias pre-aggregation states of a zero-size aggregate in arrayReduceInRanges. #115008 (Groene AI).
• Fixed an INSERT into a Log, TinyLog or StripeLog table that fails while committing the recorded file sizes. Such an insert could keep its rows in the table even though it reported an error, or leave the data files of an array column inconsistent with each other. Reading a Log or TinyLog table whose array column holds no elements while its offsets claim some now raises an error naming the column instead of reading past the end of the elements. A partially written elements column was already rejected. #115027 (Groene AI).
• Fix an adaptive aggregator abort on lazily replicated columns. #115034 (Groene AI).
• Fixed an out-of-bounds read in pointInPolygon with a constant polygon whose bounding box is unbounded, for example when a coordinate reaches +-DBL_MAX. Such a polygon is now rejected with BAD_ARGUMENTS, and a polygon with an empty ring is treated as having no interior. Previously the query could return a wrong result, or abort the server in builds with hardening or sanitizers enabled. #115035 (Groene AI).
• Fixed an out_of_range exception raised as an internal Code: 1001 error when a sequenceMatch, sequenceCount or sequenceMatchEvents pattern contains the event number 0, for example sequenceMatch('(?0)'). Such a pattern is now rejected with BAD_ARGUMENTS. A temporal condition holding a lone sign, such as sequenceMatch('(?1)(?t>+)(?2)'), was silently treated as (?t>0) and is now rejected with SYNTAX_ERROR. #115056 (Groene AI).
• Fixed Found patch part ... that intersects mutation with version ... (LOGICAL_ERROR) on tables with lightweight updates. A merge of patch parts could produce a patch whose range of data versions spans the data version of an existing part; that patch can then be neither applied nor skipped, so every later operation on the partition failed and the replication queue stalled. Such merges are no longer assigned. #115064 (Alexey Milovidov).
• Fixes a LOGICAL_ERROR (Cache file segment is in detached state) when reading through the filesystem cache with the experimental use_reader_executor setting. The regression was introduced by #110029 and is not in any release, so no backport is owed. #115070 (Groene AI).
• Arrow Flight server: accept Basic authentication credentials encoded in Base64 without padding, as sent by Go-based Flight clients such as the ADBC Flight SQL driver, and return a proper authentication error instead of a generic Unexpected error in RPC handling when the authorization header is malformed. #115085 (Alexey Milovidov).
• Fix autocompletion (hints and TAB completion from system.completions) in the embedded client used by the SSH interface and the Web terminal: suggestions were loaded through a separate connection that authenticated as the default user with an empty password, which failed on servers where default has a password, and otherwise leaked names of entities the logged-in user has no access to. Suggestions are now loaded through the session’s own connection. #115086 (Alexey Milovidov).
• Fixed a server that refuses to start after a table was created with CREATE TABLE ... AS url('http://host/**/', ...) while allow_experimental_url_wildcard_from_index_pages was enabled. Loading such a table’s metadata re-evaluated the experimental check and failed with SUPPORT_IS_DISABLED, aborting startup. #115095 (Groene AI).
• Rewrite bucketIndexForTimestamp in 64-bit arithmetic. #115097 (Vitaly Baranov).
• Fixed the formatting of ALTER TABLE ... MATERIALIZE INDEX, MATERIALIZE STATISTICS and MATERIALIZE PROJECTION, which dropped the IF EXISTS clause, so the formatted query was not the query that was parsed. Also rejected four ALTER ... STATISTICS AST JSON payloads that the SQL parser cannot produce and that formatted into a different command. #115099 (Groene AI).
• Fixed an exception (Block structure mismatch in Expression for FilterSortedStreamByRange) for some SELECT ... FINAL queries with PREWHERE over a table whose sorting key is an expression. #115113 (Shaohua Wang).
• Fix clickhouse-local CORS CLI overrides. #115129 (Alexey Milovidov).
• Disable distributed index analysis with projections (wrong results). #115132 (Azat Khuzhin).
• Fixed if and multiIf returning a large positive value instead of a negative one when a Time branch is combined with Time64, DateTime or DateTime64 and the expression is JIT-compiled. #115146 (Groene AI).
• A web terminal embedded in another page, such as the SQL console or play.html, receives its credentials from that page and no longer opens a connection of its own before they arrive. Previously it connected twice - once with an empty password, then again with the real credentials - and printed Connecting... twice. #115148 (Alexey Milovidov).
• Fixed wrong results when a WHERE clause contained a NULL expression under a NOT, for example SELECT count() FROM t WHERE NOT (id >= 20 AND NULL). Index analysis treated such a condition as provably true for the whole range, so the filter was skipped: count() returned rows the query rejects, and ALTER TABLE ... DELETE with that condition removed rows it should have kept. #115152 (Groene AI).
• Fixed EXPLAIN ANALYZE losing the join statistics of a view declared SQL SECURITY DEFINER or SQL SECURITY NONE whose body contains a JOIN. In debug and sanitizer builds the query aborted with the logical error JoinStep analyzed without the analyze mode. #115156 (Groene AI).
• Fixed an out-of-bounds read while deserializing corrupted posting lists of a text index. #115222 (Anton Popov).
• Fixed wrong results for SELECT count(arrayJoin(arr)) with the default optimize_trivial_count_query = 1. The stored row count was returned instead of the number of array elements, and on file() and url() the query returned 0. Closes #115123. #115227 (Groene AI).
• Fixes NOT_FOUND_COLUMN_IN_BLOCK on every INSERT into a table that declares STATISTICS(...) on an ALIAS or EPHEMERAL column, which made such tables read-only. #115231 (Groene AI).
• Fix the final statistics of a finished query (N rows in result, Read ... rows) being rendered in an invisible color in the Web UI. #115245 (Alexey Milovidov).
• Fixed silently incorrect results and undefined behavior when reading a malformed AggregateFunction(uniq, ...) state. The skip_degree field of the serialized state was used as a shift exponent without validation, so an out-of-range value produced an arbitrary cardinality instead of an error. Such states are now rejected with INCORRECT_DATA. #115276 (Groene AI).
• Fixed a row policy bypass where the mergeTreeIndex table function exposed primary key and minmax index values of the rows hidden by a SELECT row policy on the source table; reading mergeTreeIndex for a table with a row policy is now denied. #115304 (Yarik Briukhovetskyi).
• Fixed CREATE QUOTA and ALTER QUOTA silently discarding all but the last FOR INTERVAL block when the separating comma between intervals was omitted. Such a statement succeeded while installing none of the earlier limits. The intervals now accumulate again, as they did before 2020. Closes #115282. #115320 (Groene AI).
• Fixed wrong results from if, multiIf and CASE when one branch is a Date or Date32 and the result type is DateTime or DateTime64. With compile_expressions = 1 (the default) the compiled expression copied the date’s day count into the timestamp column without converting days to seconds, so if(cond, toDate32('1900-01-01'), toDateTime64('1970-01-01', 3)) returned 1969-12-31 23:59:34.433. Closes #115272. #115325 (Groene AI).
• Do not leave metadata referencing a dropped named collection. #115326 (Groene AI).
• Fixed timezoneOffset (alias timeZoneOffset) returning an offset wrong by exactly one day for time zones whose UTC offset differs from their 1970-01-01 offset by a whole day, such as Pacific/Kiritimati and Pacific/Apia after they crossed the international date line, and for most zones before 1970. This also fixes parseDateTime, parseDateTimeOrNull, parseDateTime64, parseDateTimeInJodaSyntax, EXTRACT(TIMEZONE_HOUR / TIMEZONE_MINUTE ...), formatDateTime’s %z, toUTCTimestamp and fromUTCTimestamp, which consume that offset. Closes #115281. #115332 (Groene AI).
• Fixed a bypass of settings constraints: SET <name> = DEFAULT reset a setting without checking the constraints from the user’s profile, so a MIN/MAX/CONST constraint could be escaped by resetting the setting instead of assigning to it, and a session could leave readonly mode with SET readonly = DEFAULT. #115334 (Alexey Milovidov).
• Arrow Flight server: reject malformed Basic authentication credentials that omit the username/password separator. #115347 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR (Join is supported only for pipelines with one output port) for a full_sorting_merge JOIN with query_plan_join_shard_by_pk_ranges enabled when one of the two sides is pruned to no parts at all - by the primary key, by column statistics, or because the table is empty. #115352 (Alexey Milovidov).
• Disable the query condition cache for filters over _part_starting_offset. #115358 (Azat Khuzhin).
• Fixed randConstant returning two different values in one query when its calls carry different aliases. randConstant is documented to hold a single value for the whole query, and SELECT randConstant() = randConstant() already returned 1, but adding aliases as in SELECT randConstant() AS a, randConstant() AS b produced two unrelated values. #115384 (Alexey Milovidov).
• Fixed uniq, uniqExact, uniqHLL12 and uniqTheta returning a wrong result for an argument wrapped in an injective function that hides nullability, such as uniqExact(tuple(x)) over a Nullable column. The optimize_injective_functions_inside_uniq optimization removed the wrapping function, after which NULL rows were skipped instead of counted. #115466 (Vladimir Cherkasov).
• Fix GroupingAggregatedTransform producing duplicate chunks for the same bucket. #115487 (Nikita Taranov).
• Fixed the error message of geoToH3 naming the wrong argument position and the wrong type when a coordinate is not Float64. Under the default geotoh3_argument_order = 'lat_lon' the message pointed at the other argument and printed that argument’s type. Closes #101813. #115503 (Groene AI).
• Reject a bare carriage return (\r), in addition to a line feed (\n), in HTTP header names and values that pass through the HTTP header filter. This closes an HTTP header injection (request/response splitting) gap for features that pass user-provided headers, such as the url table function and engine, HTTP dictionaries, and data lake catalogs. #115510 (Sergei Trifonov).
• Fixed caches configured by a maximum number of entries (SLRU policy) silently stopping to admit new entries once that many entries had been accessed at least twice. Affected caches include the Iceberg and Paimon metadata file caches, the Parquet metadata cache, the text index caches, the vector similarity index cache, the compiled expression cache and the ObjectStorageQueue file status cache. #115517 (Groene AI).
• Fixed a hang of up to one hour in BACKUP/RESTORE ... TO S3, GCS disks and the s3 table function when the named collection sets http_client = 'gcp_oauth' and the Google OAuth2 token endpoint accepts the connection but never answers. The bearer token request no longer inherits the caller’s data transfer request timeout; its response wait is bounded at 10 seconds, matching the STS AssumeRole request cap. #115531 (Groene AI).
• Fix dictGet and its variations in distributed queries: an unqualified dictionary name is now bound to the current database of the initiator, instead of being resolved against the current database of each shard, where it either was not found or resolved to a different dictionary of the same name. #115541 (Alexey Milovidov).
• Fix a race that could make a STREAM BOUNDED query return fewer rows than were committed before it started, usually none. #115543 (Alexey Milovidov).
• Fixed Keeper snapshots being written as empty files when compress_snapshots_with_zstd_format is disabled. The server logged Created persistent snapshot but the snapshot file was 0 bytes, and the node then refused to start with CORRUPTED_DATA. Closes #115388. #115553 (Groene AI).
• Fix read_rows, read_bytes, written_rows and written_bytes in system.query_thread_log reporting a thread lifetime running total instead of the values for the logged query. Rows of long-lived pooled threads, in particular the query’s initiating HTTPHandler thread, were affected. #115596 (Groene AI).
• Fixed an out-of-bounds read of the cached dictionary hashes when aggregating or joining on a LowCardinality(String) key whose dictionary grows while it is being aggregated, reachable from TTL ... GROUP BY. It could yield a wrong hash, and a reported heap-buffer-overflow in sanitizer builds. #115604 (Groene AI).
• CREATE TABLE and ALTER TABLE now reject a lossy codec such as SZ3 on a column used in the partition key. #115626 (Groene AI).
• Fix invalid backslash escape characters in text, tokenbf_v1 and ngrambf_v1 indexes. #115634 (Elmi Ahmadov).
• Fixed removePartAndCoveredParts when it is called for a part that is present in the active data part set. #115640 (Mikhail Artemenko).
• Allow KeeperMap readers to accept shared metadata when equivalent primary keys differ only by redundant outer parentheses. #115642 (Sergey Kuznetsov).
• Fixed a bug where restoring a database or table under a new name dropped the table aliases from the bodies of stored views, so reading a restored view failed with UNKNOWN_IDENTIFIER. Closes #115479. #115645 (Groene AI).
• Fixed Code: 349. Cannot convert NULL value to non-Nullable type when a has or notHas predicate carries a NULL array element and the primary key is a String, Array(String) or Map(String, String), and fixed wrong results when that element is a Dynamic or Variant holding NULL. The NULL is now dropped from the set during key analysis, as it already was for other key types. #115647 (Groene AI).
• Fixed a bug in the function formatRowNoNewline that could produce incorrect results or a logical error when a row is formatted to an empty result. #115669 (Alexey Milovidov).
• Reject zero and negative step values in one-point Prometheus range queries. #115679 (Minh Vu).
• Reject PromQL timestamps and durations that exceed the Int64 range. #115682 (Minh Vu).
• Fix server crashes on deeply nested input in readWKB, JSONMergePatch, the address parser of the url/remote/cluster/mysql/postgresql table functions, the CURSOR clause parser and the CapnProto schema parser. Fix a crash in the parser of the EXECUTE statement when an argument is not a literal. #115700 (Alexey Milovidov).
• Do not return uninitialized memory in the result of a binary string literal whose length is not a multiple of eight, and from the LZ4 decompressor when a compressed block has no body. #115704 (Alexey Milovidov).
• Bound the size of the startup message of the PostgreSQL wire protocol, which is read before authentication. #115708 (Alexey Milovidov).
• Fix the MULTIPLE_EXPRESSIONS_FOR_ALIAS error in queries where a column alias coincides with the alias of a table expression in the FROM section (an analyzer regression in version 25.2). #115714 (Alexey Milovidov).
• Fixed clickhouse-local terminating the whole process when a connection handler of a listener started with SYSTEM START LISTEN threw, for example because a client sent a malformed packet or reset a keep-alive socket. Such a process now logs the error and keeps serving, like clickhouse-server does. #115715 (Groene AI).
• Fix LOGICAL_ERROR: No available columns when a query reads no columns from a table expression that has no columns at all, for example SELECT count() FROM loop(db, tbl) where tbl is an Alias table whose target was dropped, or a parameterized view. Such a query now fails with UNSUPPORTED_METHOD instead of aborting the server in debug and sanitizer builds. #115716 (Groene AI).
• Fix LOGICAL_ERROR “Reading from materialized CTE ‘X’ before its materialization completed - DelayedPortsProcessor gate is missing in the query plan” when a MATERIALIZED CTE is read from inside an IN subquery over a Distributed table. Closes #113184. #115740 (Groene AI).
• Reject invalid names of projections and indexes. #115824 (Alexander Tokmakov).
• Reject PromQL vector matching modifiers when one operand is a scalar. #115828 (Minh Vu).
• Fixed cast_keep_nullable not preserving nullability when the target type is LowCardinality. Casting a NULL-capable value to LowCardinality(T) threw CANNOT_INSERT_NULL_IN_ORDINARY_COLUMN instead of producing LowCardinality(Nullable(T)). Closes #103485. #115851 (Groene AI).
• Add Objects’ paths to deduplication hashing, preventing incorrect deduplication. #115866 (Mikhail f. Shiryaev).
• Fix TYPE_MISMATCH (“CAST AS Array can only be performed between same-dimensional array types”) for a query with a constant of type Array(Map(...)), such as has([map('k', 'v')], m), executed over a Distributed table or with parallel replicas. #115883 (Alexey Milovidov).
• Fix Unknown identifier raised during query plan optimization under make_distributed_plan = 1 when a window query’s sort column passes through an Expression step that does not mention it. Closes #115737. #115890 (Groene AI).
• Fixed h3kRing, h3HexRing, h3Line and h3ToChildren ignoring max_execution_time and KILL QUERY while expanding a block of rows. #115893 (Alexey Milovidov).
• Reading through a Merge table that matches an Alias table whose target is a parameterized view no longer raises LOGICAL_ERROR: Table has no columns. (which aborts the server in debug and sanitizer builds). Such a table now reports STORAGE_REQUIRES_PARAMETER, the same error the parameterized view itself reports. Any other matched table that cannot supply columns now reports UNSUPPORTED_METHOD instead of an internal error. #115896 (Groene AI).
• Fixed the on-the-fly preview of a pending mutation whose expression references _sample_factor. With apply_mutations_on_fly = 1 and a SAMPLE clause, the expression was evaluated with the reading query’s sample factor instead of the value the background materialization persists, so previewed rows disagreed with the eventual materialized result, and a pending ALTER TABLE t DELETE WHERE _sample_factor > 1 could make a SAMPLE read return no rows at all. The query’s own _sample_factor column is unchanged. #115906 (Groene AI).
• Fixed a case where ATTACH TABLE carrying an ENGINE = URL(...) definition did not check the TABLE ENGINE grant of the engine the URL scheme dispatches to. A user holding only TABLE ENGINE ON URL could attach and then read a File, S3, AzureBlobStorage or HDFS backed table, which CREATE TABLE correctly rejects. #115914 (Groene AI).
• Do not let an unparsable KeeperMap metadata node stop the server from starting. #115941 (Groene AI).
• Fix a crash of the showCertificate function on a server that obtains its TLS certificate with ACME (Let’s Encrypt). The function now reports the certificate that is actually served, also after a certificate reload. #115976 (Alexey Milovidov).
• Fixes a server abort in debug and sanitizer builds when inserting into a materialized view whose TO target is another materialized view. Such an insert already failed on released versions, with an uninformative Code: 1001 STD_EXCEPTION; it now reports NOT_IMPLEMENTED naming the offending view. #115985 (Groene AI).
• Statistics files with format version V3 are now read instead of rejected with ILLEGAL_STATISTICS. Such files were produced only by builds of master from a narrow window, but parts containing them were completely unreadable, and on a readonly disk ALTER TABLE ... MATERIALIZE STATISTICS cannot regenerate them. #116030 (Alexey Milovidov).
• Fixed LOGICAL_ERROR: Index with name auto_minmax_index_<column> already exists when a single ALTER TABLE statement combined RENAME COLUMN with another command, such as MODIFY SETTING, on a table with implicit minmax indices (add_minmax_index_for_numeric_columns and friends). Renaming a plain column also no longer leaves a following DROP COLUMN failing with UNKNOWN_IDENTIFIER, nor a following ADD COLUMN without its implicit index. #116063 (Groene AI).
• Fix stateless workers crash with object storage. #116067 (Konstantin Vedernikov).

​

Build/Testing/Packaging Improvement

• Provide signed artifacts for macOS. #115169 (Konstantin Bogdanov).
• Fix building from source with clang 23. #112288 (Azat Khuzhin).
• Reduced the dependencies and the code size of the SQL parser, which can now be built on its own — see utils/wasm-parser for a standalone WebAssembly build of it. The tables of SQL keywords and access types are now built at compile time instead of on every server start. #112067 (Alexey Milovidov).
• Building from source now fails during CMake generation instead of corrupting StorageSystemLicenses.generated.cpp when license generation fails. #111222 (Mikhail Artemenko).
• Updated the bundled Apache Thrift library to version 0.24.0. #111278 (Groene AI).
• Add libucontext for musl builds, enabling the boost::context ucontext backend used by sanitizer builds. #111437 (Konstantin Bogdanov).
• Fix source-level portability issues when building with musl. #111631 (Konstantin Bogdanov).
• Editing Core/ProtocolDefines.h no longer triggers a near-full rebuild when building from source: the include was removed from BlockInfo.h and added only in the files that use its constants. #111650 (Alexander Gololobov).
• Fix a startup abort in macOS debug builds by disabling memory-tracker debug checks for allocations made by system libraries. #112080 (Arsen Muk).
• Updated the bundled libarchive to version 3.8.9. #112434 (Konstantin Bogdanov).
• Updated StringZilla to version 5.0.5. #112319 (Konstantin Bogdanov).
• Update jwt-cpp to v0.7.2. #112595 (Konstantin Bogdanov).
• Update the bundled SQLite amalgamation from 3.41.2 to 3.53.4. #112599 (Konstantin Bogdanov).
• Updated aws-c-http to version 0.11.0. #112413 (Govind R Nair).
• Update libcotp to v4.2.1. #112596 (Konstantin Bogdanov).
• OpenSSL can now be built from its portable C sources on architectures that have no hand-written assembly in the build description, instead of failing at the final link with undefined symbols. The new OPENSSL_NO_ASM CMake option selects the same path explicitly. #112863 (Alexey Milovidov).
• Added an ENABLE_ORC CMake option (on by default), so a build can have Parquet without ORC and without the protobuf and protoc that ORC brings with it. #112864 (Alexey Milovidov).
• Common/Exception.cpp no longer requires ClickHouse’s patched libc++ to compile, so a build linking a different C++ standard library gets exceptions without a throw-site stack trace rather than not compiling. #112865 (Alexey Milovidov).
• Stack unwinding on aarch64 musl builds now continues through signal frames and blocked syscalls instead of stopping at the libc frame, improving stack traces from the fatal error handler and the sampling query profiler. #112592 (Konstantin Bogdanov).
• Update the vendored libexpat to 2.8.2. #112593 (Konstantin Bogdanov).
• Update simdjson to v4.6.5. #112594 (Konstantin Bogdanov).
• Update Apache Arrow to 25.0.0 and Apache ORC to 2.3.1. #112604 (Konstantin Bogdanov).
• Coverage builds (WITH_COVERAGE) now link the profile runtime built in-tree from the vendored LLVM sources instead of the archive bundled with the host clang toolchain. #112607 (Konstantin Bogdanov).
• Update Protocol Buffers to v35.1 and gRPC to v1.83.0. #112614 (Konstantin Bogdanov).
• Added an experimental WebAssembly (wasm64, via Emscripten) target: emcmake cmake now configures the tree and the lower layers build. Also fixed -DENABLE_LIBFIU=OFF, which did not build on any platform. #112911 (Alexey Milovidov).
• Fixed every -Wunique-object-duplication site: header-defined mutable singletons that would get one copy per shared object instead of one per process. Mostly by moving the definition into a .cpp. #112995 (Alexey Milovidov).
• Installing a tgz package over a dangling symlink now preserves the symlink chain and applies the packaged target’s mode and ownership instead of a umask-derived mode. #112766 (Alexey Milovidov).
• In aarch64 musl builds, memmove now uses the ARM optimized-routines assembly implementation (shared with memcpy) instead of the generic C word-at-a-time loop. #113502 (Konstantin Bogdanov).
• In musl builds, getaddrinfo with AI_ADDRCONFIG now follows glibc semantics, so localhost no longer resolves to an unusable ::1 first in IPv4-only environments, and sysconf returns glibc’s buffer-size hints for getpwuid_r/getgrgid_r instead of -1, unbreaking libhdfs3. #113485 (Konstantin Bogdanov).
• In musl builds, sched_getcpu is now an rseq read (~1ns) instead of a syscall, matching glibc and removing the system-time overhead of jemalloc’s percpu_arena on allocation-heavy workloads; musl also registers rseq eagerly and exports the glibc __rseq_* ABI. #113501 (Konstantin Bogdanov).
• Fix local (non-CI) builds compiling all vendored Rust code (the PCO codec, Delta Lake reads via delta-kernel-rs, prqlc, chdig, wasmtime, the client’s fuzzy history search via skim) with opt-level = 0. The SANITIZE CMake variable was declared with option, so its default value was the BOOL OFF, which failed the STREQUAL "" “no sanitizer” checks, and every build configured without an explicit -DSANITIZE=... took the sanitizer workaround path that disables Rust optimization. #114083 (Alexey Milovidov).
• Push docker image layers zstd-compressed. #114342 (Alexey Milovidov).
• Build Debian packages for the release builds only. #114739 (Alexey Milovidov).
• Fixed building unit_tests_dbms in configurations without the full library set (ENABLE_LIBRARIES=0 or ENABLE_ROCKSDB=0). #115655 (Alexey Milovidov).

​

ClickHouse release 26.7, 2026-07-22. Presentation, Video

​

Backward Incompatible Change

• Mark rotated non-replicated MergeTree system log tables without TTL as table_readonly to avoid unnecessary background operations. The table_readonly MergeTree setting now also suppresses all background work on a plain MergeTree table — regular, TTL (DELETE/MOVE/recompression) and recompression merges, background mutations, and background part moves — and rejects the mutating partition commands (ATTACH/MOVE/DROP/DROP DETACHED/FETCH/REPLACE PARTITION and MOVE PARTITION ... TO TABLE into the table), in addition to the inserts, mutations, and OPTIMIZE it already rejected. As a result, a table_readonly table with a TTL no longer reclaims its expired data while the setting is enabled. Downgrade warning: the rotated system log tables (for example system.text_log_0) now store table_readonly = true in their metadata. The table_readonly setting exists since 26.3, so servers older than 26.3 fail to start after a downgrade with UNKNOWN_SETTING (Unknown setting 'table_readonly': for storage MergeTree: Cannot attach table). Before downgrading to a version older than 26.3, drop the rotated system.*_log_N tables (it is safe to drop them), or run ALTER TABLE system.<table>_log_N RESET SETTING table_readonly on each of them. #95079 (Mathuranath Metivier).
• S3 access originating from user SQL no longer resolves the server’s own cloud credentials (environment, IMDS/IRSA, instance profile, AWS config files, role_arn-based STS, GCP OAuth metadata) by default; such a request must use explicit credentials or NOSIGN. Named collections now default use_environment_credentials to 0. The previous behaviour can be restored per request with use_environment_credentials = 1 (or globally via the <s3> config) together with the new setting s3_allow_server_credentials_in_user_queries (disabled by default). Explicitly supplied and operator-provisioned config credentials are unaffected. On server startup or RESTORE, a persistent S3/S3Queue table, dynamic S3 disk, or DataLakeCatalog whose definition resolves such credentials is loaded anonymously (inaccessible until re-credentialed) instead of aborting startup, controlled by the new server setting s3_load_table_anonymously_if_credentials_restricted (enabled by default). #106855 (Raúl Marín).
• The **/ glob (any number of directories) now also matches the same directory. Previously, **/ in glob patterns was not handled as a special case, so data/**/file.txt would not match data/file.txt (zero directory levels). #97676 (Alexey Milovidov).
• Reject the zip/zipx backup archive format for backups stored on object storage — direct S3(...)/AzureBlobStorage(...) destinations and Disk(...) destinations backed by S3 or Azure — for both BACKUP and RESTORE. Zip requires seeking to read its central directory, which is very slow over object storage. Use a tar-based format such as tar.gz instead. #101770 (Yash ).
• Extended the supported range of DateTime64 from [1900-01-01, 2299-12-31] to [0000-01-01, 9999-12-31]. Values outside the former range are now computed correctly (via cctz) instead of being clamped to the boundary. With precision 8 or 9 the range remains narrower because the ticks are stored as Int64 (with nanosecond precision the maximum is still 2262-04-11). Backward compatibility: if you used out-of-range date-times before, but relied on the very specific saturation rules inside the old range, keep in mind that now the results are changing to be more correct. #107907 (Alexey Milovidov).
• AggregatingMergeTree now rejects, at table creation time, schemas where a column is neither part of the sorting key nor an aggregate-state measure (AggregateFunction/SimpleAggregateFunction). Such columns are silently collapsed to an arbitrary value during background merges, producing wrong results for queries that GROUP BY or filter on them. Set allow_dimensions_outside_sorting_key = 1 to restore the previous behavior. Closes #751. #108087 (Alexey Milovidov).
• Removed the config-based workload scheduling configuration (the resources and workload_classifiers server config sections). Use CREATE RESOURCE and CREATE WORKLOAD queries instead. The obsolete config sections are ignored with a warning. #108286 (Sergei Trifonov).
• ClickHouse now always uses the unified insert deduplication hash for both synchronous and asynchronous inserts; the legacy per-insert deduplication behaviours are removed. The server setting insert_deduplication_version is kept as a migration guard: the server refuses to start if it is set to a legacy value (old_separate_hashes or compatible_double_hashes). To upgrade from a version that used a legacy value, first run a release that supports compatible_double_hashes (which writes both the legacy and unified hashes). For replicated tables run it for at least replicated_deduplication_window_seconds (one hour by default; the default windows retain the unified hashes of all inserts for that window, which is considered enough for an insert retry loop). For non-replicated tables with non_replicated_deduplication_window > 0 that window is count-based rather than time-based, so run compatible_double_hashes for at least that many inserts. Then remove the setting (or set it to new_unified_hash) before upgrading to this version. #108361 (Sema Checherinda). If you never set insert_deduplication_version, you can ignore this item.
• Removed the long-deprecated functions snowflakeToDateTime, snowflakeToDateTime64, dateTimeToSnowflake and dateTime64ToSnowflake. They were deprecated back in v24.6 in favor of snowflakeIDToDateTime, snowflakeIDToDateTime64, dateTimeToSnowflakeID and dateTime64ToSnowflakeID, which should be used instead. The setting allow_deprecated_snowflake_conversion_functions (which used to re-enable them) is now obsolete and has no effect. #108711 (Alexey Milovidov).
• The setting use_legacy_to_time is now 0 by default, so toTime converts values into the Time data type instead of converting a date with time to a fixed date. The legacy behavior is still available via the toTimeWithFixedDate function or by setting use_legacy_to_time = 1. #108729 (Alexey Milovidov).
• Naive Bayes models (used by naiveBayesClassifier) are now configured as a dictionary with the NAIVE_BAYES layout, built at load time from a table of pre-aggregated per-class n-gram counts, instead of the previous server-side configuration (an XML config file referencing serialized .bin model files), which is no longer supported — existing models must be recreated as dictionaries. Three new functions are added alongside naiveBayesClassifier: naiveBayesClassifierWithProb returns the predicted class together with its probability, naiveBayesClassifierWithAllProbs returns every class with its probability, and naiveBayesNgrams tokenizes text into n-grams the same way the dictionary does, for building the training data from raw labelled text. Additionally, several performance optimizations have been applied: on a 9.7 MiB code-point trigram language model, it uses ~49× less memory (from 1.84 GiB to 38 MiB), loads ~11× faster, and classifies ~11× faster. #108773 (Nihal Z. Miaji).
• The hasColumnInTable function no longer accepts the optional hostname, username, and password arguments for checking a column on an arbitrary remote server; only the hasColumnInTable(database, table, column) form remains. The removed remote mode was a security concern: it let any user trigger outbound connections to arbitrary hosts and leaked credentials into query logs, and it was not gated by any privilege. #110881 (Alexey Milovidov).

​

New Feature

• Added EXPLAIN ANALYZE for examination of query performance: the query is executed, and the actual execution metrics are rendered in the familiar query plan format. #106586 (Kirill Kopnev). #110668 (Kirill Kopnev).
• Added support for WHERE clauses in projection definitions. Projections with WHERE only materialize rows matching the predicate, and the optimizer can use them (cost-based) when the query’s WHERE implies the projection’s WHERE. #102347 (S Bala Vignesh).
• Added the skip_unavailable_shards_mode setting (also available as a Distributed engine setting) to control which exceptions from a remote shard are silently ignored when skip_unavailable_shards is enabled. #79091 (cjw).
• Add groupFormat aggregate function that formats rows in each group using a specified output format and returns the result as a string. #93201 (Yang Hu).
• Added AWS_MSK_IAM as a supported value for kafka_sasl_mechanism, enabling ClickHouse to authenticate with Amazon MSK using IAM roles without managing SASL/SCRAM credentials. #96100 (kalavt).
• Add aggregate function combinator -Tuple, which applies the underlying aggregate function to each element of a Tuple column independently and returns a Tuple of the results, preserving element names: sumTuple(t) for t = (a, b) returns (sum(a), sum(b)). Aggregate functions with several arguments take one tuple per argument, paired by position: corrTuple((a1, a2), (b1, b2)) returns (corr(a1, b1), corr(a2, b2)). Unlike -ForEach over arrays, elements may have different types, and per-element result types and names are preserved. #98190 (RinChanNOW).
• Text indexes now support a postprocessor argument — an arbitrary expression that transforms each token after tokenization (for example, lower). #98939 (Jimmy Aguilar Mena). #108606 (Jimmy Aguilar Mena).
• Added the system.stemmers system table, which lists all languages that can be specified for the stem function. #100611 (Jimmy Aguilar Mena).
• Add support for WITH TIES for negative LIMIT. #100930 (Nihal Z. Miaji).
• Added support for the standard SQL AT TIME ZONE and AT LOCAL postfix operators as syntactic sugar for toTimeZone. The expression expr AT TIME ZONE zone is now equivalent to toTimeZone(expr, zone), and expr AT LOCAL is equivalent to toTimeZone(expr, timeZone()). #106092 (lizepeng).
• The URL table engine and url table function now dispatch to the appropriate backend based on the URL scheme: file:// is served by the File engine, s3:///gs:///gcs:///oss:// by S3, az:///azure:///abfss:///abfs:// by AzureBlobStorage, hdfs:// by HDFS, and http(s):// by the URL engine as before. The url_base setting is applied before scheme dispatch. Only the S3 schemes resolved by the default url_scheme_mappers are dispatched; other S3-compatible vendor schemes (cos, obs, …) are not, and require using the s3 engine/function directly. #106093 (Alexey Milovidov).
• Added the Remote and RemoteSecure table engines, the persistent counterparts of the remote and remoteSecure table functions. CREATE TABLE ... ENGINE = Remote('addresses', db, table, ...) now works in addition to CREATE TABLE ... AS remote(...). #106189 (Alexey Milovidov).
• Added engine-agnostic SYSTEM STOP, SYSTEM START, SYSTEM PAUSE, SYSTEM CANCEL, and SYSTEM REFRESH commands, and their ... ALL BACKGROUND server-wide forms, to control the background activity of Kafka, RabbitMQ, NATS, S3Queue/AzureQueue tables and refreshable materialized views through one unified interface. For refreshable materialized views they alias the existing SYSTEM ... VIEW commands. As part of supporting these controls on NATS, JetStream tables now acknowledge messages only after a successful insert (at-least-once, previously messages were auto-acknowledged on delivery and could be lost if the insert failed). A new nats_wait_for_flush_interval setting (default false, preserving the previous low-latency behaviour) optionally keeps a consumption cycle open for the whole flush interval instead of flushing as soon as the queue drains. A new nats_commit_on_select setting makes a direct SELECT on a JetStream table consume (acknowledge) the messages it reads. #107476 (Samuel Krempaský).
• Added support for the mysql, postgresql, and sqlite table functions and table engines to accept a user’s query (instead of a table name) and pass it to the external database as is, written either as a subquery (SELECT ...) or as query('SELECT ...'). The structure of the resulting table is inferred from the query result, and such a table is read-only. Closes #46758. Based on the initial implementation by Denis Vidiaev (#79652). #107740 (Alexey Milovidov).
• Added tabs to the ClickHouse Web UI (Play) for working with multiple queries at once. Tabs, their titles, parameters, active state, and small result snapshots are persisted across reloads (large and image results are not restored), integrate with browser history and the URL, and can be switched with the mouse wheel. #107826 (Alexey Milovidov).
• Introduce the QueryRunner table engine. Records inserted into a QueryRunner table represent queries that the engine executes. The engine can be used for asynchronous query execution, batch execution of generated queries, directing queries to remote clusters, benchmarks, fuzzing, and testing with shadow traffic. #107888 (Miсhael Stetsyuk).
• Add the GeoJSON output format for writing GeoJSON FeatureCollection documents, producing one feature per row. The single geo-typed column (Point, LineString, MultiLineString, Polygon, MultiPolygon, Ring, or Geometry) becomes each feature’s geometry; a column named id becomes the feature id; and any remaining columns become the feature properties, where a lone object-typed column named properties (a JSON, Map, or named Tuple) is written directly as the properties object. A Ring is written as a single-ring Polygon. The new format_geojson_validate_geometry setting (enabled by default) controls whether geometries that are not valid GeoJSON shapes — such as a Polygon with fewer than four points or an unclosed Polygon ring — are rejected when reading and writing. Additionally, the GeoJSON input format now infers the id column as Nullable(String) instead of String, so a feature with an absent or explicit "id": null member is read as NULL and kept distinct from an empty string "". Based on the initial implementation by Mark Needham (#98124). #108065 (Nihal Z. Miaji).
• Added the dotProductTransposed function (alias scalarProductTransposed) that computes the approximate inner product between a QBit column and a reference vector, complementing the existing L2DistanceTransposed and cosineDistanceTransposed functions. #108100 (Alexey Milovidov). Added quantized transposed distance functions cosineDistanceTransposedQuantized, L2DistanceTransposedQuantized and dotProductTransposedQuantized that operate on a QBit(Int8) of quantizeBFloat16ToInt8 Lloyd-Max codes, dequantizing the stored codes on the fly. A floating-point reference vector is the full-precision query, compared at Float32 precision (a Float64 query is narrowed); an Array(Int8) reference is itself dequantized, for a symmetric quantized-vs-quantized distance. #109405 (Alexey Milovidov).
• Added an optional stride parameter to the QBit data type (QBit(T, dimension, stride)) that stores groups of dimensions in separate streams, so a vector search can read only the first dimensions efficiently (e.g. for Matryoshka embeddings). The transposed distance functions accept an optional fourth used_dims argument to read a reduced number of dimensions. #108103 (Alexey Milovidov).
• Added support for the Int8 element type in the QBit data type, enabling storage and transposed-distance vector search (L2DistanceTransposed, cosineDistanceTransposed) over quantized 8-bit integer vectors. #108105 (Alexey Milovidov).
• New function randomHadamardTransform(vector[, seed[, output_dims]]): a deterministic randomized Hadamard transform of a float vector — an orthogonal, norm-preserving rotation useful for preprocessing embeddings before quantization, and (when truncated) as a Johnson–Lindenstrauss / subsampled-randomized-Hadamard random projection. #108227 (Alexey Milovidov).
• Support configuring lazy loading of dictionaries on a per-dictionary basis with the dictionary_lazy_load setting in the dictionary definition, overriding the global dictionaries_lazy_load server setting, so that some dictionaries can be loaded lazily while others are loaded eagerly. #108314 (Miсhael Stetsyuk).
• Added a built-in documentation search page, available at the /docs path of the HTTP interface, that provides instant search over the system.documentation table and renders the reference documentation (with syntax highlighting, math, and cross-links). #108345 (Alexey Milovidov).
• Added function xxHash64Spark, which computes Spark-compatible xxHash64 values for String and NULL inputs using seed 42 and returns Int64. #108436 (Lalit Yadav).
• ALTER USER, ALTER ROLE and ALTER SETTINGS PROFILE now accept SET name = value as an alias for MODIFY SETTING name = value. It changes individual settings in place while keeping the rest, unlike the bare SETTINGS clause which replaces the whole settings list. This makes it less likely to accidentally wipe a user’s other settings. #108722 (Groene AI).
• Support ALTER TABLE ... MODIFY CONSTRAINT [IF EXISTS] name CHECK expr to change the expression of an existing constraint in place. #108768 (Alexey Milovidov).
• Web UI: added a per-column color coding switch (the 🌈 icon in a column header) to toggle between bar, heatmap, categorical, and no visualization. The selected modes are remembered in the URL and browser history. #108873 (Alexey Milovidov).
• Keeper now correctly handles ZooKeeper’s CreateContainer opcode (19), gated behind a new opt-in create_container feature flag (disabled by default, like create_ttl). Once enabled after a full ensemble upgrade, external ZooKeeper clients (e.g. Java ZooKeeper 3.9+) can create container nodes against Keeper instead of getting ConnectionLoss; container nodes report the expected ephemeralOwner sentinel value, and a background GC thread on the leader auto-deletes childless containers. A plain Create request carrying the CONTAINER create-mode flag without opcode 19 continues to be rejected (ZBADARGUMENTS): no known ZooKeeper client emits that combination. This is server-side protocol compatibility only: ClickHouse’s own ZooKeeper client (zkutil) is unchanged and has no API to create container nodes itself. #108908 (unintended).
• Added functions geoToUTM, UTMToGeo, geoToMGRS and MGRSToGeo for converting between WGS84 geographic coordinates and the UTM and MGRS coordinate systems. #108939 (Alexey Milovidov).
• Added the digits(n, offset[, length]) function, which returns the digits of n starting at the 1-based offset and spanning length digits, or to the end of the number if length is omitted. #109012 (Umang Agrawal).
• Add the sqr arithmetic function for calculating the square of a number. #109061 (Lalit Yadav).
• Added a new function dictGetRoot which returns the topmost ancestor (the root) of a key in a hierarchical dictionary. It is a convenient equivalent of dictGetHierarchy(dict_name, key)[-1]. #109459 (Alexey Milovidov).
• Add SYSTEM UNLOAD DICTIONARY and SYSTEM UNLOAD DICTIONARIES commands to release dictionary memory without dropping the dictionary definition. Dictionaries will be reloaded lazily on the next access. #109639 (Matheus Nerone).
• Added functions geometryIntersectCartesian and geometryIntersectSpherical that return whether two geometries intersect. Unlike polygonsIntersectCartesian/polygonsIntersectSpherical, they accept any geometry data type (Point, LineString, MultiLineString, Ring, Polygon, MultiPolygon), including the common Geometry type, and the two arguments may be of different types. #110062 (Alexey Milovidov).
• Query parameters ({name:Type} substitutions) can now be used as setting values, both in the SETTINGS clause of a query (such as SELECT or INSERT) and in standalone SET queries, e.g., SELECT ... SETTINGS max_threads = {threads:UInt64} and SET max_threads = {threads:UInt64}. #108760 (Alexey Milovidov).

​

Experimental Feature

• Wildcard expansion for the url table function and the URL table engine: wildcards in the URL path are expanded by listing HTTP index pages (HTML or plaintext directory listings) and extracting matching URLs, with limits on the index page size and on the number of directories read. Enabled by the allow_experimental_url_wildcard_from_index_pages setting. #95181 (Yue Ni).
• Text indexes can now store token positions (the support_phrase_search index argument) to enable exact phrase matching with direct reads for the hasPhrase function. Requires enabling the MergeTree setting allow_experimental_text_index_phrase_search. #103172 (Elmi Ahmadov). The argument was initially introduced under the name positions. #109900 (Elmi Ahmadov).
• Support manifest file compaction for Iceberg tables via OPTIMIZE TABLE ... MANIFEST (guarded by the allow_experimental_iceberg_compaction setting). Closes #95174. #98178 (Smita Kulkarni). Fix OPTIMIZE TABLE ... MANIFEST for Iceberg tables when manifest partition tuples do not match their partition spec. #111368 (Smita Kulkarni).
• Add a Real Doubles (RD) variant and an optional ALP(AUTO|STD|RD) argument to the experimental ALP codec. Bare ALP now auto-selects between the existing STD scheme and RD instead of always using STD. #99654 (Nazarii Piontko).
• Add an in-memory SLRU cache for deserialized Paimon metadata files (manifest lists and manifests). When enabled via the use_paimon_metadata_files_cache setting, repeated queries against the same Paimon table skip re-downloading and re-parsing metadata from object storage. #104657 (XiaoBinMu).
• Added “drivers” for executable user-defined functions. A driver declared in <user_defined_executable_function_drivers_config> can be used in CREATE FUNCTION name ARGUMENTS (...) RETURNS T ENGINE = DriverName(...) AS '...code...' to compile or otherwise process a user code snippet at function-creation time and produce a runnable executable UDF. The resulting configuration is stored in <dynamic_user_defined_executable_functions_path>, and the originating query is persisted as ATTACH FUNCTION so the function survives server restarts. A proof-of-concept c_function_body driver compiles and runs C function bodies inside sandboxed Docker containers in executable_pool mode. Based on the initial implementation by Daniil Timižev (#77128). #105131 (Alexey Milovidov).
• Serve SELECT count(*) FROM t WHERE col <op> default(col) from per-column sparsity statistics in serialization.json without any data scan when the predicate exactly partitions the column into defaults and non-defaults (under the new experimental setting optimize_trivial_count_with_sparsity_filter, off by default). #105890 (Raúl Marín).
• Prometheus Query API requests to /api/v1/query and /api/v1/query_range are now recorded in system.query_log with read_rows and read_bytes metrics. #106611 (James Cunningham).
• The experimental ReaderExecutor read path (use_reader_executor, off by default) can now hold a remote source connection open and reuse it across sequential reads, reducing the number of object storage requests on scans. #107735 (Sema Checherinda). It now also supports reading encrypted files, with a global encryption-header cache configured by the encryption_header_cache_size server setting and cleared by SYSTEM DROP ENCRYPTION HEADERS CACHE. #109702 (Sema Checherinda).
• Support distributed query-plan reads for SELECT ... FINAL on MergeTree-family tables when make_distributed_plan is enabled. #108148 (Alexander Gololobov).
• Add experimental plan-based parallel replicas execution for MergeTree queries. #108504 (Igor Nikonov).
• Added the experimental Quantize vector codec family and an opt-in two-stage approximate vector-search rewrite over quantized companion streams. #108565 (Shankar Iyer).
• Revive the experimental SZ3 error-bounded lossy compression codec for Float32/Float64 (and arrays of them) columns, originally implemented by Konstantin Vedernikov (#83088). It requires allow_experimental_codecs. #108788 (Alexey Milovidov). Fixed reading data compressed with SZ3('ALGO_LORENZO_REG', ...), which previously failed with CORRUPTED_DATA, and avoided undefined behavior when compressing non-finite floating-point values (NaN and infinities). #110762 (Alexey Milovidov).
• Added the experimental ZXC codec — an asymmetric LZ codec with slow compression and very fast decompression, with a compression ratio between LZ4 and ZSTD. It requires allow_experimental_codecs = 1, and system.compression_codecs reports it as experimental. #110620 (Alexey Milovidov). #111007 (Alexey Milovidov).
• You can now authenticate to OneLake using a pre-obtained bearer token via the onelake_bearer_token setting, instead of a onelake_client_id and onelake_client_secret. This avoids sharing a long-lived client secret. The token is not refreshed, so the database must be recreated once it expires. #109104 (Asya Shneerson).
• AI functions no longer take the named collection as the first positional argument. Credentials come from the ai_function_text_default_credentials / ai_function_embedding_default_credentials settings or from a credentials key in an optional trailing Map(String, String) argument, which also carries tunables (model, max_tokens, temperature, system_prompt, instructions, dimensions). #109232 (George Larionov). The aiEmbed function now takes model as a required positional argument (aiEmbed(text, model[, params])) instead of reading it from the named collection, which ensures reproducibility for embeddings. #110619 (George Larionov).
• Fix data_uncompressed_bytes for skip indices packed into skp_idx.packed (via packed_skip_index_max_bytes): the reported uncompressed size was the compressed size, which could also prevent distributed_index_analysis from activating. #109272 (Raúl Marín).
• Added experimental packed data part storage for MergeTree tables, which stores most of a part’s files in a single data.packed archive instead of a file per stream (projections and a few service files, such as txn_version.txt, are still written separately). It is controlled by the min_bytes_for_full_part_storage, min_rows_for_full_part_storage, and min_level_for_full_part_storage settings and is disabled by default. Once a table writes packed parts, older server versions cannot read them, so enabling this format prevents downgrading the server. #108118 (Raúl Marín).
• Added an experimental table function eval, which evaluates a constant expression to a query string and executes the resulting single SELECT query. The feature is disabled by default and can be enabled with the setting allow_experimental_eval_table_function. Author: Yue Ni. #110132 (Alexey Milovidov).
• Add zstd compression support to the experimental Prometheus remote-write v1 HTTP handler. #110907 (James Cunningham).
• Writing to DeltaLake tables (the allow_experimental_delta_lake_writes setting) was promoted to Beta. #107034 (Kseniia Sumarokova).
• Added ProfileEvents DistributedPlanRemoteTasks, DistributedPlanLocalExecution, and DistributedPlanHostsUsed to observe the execution of experimental distributed query plans (make_distributed_plan). #107985 (Shankar Iyer).
• Removed the experimental RocksDB-based storage in Keeper (experimental_use_rocksdb). A better new on-disk storage is coming soon. #108000 (Michael Kolupaev).
• Distributed index analysis (an experimental feature controlled by the distributed_index_analysis setting) now sends big lists of data part names to remote servers as scalars instead of embedding them into the query text. This avoids possible failures due to the max_ast_elements / max_query_size limits and removes the AST parsing overhead. #110419 (Azat Khuzhin).
• Optimize the experimental timeSeries*ToGrid aggregation functions: buckets are now aligned to both step and window, and the final aggregation is much faster when the window is much larger than the step. The serialization format of the aggregation state was changed in an incompatible way (these functions are experimental). #106724 (Vitaly Baranov).
• External target tables of a TimeSeries table are now registered as referential dependencies: similarly to materialized views, this disallows dropping an external target table before dropping the TimeSeries table that references it (if check_referential_table_dependencies is enabled). #108388 (Vitaly Baranov).
• Add the PromQL function increase. #111023 (Vitaly Baranov).

​

Performance Improvement

• JOINs can now use the primary key index or skip indexes on the left-hand side table to prune granules. Controlled by the setting enable_join_runtime_filters_index_analysis. #109085 (Shankar Iyer).
• Serialization and deserialization of vectors of trivially serializable types now use a single buffer I/O call instead of processing individual elements. #89842 (Jimmy Aguilar Mena).
• Optimize inverse dictionary lookups: a constant equality predicate such as WHERE dictGet(dict, attr, key_expr) = value is now constant-folded directly into a key filter (key_expr = const, key_expr IN [...], or WHERE 0) instead of being rewritten to an IN (SELECT ... FROM dictionary(...)) subquery, and the constant path of dictGetKeys now executes in parallel. #91164 (Nihal Z. Miaji).
• Add a new setting merge_tree_generic_exclusion_search_max_steps that limits the number of steps the generic exclusion search algorithm spends analyzing the primary key index of each data part. The budget is spent on the largest remaining key ranges first, so even a small budget prunes the bulk of the data; ranges that were not fully analyzed are read whole, so query results stay correct but more granules may be read. The limit is applied loosely: it may be exceeded by at most merge_tree_coarse_index_granularity steps, plus one step for each range the part is already divided into (for example, by the query condition cache). The default value 0 means unlimited steps. #92779 (Michael Jarrett).
• Queries with ORDER BY and LIMIT on views over Distributed tables now use the merge-sorted-streams optimization: the outer ORDER BY/LIMIT is pushed into a simple view’s inner query, so each shard sorts its data locally and the coordinator merges pre-sorted streams instead of performing a full sort. #94102 (matanper).
• Enable optimize_or_like_chain by default. Closes #87779. #94517 (Alexey Milovidov).
• Improve the performance of fetching a single user record from system.users when the server has many users. #96699 (Alistair Evans).
• Reading in order with parallel replicas now uses the same logic of splitting the table into max_threads parts as local reading, for better parallelism. #101434 (Nikita Taranov).
• Use partition minmax index bounds to prune more granules during primary key analysis for MergeTree tables, when a primary key column is also an input column of the partition key. For example, in a MergeTree table with ORDER BY (id, event_time) and PARTITION BY toYYYYMM(event_time), ClickHouse will use the partition minmax index on event_time during primary key index analysis to make more informed granule-pruning decisions. Controlled by the new setting use_partition_minmax_for_primary_key_pruning (enabled by default). #103480 (UnamedRus).
• Enable the query condition cache for queries that use Top-K dynamic filtering (ORDER BY ... LIMIT n). The cache entry is partitioned by the Top-K plan parameters, so the same query reuses cached WHERE filter results, while a different LIMIT, sort column, or sort direction produces a fresh entry. #104478 (Alexey Milovidov).
• Added the use_constant_folding_in_index_analysis setting (disabled by default). When enabled, MergeTree primary-key, MinMax, and skip-index analysis fold partition-level constants into the filter predicate separately for each part, improving pruning for filters whose branches depend on partition values, e.g. (a = 1 AND b >= 1) OR (a = 2 AND b > 10) with PARTITION BY a. #104582 (Mikhail Artemenko).
• Push LIMIT into aggregation-in-order to enable early termination when the GROUP BY key matches the ORDER BY key and the table sorting key, significantly reducing the number of rows read. #104859 (Konstantin Bogdanov).
• Updating profile events is now up to 30 times faster, using per-CPU atomics for server-wide and per-user counters. Controlled by the user_profile_events_per_cpu server setting (it uses around 640 KiB per user on 64 cores, so with a very large number of users it may make sense to keep it disabled). #105056 (Azat Khuzhin).
• Enable allow_aggregate_partitions_independently by default. When a GROUP BY key suits the partition key, ClickHouse can aggregate each partition independently and skip the global merging step. Runtime heuristics automatically skip the optimization when the partition layout would make it unfavorable (too few partitions, too many partitions, or significantly skewed partition sizes). #105128 (Alexey Milovidov).
• Vector search queries with rescoring (vector_search_with_rescoring = 1) now compute the exact distance only for the rows returned by the vector index, applying an exact row-position filter instead of brute-force rescoring neighbouring rows from the same MergeTree granule. #105591 (Sergey Kuznetsov). #108846 (Sergey Kuznetsov).
• Improve the performance of the partial_merge join algorithm on FixedString keys by comparing runs of values at once instead of one comparison call per value. #105737 (Artem Zuikov).
• SHOW TABLES and system.tables on data lake catalogs (Iceberg REST, Glue, Unity, Hive, Paimon) now push namespace-bound predicates down to the catalog, avoiding a full catalog scan. Closes #105022. #106029 (Smita Kulkarni).
• Bound untracked memory per CPU to avoid overcommit and out-of-memory errors (the limits are configured with the max_per_cpu_untracked_memory and per_cpu_untracked_memory_thread_buffer server settings). #106055 (Azat Khuzhin).
• Speed up operators that scan a sorted stream for runs of equal key values — DISTINCT in order, LIMIT BY in order, negative LIMIT BY in order, full_sorting_merge and partial_merge joins. #106502 (Nihal Z. Miaji).
• Implemented a native reader and writer for the Arrow and ArrowStream formats that does not use the Apache Arrow library, avoiding extra data copies and conversions. It is now the default (settings input_format_arrow_use_native_reader and output_format_arrow_use_native_writer) and is faster for both reading and writing. #106522 (Alexey Milovidov).
• Re-enable LTO for jemalloc after fixing a thread-cache corruption that appeared with LTO, which improves performance. #106898 (Azat Khuzhin).
• Reduced memory usage and improved the performance of JOINs. The right-hand side of a hash join now uses a compact 8-byte index-based row reference, so the hash-table entries of ALL joins are as small as those of ANY joins for every key type. Benchmarked on large parallel_hash join queries — INNER and ANY joins of 100–300 million row tables on UInt64 and String keys, with both unique and duplicated keys: the median query became about 12% faster and used about 15% less peak memory, with no query becoming slower. INNER joins on UInt64 keys gained the most (median about 21% faster with 38% less memory). #107189 (Harikrishnan Prabakaran).
• Single-column non-Nullable LowCardinality(String) join keys are now natively supported in the hash join. #107264 (Nikita Taranov).
• Added the dpsub join-order enumeration algorithm (dynamic programming over subsets), providing optimal join plans with lower optimization overhead than dpsize and support for non-inner joins. #107351 (Fisnik Kastrati).
• Speed up filesystem cache loading on startup by encoding each cache file’s size in its name (<offset>_<size>), avoiding a stat per file. Loading of cache files written by older versions remains supported. #107415 (Alexey Milovidov).
• Improve the performance of arrayElement on Array(LowCardinality(String)) and of the map LIKE functions on Maps with LowCardinality(String) keys or values, by avoiding unnecessary string materialization. #107450 (Michael Jarrett).
• Avoid converting single-level aggregation to two-level aggregation when memory usage is small, by tracking the memory of the aggregation state itself instead of the memory of the whole query. #107490 (Hechem Selmi).
• Speed up analysis of aggregating queries: the cache key for the aggregation hash-table size statistics is now computed from the query plan instead of building the full query AST. #107643 (Dmitry Novik).
• Improve the performance of INTERSECT ALL and EXCEPT ALL (the default mode for INTERSECT and EXCEPT) by several times, by keying the multiset on the row value instead of hashing each row with SipHash. #107649 (Raúl Marín).
• Rebuild row policy filters without holding the cache lock, improving query scalability under concurrent CREATE ROW POLICY and DROP ROW POLICY. #107917 (Azat Khuzhin).
• Speed up the match, extract, extractAll, replaceRegexpOne and replaceRegexpAll functions for simple regular expressions by compiling them to native code with LLVM. Controlled by the new setting compile_regular_expressions (enabled by default); patterns outside the supported subset transparently fall back to the RE2 engine. #108004 (Alexey Milovidov).
• Sped up ZSTD decompression on AArch64 (ARM) for columns with small match offsets, such as fixed-width integer columns, by vectorizing short-offset overlapping copies with NEON. For example, decompression of a UInt64 column is up to ~2.3x faster on AWS Graviton 4. #108049 (Alexey Milovidov).
• Use the libdeflate library for gzip/zlib/deflate compression and decompression, making it faster (compression ~1.15× with a better ratio, decompression ~1.4–1.5×) for .gz/HTTP/url/s3 data and the Parquet GZIP codec. #108074 (Alexey Milovidov).
• Improved performance of text-search queries combined with a primary key filter. #108114 (Anton Popov).
• Use runtime statistics collected during the first run of a query to optimize subsequent runs: the hash join algorithm can automatically switch to parallel_hash based on the collected statistics. #108125 (Hechem Selmi). The size of join runtime filters is optimized using the hash table size from runtime join statistics, controlled by the new setting join_runtime_filter_size_from_hash_table_stats. #108313 (Hechem Selmi).
• LZ4 decompression speed was improved for fixed-size, low-cardinality columns. #108175 (Nikita Taranov).
• Speed up the default-whitespace trimLeft/trimRight/trimBoth functions (and their aliases ltrim/rtrim/trim): consecutive rows that need no trimming are copied in a single batch, and the per-row space scan runs only for rows that actually have a leading or trailing space. #108177 (Groene AI).
• Improve the performance of replaceAll and replaceRegexpAll when the pattern is a single character and the replacement is a single character (for example replaceRegexpAll(s, ' ', '_')). Such replacements no longer change the string layout, so the column is now copied once and matching bytes are rewritten in place instead of running a per-match search loop. #108178 (Groene AI).
• Improve performance of the pointInPolygon function with a constant polygon. The preprocessed-polygon cache is now keyed on the raw constant arguments, so a constant polygon is parsed only once (on a cache miss) instead of being re-parsed on every input block. #108184 (Groene AI).
• Speed up parsing of the canonical YYYY-MM-DD hh:mm:ss date-time representation in best-effort mode (date_time_input_format = 'best_effort', cast_string_to_date_time_mode = 'best_effort'), which is the default. This recovers a parsing performance regression introduced when those defaults were switched from basic to best_effort. #108187 (Groene AI).
• Avoid maintaining system.predicate_statistics_log selectivity counters on the MergeTree read path when the feature is disabled (predicate_statistics_sample_rate = 0, the default). This removes a per-granule atomic update and an O(rows) filter popcount from every read, recovering a performance regression that was most visible on AArch64. #108190 (Groene AI).
• Speed up parsing of floating-point numbers from text with precise_float_parsing = 1, making it as fast as or faster than the default parser on almost all inputs. Also bumps the bundled fast_float library to v8.2.10. #108205 (Raúl Marín).
• Use the simdutf library for base64 encoding and decoding functions, improving their performance. #108333 (Konstantin Bogdanov).
• Optimize DISTINCT for expensive high-cardinality keys. #108366 (Nihal Z. Miaji).
• On non-Linux builds (macOS and FreeBSD), jemalloc no longer purges dirty pages eagerly (a madvise syscall per free); it now uses a background purging thread with a finite dirty_decay_ms, like on Linux. This significantly speeds up allocation-churning workloads such as recursive CTEs (~2x on a recursive-CTE benchmark on macOS) with negligible RSS impact. #108430 (Alexey Milovidov).
• Speed up query planning under the analyzer for expressions that reference the same WITH alias or repeat the same subexpression many times (for example deeply nested if/multiIf chains), by not rebuilding shared subexpressions when constructing the actions DAG. #108523 (Dmitry Novik).
• Runtime filters are now built on equality keys even when the ON clause also contains non-equality predicates, reducing the number of rows entering the hash join for joins that mix equality and non-equality conditions. LEFT ANTI JOIN is unaffected. #108579 (Antonio Álvarez Caballero).
• Enable -fno-math-errno globally across the codebase, which allows the compiler to optimize math functions better. #108628 (Nikita Taranov).
• Lower the software-prefetch threshold for aggregation and join hash tables from 4 * L2 cache size to L2 cache size, so prefetch is enabled once the hash table no longer fits in L2. This recovers a regression on GROUP BY and JOIN over medium-sized hash tables (~1-8 MiB) on platforms whose reported L2 is large (e.g. AArch64 with a 2 MiB L2). #108655 (Groene AI).
• Improve window function performance. Window queries over wide tables that include SELECT * now use significantly less memory and run faster. Window aggregates, including quantile*, uniq*, and groupArray, are faster with OVER () and OVER (ORDER BY ... RANGE ...). The cume_dist function is also faster, and windows using PARTITION BY or ORDER BY now perform better on columns with many repeated values. #108688 (Nihal Z. Miaji).
• Sped up query analysis for queries with many or very large AND-chains of comparisons: the optimize_and_compare_chain optimization is now bounded by a work budget (new setting optimize_and_compare_chain_max_hash_work) instead of hashing a large fraction of the query tree, and lambda resolution no longer recomputes lambda-body hashes for its recursion guard. #108757 (Alexey Milovidov).
• Fix a performance regression when reading JSON/Dynamic columns with multiple threads: a per-value lookup of the input_format_binary_max_type_complexity setting caused reference-count contention on the shared query context, serializing parallel reads; the setting value is now cached per thread. #108797 (Seva Potapov).
• Fixed a performance regression (introduced in #71781) where reading many small files from object storage via the s3 and other table functions stopped prefetching and fell back to synchronous reads, significantly slowing single-threaded and low-concurrency reads of lots of tiny files. #108872 (Nikita Fomichev). The initial small-object prefetch is now also issued when object storage reads go through the filesystem cache (filesystem_cache_name), so reads of many small files (e.g. S3Queue ingestion) with the cache enabled are no longer synchronous and latency-bound. #109478 (Nikita Fomichev).
• Parse deeply nested array and tuple literals in linear instead of quadratic time. #108892 (Alexey Milovidov).
• Reduce the overhead coming from lock contention in the parallel_hash join algorithm. #108938 (Hechem Selmi).
• Speed up estimateCompressionRatio for T64-encoded columns by computing the compressed size analytically instead of compressing. #109054 (Raufs Dunamalijevs).
• Remove a single-threaded bottleneck in the shuffle step of distributed query plans: each upstream stream now scatters its rows by destination bucket independently instead of funneling the whole data stream through a single transform. #109206 (Alexander Gololobov).
• Reduced Keeper server idle CPU usage. #109255 (Michael Kolupaev).
• Speed up serialization and merges of JSON columns whose shared data contains many sparse paths: when flattening shared data into per-path columns, the per-row scan over all accumulated path columns is removed, and gaps are backfilled in bulk instead of inserting default values one cell at a time. #109341 (Groene AI).
• Reduce CPU overhead of expression evaluation for queries with a large number of columns (e.g. vector search over a QBit column with a small stride, where each vector expands into hundreds of bit-plane sub-columns that are all fed into a single *DistanceTransposed call). #109380 (Alexey Milovidov).
• Improve the performance of H3 geo functions (h3ToGeoBoundary, h3ToGeo, h3CellAreaM2/h3CellAreaRads2, geoToH3) by computing paired sine/cosine together and eliminating redundant trigonometric calls in the coordinate transforms. Results are unchanged (bit-for-bit identical). #109399 (Alexey Milovidov).
• Improve performance of comparisons (<, >, <=, >=) of wide integer types (Int128, UInt128, Int256, UInt256 and types based on them, such as Decimal128) by up to 7 times, and of converting signed integers to Int256 (up to 7 times on mixed-sign data), by making the comparison and sign-extension code branchless. #109474 (Manuel).
• Optimized uniqCombined (including the case of aggregation without keys) and uniqHLL12. #109794 (Anton Popov). #109831 (Anton Popov).
• Functions arraySort and arrayReverseSort are several times faster over numeric arrays (including Decimal and DateTime64) when called without a lambda. #109832 (Manuel).
• Speed up addDays, addWeeks, subtractDays, and subtractWeeks on DateTime and DateTime64 values in fixed-offset time zones (such as UTC) by taking an arithmetic fast path. #109836 (Manuel).
• Optimized analysis of the text index. #109886 (Anton Popov).
• Parallelize processing of non-joined rows in the parallel_hash join algorithm in two more cases: when there is a residual filter, and in the case of small join keys (single-level hash map). #110008 (Hechem Selmi).
• Built-in web UI responses (the Play UI, dashboards, and static-file handlers) are now compressed according to the client’s Accept-Encoding header, using zstd, gzip, deflate, brotli, lz4, xz, snappy, or bzip2. For example, the /play page shrinks from ~350 KB uncompressed to under 120 KB with zstd, reducing load times on slow connections. Already-compressed responses are left untouched, and Vary: Accept-Encoding is emitted for correct CDN and cache partitioning. #110108 (Sayantanu Dey).
• Push a filter below a window function or below LIMIT BY when it is safe: a predicate written above them (an outer WHERE around the subquery, or a conjunct in QUALIFY) that references only the window PARTITION BY columns, or only the LIMIT BY key columns (for LIMIT n BY with n >= 1 and no OFFSET), now reaches storage and enables primary key and partition pruning, skip indexes, and projections. For example, SELECT ... row_number() OVER (PARTITION BY key ORDER BY ts) AS rn ... QUALIFY rn = 1 AND key = 'x' no longer reads and windows the whole table. #110114 (Groene AI). #110116 (Groene AI).
• With join_use_nulls = 1, a null-rejecting WHERE that references columns from both sides of an outer join (e.g. WHERE l.k = 42 AND r.k = 42) now converts the join to INNER (or LEFT/RIGHT) and prunes the primary key, the same as with join_use_nulls = 0. Previously the conversion was skipped in this case and both tables were read in full. #110121 (Groene AI).
• Functions with a single non-const Nullable argument now share the argument’s null map with the result instead of allocating and merging a new one. #110151 (Manuel).
• Avoid a heap allocation per hyperrectangle check in primary key index analysis, speeding up mark filtering by 5-26% depending on the query shape. #110153 (Manuel).
• Improve performance of functions arrayMin and arrayMax over numeric arrays by ~1.3-1.5x by using a vectorized reduction instead of a per-element comparison loop. #110163 (Manuel).
• Vectorize decompression of the Delta codec. Decoding is now 1.5–5 times faster for 8/16/32-bit data types, making scans of Delta-compressed columns up to 20% faster. #110189 (Manuel).
• Speed up string search functions (like, position, match, countSubstrings, hasToken, etc.) over Enum columns with a constant needle by searching only the distinct enum names and mapping the results back per row, instead of searching every row. #110325 (Alexey Milovidov).
• Speed up aggregate functions groupBitOr/groupBitAnd/groupBitXor and the variance family (varPop, varSamp, stddevPop, stddevSamp, skewPop, skewSamp, kurtPop, kurtSamp, covarPop, covarSamp, corr) with vectorized batch processing: up to 4x faster with the -If combinator on unpredictable conditions, and up to 3x faster for the variance family without it. #110461 (Manuel).
• Speed up the analysis of the text index (the stage that selects which granules to read): use exclusion search instead of iterating over all granules, and avoid creating a temporary bitmap for each range check. #110530 (Anton Popov).
• Account for the on-disk size of Map subcolumns in the automatic PREWHERE optimization. This fixes a performance regression in PREWHERE introduced in #99200. Part of #110462. #110623 (Pavel Kruglov).
• Speed up parsing of UUID values from text (e.g. JSONExtract into LowCardinality(UUID), toUUID, CAST AS UUID) by validating and converting hex digits in a single pass instead of two. #110625 (Groene AI).
• Significantly improve performance of the timeSeries*ToGrid aggregate functions by using faster sorting and hash table implementations. #110875 (Nikita Mikhaylov).
• Queries like SELECT * FROM t WHERE id now use index skipping on the id column. Closes #89222. #89603 (Aditya Chopra).
• Added the query_plan_merge_expression_into_join setting (enabled by default) to allow merging expression steps into JOIN steps during the join reordering optimization. This enables join reordering across subqueries that wrap joins (e.g., when a JOIN is inside a subquery with computed columns), leading to better optimization of complex join trees. #98533 (Vladimir Cherkasov).
• Optimize AND chains with multiple comparison conditions on the same expression: detect contradictions (e.g., a < 3 AND a > 5 → false) and prune redundant conditions (e.g., a = 3 AND a < 5 → a = 3). #99736 (Xiaozhe Yu).
• Materialized CTEs referenced from multiple branches of a UNION query are now materialized once and shared across all branches. Previously, each branch received its own copy of the CTE, which was inlined and evaluated separately. #102107 (Dmitry Novik).
• Threads in BackgroundSchedulePool are now created lazily on demand instead of all at once at server startup. A new server setting background_schedule_pool_initial_size (default 16) controls how many workers are pre-spawned; the pool grows up to background_schedule_pool_size when needed. This reduces the number of idle threads on lightly-loaded servers. Closes #85265. #105066 (Alexey Milovidov).
• Skip unnecessary mark file loads for the JSON type’s advanced shared data serialization. #107051 (Pavel Kruglov).
• The cache of preprocessed constant polygons for the function pointInPolygon is now shared between queries and threads and bounded by the new server setting point_in_polygon_cache_size (default 256 MiB, 0 disables the cache, changeable at runtime). Previously, the cache kept up to max_threads copies of each preprocessed polygon, grew without bound across distinct polygons, and was never released until server restart. It can be cleared with SYSTEM DROP POINT IN POLYGON CACHE, and the current usage is reported in system.metrics as PointInPolygonCacheBytes, PointInPolygonCacheCells, and PointInPolygonCacheSizeLimit. Closes #106393. #107247 (Nihal Z. Miaji).
• Keeper no longer holds its internal snapshots lock while serializing and writing a snapshot to disk or moving snapshots between disks, so snapshot transfers to other nodes and Raft processing are no longer stalled by local snapshot creation. #107595 (Antonio Andelic).
• Parallelize the filesystem cache background eviction (controlled by the keep_free_space_ratio settings). Added a new setting keep_free_space_eviction_threads to control the parallelism, and increased the default of keep_free_space_remove_batch from 100 to 250. #108147 (Kseniia Sumarokova).
• Limit the number of concurrently staged parts in the Azure Blob Storage read-then-write copy (used by backups when native copy is unavailable) to max_inflight_parts_for_one_file, preventing excessive memory usage when many large files are copied in parallel. #108232 (Smita Kulkarni).
• Enabled all three text index caches globally — previously, they were only enabled within queries. Also, the posting lists cache size is now zero, which effectively disables it again, because posting lists are large and caching them is too costly. #108274 (Robert Schulze).
• Add streaming in-memory marks compression during marks loading to reduce peak memory usage. Closes #108285. #108325 (Pavel Kruglov).
• The new filesystem cache setting reserve_granularity (default 4 MiB) reserves space ahead in coarser granules, reducing lock contention on the filesystem cache space reservation hot path. #108369 (Kseniia Sumarokova).
• Reduce the amount of memory used by the filesystem cache metadata (per file segment and per key). #108477 (Kseniia Sumarokova).
• Queries to system.iceberg_history with WHERE filters on the database and table columns now access only the matching databases, which can save time if you have many unrelated remote databases on the server. #108492 (Den Kalantaevskii).
• Removed unnecessary locking in the filesystem cache to reduce lock contention. #108932 (Kseniia Sumarokova).
• Reduced memory usage of BACKUP and RESTORE metadata handling. Opening a backup (for RESTORE, or as the base of an incremental BACKUP) now parses the .backup metadata as a stream instead of loading it into an in-memory XML document tree, which for large (especially incremental) backups avoids allocating a multi-gigabyte DOM tree. #109107 (Julia Kartseva). Finalizing a BACKUP no longer copies the file infos of all files into a temporary vector, which, for backups of millions of files, transiently cost several gigabytes. #109861 (Julia Kartseva). Writing backup entries also no longer copies the current host’s file-info list. #111162 (Julia Kartseva).
• Reduce memory consumption in the aggregator. #109224 (Konstantin Vedernikov).
• Support the query condition cache for local Parquet files read via the File table engine (previously supported only for object storage and data lakes). #109247 (Alexey Milovidov).
• Reduced the memory usage of the text index header cache by 2-2.5 times. More headers of data parts fit into the cache (setting text_index_header_cache_size), reducing disk reads for text search queries. #109332 (Anton Popov).
• The transposed vector distance functions L2DistanceTransposed, cosineDistanceTransposed and dotProductTransposed now apply the partial bit-plane read optimization to Nullable(QBit) columns, reading only the requested bit planes instead of the whole column. #109358 (Alexey Milovidov).
• IS NOT DISTINCT FROM and IS TRUE now use primary-key and minmax indexes to prune granules, the same as =. Previously k IS NOT DISTINCT FROM 42 and (k = 42) IS TRUE scanned all granules. #110006 (Groene AI).
• Pooled connections are no longer pinged before each use. This removes a Ping-Pong round trip that was added to every reused connection, reducing the latency of distributed queries and of clickhouse-benchmark. A stale pooled connection is detected with a zero-timeout poll (a non-blocking check that adds no round trip) and recovered by reconnecting. #110068 (Alexey Milovidov).
• Fixed spurious reconnections of pooled HTTP keep-alive connections over TLS (HTTPS, S3). The stale-connection check used poll on the socket, which misreports a live secure connection carrying an unread TLS post-handshake record (a session ticket or KeyUpdate) as closed; it now uses a non-blocking MSG_PEEK on plain sockets and SSL_peek/SSL_has_pending on TLS sockets, which also correctly detects an orderly TLS shutdown and a TLS record that has only partially arrived. #110402 (Alexey Milovidov).
• Queries accelerated by TopK dynamic filtering (ORDER BY ... LIMIT k) now make fuller use of the query condition cache: the cache is populated even when lazy materialization is applied to the query, and such queries can reuse entries previously written by an ordinary query that had the same WHERE predicate. #110507 (Shankar Iyer).
• A LIKE/NOT LIKE pattern without wildcards (%, _) now uses an exact primary key range, so it reads the same number of granules as the equivalent =/!= predicate instead of a wider prefix range. #107077 (Groene AI).
• Fix redundant file segment lock contention (a performance regression introduced in 26.1) when reading data through the filesystem cache. #109577 (Kseniia Sumarokova).
• Fixed skip indexes defined on Tuple subcolumns not being used when the field is accessed via tupleElement(t, 'name'), tupleElement(t, N), or t.N while the full tuple is also read in the same query (e.g. SELECT *). All of these forms now prune granules the same way as the named subcolumn access t.name. #110056 (Groene AI).
• Fixed slow ORDER BY ... LIMIT queries on Distributed tables when prefer_localhost_replica selects a local replica: the preliminary limit was removed from the local-shard plan, preventing lazy materialization and dynamic top-K filtering. #110136 (Michael Jarrett).
• Fixed the text index not being used for the ILIKE operator when the index is defined over an expression rather than a bare column (for example assumeNotNull(col) with a lower(...) preprocessor): such queries were reading all granules instead of using the index. This closes #110350. #110595 (Elmi Ahmadov).
• Reduce lock contention in the filesystem cache: queries reading from the cache no longer block on the priority queue lock when updating an entry’s LRU priority. #109065 (Antonio Andelic).
• Text index queries that use a filter-only postprocessor expression (for example if(..., '', token)) now discard empty tokens up front instead of materializing them row by row, improving performance. #109049 (Elmi Ahmadov).
• Reduce lock contention in system log queues, lowering query tail latency on busy clusters. #110539 (Sean Haynes).
• Improved the performance of substring search functions (position, countSubstrings, multiSearch*, replaceAll, replaceOne, and their case-insensitive and UTF-8 variants) with a non-constant needle by up to 1.5x by using SIMD-backed string searchers on the per-row path instead of a naive implementation. #110580 (Raúl Marín).
• Lazy materialization is now applied to ReplacingMergeTree queries with FINAL, a filter, and a small LIMIT even without ORDER BY. Columns not needed by the filter and the FINAL merge are fetched only for the rows that remain after the limit, which makes such queries much faster when they select wide columns with a selective filter. Controlled by query_plan_optimize_lazy_materialization and query_plan_max_limit_for_lazy_materialization. #110722 (Nikolai Kochetov).
• Improved async insert performance when deduplication is enabled and one flush is split across multiple partitions: ClickHouse now reuses the computed deduplication hashes across partitions instead of recomputing them for each one. #111150 (Valery Petrov).

​

Improvement

• Added a setting input_format_csv_missing_nullable_as_empty_string (disabled by default). When enabled, a missing value of a Nullable(String) column in CSV input is read as an empty String instead of NULL, regardless of input_format_csv_empty_as_default. #58225 (kevinyhzou). #107577 (Alexey Milovidov).
• Functions toDateOrNull, toDateTimeOrNull and toDateTime64OrNull now accept integer arguments of all native integer types (interpreted the same way as by toDate, toDateTime and toDateTime64, with an optional timezone argument), returning NULL for values out of range of the result type. For example, toDateTimeOrNull(1583851242, 'Asia/Shanghai') returns 2020-03-10 22:40:42 and toDateTimeOrNull(4294967296) returns NULL. #79791 (Jitendra).
• Added an is_wildcard column to the system.grants table that indicates whether a grant uses wildcard prefix matching (e.g., GRANT SELECT ON db*.*). Previously, wildcard and exact grants on the same name were indistinguishable in system.grants. Closes #92835. #98577 (DQ).
• Allow Snappy compression in the HTTP interface (Accept-Encoding: snappy) and add the snappy_mode setting to choose between the Hadoop Snappy block format and the Snappy framing format for generic file/url snappy I/O. #100752 (Alexey Milovidov).
• Support compatibility with the old analyzer under the analyzer_compatibility_allow_non_aggregate_in_having setting. If enabled, non-aggregate conjunctions are moved from HAVING to WHERE. #104232 (Dmitry Novik).
• Added a new server setting memory_worker_rss_speculative_reserve_ratio (default 1.0) which makes the global memory tracker speculatively reserve memory on top of the observed RSS when the RSS growth outpaces the tracker’s bookkeeping between samples. With it, allocations get MEMORY_LIMIT_EXCEEDED earlier, and the kernel OOM-killer is less likely to fire first. Set the ratio to 0 to disable the speculation. #104976 (Alexey Milovidov).
• Added opt-in Prometheus metrics for filesystem cache eviction activity (filesystem_cache_evictions_total, filesystem_cache_evicted_bytes_total, filesystem_cache_evicted_segment_hits, filesystem_cache_evicted_segment_size_bytes, and their per-user variants labeled with user_id), exposed per cache via system.dimensional_metrics, system.histogram_metrics, and the Prometheus endpoint. They are controlled by the cache disk-config settings expose_prometheus_eviction_metrics and expose_prometheus_eviction_metrics_per_user (both off by default), which can be toggled at runtime via SYSTEM RELOAD CONFIG. #105020 (Sacheendra Talluri).
• Make EXPLAIN [PLAN] actions=1, compact=1, pretty=1 the default. #105036 (Kirill Kopnev).
• Bump delta-kernel-rs (the library behind the DeltaLake integration) to v0.23.0. #105861 (Smita Kulkarni).
• Add du and wc commands to clickhouse-disks. They, respectively, print the total size in bytes for a given file or directory, and list the number of bytes, lines, and words in a file. #106268 (Asya Shneerson).
• Added a new asynchronous metric UntrackedMemory (visible in system.asynchronous_metrics) that reports memory already allocated by threads but not yet accounted in the global memory tracking counter: each thread accumulates small allocations locally and reports them in bulk. This helps explain discrepancies between MemoryTracking and the process’s actual memory usage. MEMORY_LIMIT_EXCEEDED error messages now also include the amount of untracked memory, making it easier to understand why a query or the server hit its memory limit. #106386 (Miсhael Stetsyuk).
• Upgrade ClickStack (the observability UI) to version 2.28.0. #106406 (Aaron Knudtson).
• Added support for BFloat16 in dotProduct and improved its performance by batching the SIMD path. #106569 (Nikita Taranov).
• Made the settings max_named_collection_num_to_throw, max_table_num_to_throw, max_replicated_table_num_to_throw, max_view_num_to_throw, max_dictionary_num_to_throw, and max_database_num_to_throw changeable without a server restart. #106821 (Maxim Orlovsky).
• Support the REDUCED_REDUNDANCY, STANDARD_IA, ONEZONE_IA, GLACIER_IR, and EXPRESS_ONEZONE values (in addition to STANDARD and INTELLIGENT_TIERING) for the s3_storage_class_name setting. #107251 (Aditya Kumar).
• Added the create-time materialized_postgresql_use_extended_date_and_time_types setting for the MaterializedPostgreSQL database engine. By default (enabled), PostgreSQL date/timestamp columns are inferred as Date32/DateTime64; setting it to 0 at CREATE DATABASE time infers the narrower Date/DateTime types. The setting is not applicable to the MaterializedPostgreSQL table engine. #107428 (Alexey Milovidov).
• The SOME / ALL array quantifier (expr OP SOME(array) / expr OP ALL(array)) now also supports the keyword comparison predicates IS DISTINCT FROM and IS NOT DISTINCT FROM, and the string-search predicates LIKE, ILIKE, NOT LIKE, NOT ILIKE, and REGEXP, rewritten to arrayExists / arrayAll. #107454 (Alexey Milovidov).
• Fixed a memory leak that occurred when opening a SQLite database failed (for example, when the sqlite table function or the SQLite database engine is given a path that cannot be opened). #107807 (Alexey Milovidov).
• In the Web UI (play.html), the run shortcut hint now shows Cmd+Enter on Mac and Ctrl+Enter on other platforms instead of always showing Ctrl/Cmd+Enter, and the , +Shift to run all part of the hint is shown only when there are multiple queries. #107817 (Alexey Milovidov). #108957 (Alexey Milovidov).
• The Alias table engine now supports reading with parallel replicas when the target table is of the MergeTree family. #107830 (Kai Zhu).
• Added the uniq_v2 column statistics type — a lightweight alternative to the uniq statistics, based on the uniqCombined64 sketch. #107863 (Han Fei). Deprecated the minmax column statistics type and changed the default value of auto_statistics_types to basic, uniq_v2. #108680 (Han Fei).
• AI functions (aiGenerate, aiEmbed, aiClassify, aiExtract, aiTranslate) now retry transient network failures (connection resets, TLS connect failures, timeouts, unreachable addresses) when ai_function_max_retries is set, matching the retry behavior of the url table function. Previously, only provider-side HTTP error responses were retried. #107927 (Alexey Milovidov).
• INSERT into a MergeTree table now honors query cancellation and max_execution_time while writing many parts, instead of potentially running long after being killed. #107929 (Michael Kolupaev).
• In the Web UI (play.html), the Web Terminal icon can now be opened in a new browser tab with a middle-button click or Ctrl/Cmd/Shift+click. #108006 (Alexey Milovidov).
• AI SQL generation in the client no longer sends the temperature parameter unless it is explicitly set via ai.temperature in the configuration. This fixes AI SQL generation for models that reject the temperature parameter. #108014 (Alexey Milovidov).
• Added the system.masking_policies table and SHOW MASKING POLICIES introspection to open-source builds. Masking policies themselves remain a ClickHouse Cloud feature, so the table is empty in open-source builds, but introspection queries no longer fail with an error. #108030 (Alexey Milovidov).
• The Keeper-based Kafka consumer now randomizes the order in which it acquires temporary partition locks across replicas, improving fairness of partition distribution and reducing lock contention. Previously, the intended shuffle had no effect. #108033 (Alexey Milovidov).
• Added autocompletion to the Web UI (play.html), based on system.completions and the WASM-based SQL lexer. #108059 (Alexey Milovidov).
• clickhouse-client and clickhouse-local now show as-you-type autocompletion hints (inline “ghost” text) for the best matching suggestion when the cursor is at the end of the input. The most relevant suggestions (recently used and identifiers already present in the query) are ranked first. Navigate with Up/Down (or Ctrl-Up/Ctrl-Down); accept a single or selected hint with Tab or Right, and accept a selected hint with Enter; Tab also opens the classic completion list. Controlled by the new --hints option (on by default; requires --highlight). #108070 (Alexey Milovidov).
• Support the length function for the QBit data type — it returns the dimension of the vector as a constant. #108071 (Alexey Milovidov). Support CAST from a QBit to an Array, reconstructing the original vector — the inverse of the existing Array to QBit conversion. #108072 (Alexey Milovidov).
• Requests to REST data lake catalogs (such as OneLake) now include a ClickHouse User-Agent header. #108117 (Konstantin Vedernikov).
• arrayFold now respects query cancellation and max_execution_time. Previously, a fold over a very long array ran entirely inside a single function call and could not be interrupted, so KILL QUERY and time limits were ignored until the fold finished. #108192 (Groene AI).
• MVTEncodeGeom now snaps geometry to the integer pixel grid before clipping and clips polygons with the wagyu library, so the clipped output is valid (self-intersecting rings are repaired) and edge-aligned, matching PostGIS ST_AsMVTGeom. #108248 (Saarthak Gupta).
• Add per-phase query pre-execution ProfileEvents: QueryParseMicroseconds, QueryAnalysisMicroseconds, QueryPlanBuildMicroseconds and QueryPipelineBuildMicroseconds. They expose where time is spent before query execution (parsing, analysis, query plan building, pipeline building) and are available in system.query_log and system.events. #108282 (Jordi Villar).
• ALTER TABLE operations that would produce table metadata exceeding max_query_size are now rejected upfront, preventing tables from becoming unloadable by components such as DDL distribution and replica recovery. #108283 (Andrew Kravchuk).
• Added settings and engine_settings columns to system.backups and system.backup_log. settings exposes the backup/restore-specific settings requested for an operation (e.g., allow_s3_native_copy, deduplicate_files, structure_only), and engine_settings exposes the settings effectively used by the backup engine’s reader and writer (e.g., the S3 request settings such as allow_native_copy, which may differ from what was requested after merging the endpoint configuration). This makes it possible to see which settings a BACKUP/RESTORE operation actually ran with. #108334 (Julia Kartseva).
• Added a source column to system.documentation containing the path of the source file where each entity’s documentation is defined. The table now also documents compression codecs, profile events, current metrics, asynchronous metrics, and the system tables themselves (with their columns), and the documentation of settings now includes their type and default value. #108346 (Alexey Milovidov). An empty setting default value is now rendered as empty string instead of empty backticks. #108708 (Alexey Milovidov).
• Support asynchronous reads from remote replicas (async_socket_for_remote) on macOS by implementing the epoll-based polling primitives on top of kqueue. This lets distributed queries read shards in parallel on macOS instead of serially. #108403 (Raúl Marín).
• Support BFloat16 in binary math functions. #108442 (Zhang Yifan).
• MergeTree can now read a compressed stream whose blocks use different codecs. This is the read-side prerequisite for adaptive codec selection (#105404). #108592 (Raufs Dunamalijevs).
• The automatic value of max_threads and similar settings is now shown in system.settings as auto(8) instead of 'auto(8)'; the surrounding single quotes were a long-standing artifact baked into the value. Cross-version compatibility is preserved: the legacy quoted form is still accepted when parsing settings received from older servers. #108657 (Alexey Milovidov).
• Several usability fixes for the /schema Web UI: load the schema on Enter, theme-aware scrollbars, readable table names, consistent per-database grouping of independent tables, no text selection while dragging, and a play-style authentication form (Credential Management API, URL-supplied credentials with password stripping, and live credential validation). #108724 (Alexey Milovidov).
• The text index lazy posting-list apply mode is no longer experimental and can be selected with text_index_posting_list_apply_mode = 'lazy' without allow_experimental_text_index_lazy_apply. The density-threshold setting was renamed from text_index_density_threshold to text_index_lazy_intersection_density_threshold. #108814 (Anton Popov).
• Add a delta_sharing catalog type to DataLakeCatalog for Databricks Delta Sharing’s Iceberg REST endpoint, whose namespaces are flat and which ignores the parent listing filter. Use it instead of catalog_type = 'rest' for such endpoints, where SHOW TABLES would otherwise hang. #108865 (Seva Potapov).
• Added a new clickhouse-client/clickhouse-local option --echo-query-separator that prints a custom separator before the formatted echoed query, making it easier to tell the typed query apart from its reformatted echo. Disabled by default. #108888 (David Meng).
• Improvements to the database panel of the Web UI: the number of tables is shown in parentheses next to a database name when it is expanded. #108891 (Alexey Milovidov). A refresh button next to an expanded database reloads its list of tables. #108958 (Alexey Milovidov). When the panel is reopened, the previously expanded databases are restored. #108964 (Alexey Milovidov).
• Allow calculating the combined skip-index benefit in EXPLAIN WHATIF: it shows the data ratio for the intersection of all existing suitable hypothetical indices. #108934 (Yarik Briukhovetskyi).
• Fixed the horizontal scroll position of the result table being reset in the Web UI when clicking a link in a cell and returning to the page. #108941 (Alexey Milovidov).
• The MySQL database engine, table engine, and table function now map MySQL’s spatial column types (LINESTRING, POLYGON, MULTILINESTRING, MULTIPOLYGON, and the generic GEOMETRY) to the corresponding ClickHouse geometric types instead of String. This is controlled by the new geometry flag of the mysql_datatypes_support_level setting, enabled by default. POINT is still always converted to Point. The generic GEOMETRY column maps to the umbrella Geometry type; reading a value whose subtype has no ClickHouse counterpart (MULTIPOINT, GEOMETRYCOLLECTION) throws an exception at read time. #108944 (Alexey Milovidov).
• Web UI (play.html): full keyboard navigation — arrow-key navigation in the database panel and result table, keyboard access to the toolbar buttons, theme switcher, and download menu, and visible focus outlines. #108972 (Alexey Milovidov).
• HTTP interface: when credentials are provided both via URL parameters (user/password) and an Authorization header, the URL parameters now take precedence instead of the request being rejected. This fixes downloading results from the Web UI (play.html) failing with AUTHENTICATION_FAILED (error 516) when the browser has remembered Basic credentials. #108980 (Alexey Milovidov).
• Restore show_data_lake_catalogs_in_system_tables as the setting that controls only DataLakeCatalog visibility in system.tables, system.columns, and system.completions. Added show_remote_databases_in_system_tables, enabled by default, to let users hide MySQL and PostgreSQL databases from those system tables separately. #109082 (Pablo Marcos).
• Use the precise (closest-representable) float parsing algorithm by default and apply the precise_float_parsing setting to input formats (CSV, TSV, JSON, VALUES, …) and numeric literals, not just toFloat*/CAST. Set precise_float_parsing = 0 for the previous, faster in some cases but less accurate, behavior. Closes #60146. Closes #74647. Closes #68914. #109086 (Raúl Marín).
• A query with a single CTE is now formatted with the same newline and indentation as a query with multiple CTEs. Previously WITH a AS (...) kept the CTE on the same line as WITH, while two or more CTEs put WITH on its own line with each CTE indented. #109092 (Groene AI).
• Web UI: when the highlighted completion is the same as the already-typed word, or the same word in a different case (e.g. HASH offered for a typed hash), pressing Right/Tab/Enter no longer rewrites the word or gets swallowed by the autocompletion — the key moves the caret or inserts a newline as expected. #109106 (Alexey Milovidov). #109384 (Alexey Milovidov).
• Added rainbow parentheses, matched-bracket, matching-identifier, and digit-group highlighting to the Web UI (play.html), matching clickhouse-client. #109108 (Alexey Milovidov).
• In the Web UI, Tab and Shift+Tab now indent and unindent the selected lines by 4 spaces. #109110 (Alexey Milovidov).
• Web UI: highlight the position of a syntax error in the query editor with a red background. The highlight is cleared as soon as the user returns focus to the editor. #109112 (Alexey Milovidov).
• clickhouse-compressor --stat and default-codec detection for old parts now report corrupted block headers as corruption instead of a misleading end-of-file error. #109157 (Raufs Dunamalijevs).
• ClickHouse Keeper now respects the os_collect_psi_metrics setting and skips PSI metrics collection when it is disabled. #109179 (Maxim Orlovsky).
• In the Web UI, when query tabs are shown, the connection parameters (host, user, password) are hidden behind a key button in the top-right corner and shown in a drop-down on demand. #109243 (Alexey Milovidov).
• Web UI: show a single-value result (one row, one column, e.g. SHOW CREATE TABLE) with full height instead of clamping it to three lines. #109245 (Alexey Milovidov).
• In the Web UI, clicking a table in the databases panel with the middle mouse button, with Shift, or with the platform’s new-tab modifier (Cmd on macOS, Ctrl elsewhere) now opens its query in a new Web UI tab. #109246 (Alexey Milovidov).
• Key the query condition cache for remote (object storage) Parquet files by ETag in addition to the path, so that overwriting an object in place no longer serves stale cached results. #109310 (Alexey Milovidov).
• In the Web UI, empty strings and NULLs are no longer colorized in the categorical coloring mode. #109342 (Alexey Milovidov).
• Web UI: clicking a table’s icon in the database panel now shows the list of its columns, with type icons, a size bar proportional to the compressed size, and a tooltip with the compressed/uncompressed size and compression ratio. #109346 (Alexey Milovidov).
• Web UI: show the full tab title as a tooltip on hover when it is truncated. #109354 (Alexey Milovidov).
• In the Web UI, middle-clicking a tab’s title (or Ctrl/Shift/Cmd+clicking it) now duplicates the tab. #109357 (Alexey Milovidov).
• Web UI: resize the final progress and query statistics areas uniformly so they do not wrap awkwardly when the browser window is narrow. #109363 (Alexey Milovidov).
• Web UI: selecting a database in the database panel now uses it as the default database for queries run from the editor. #109372 (Alexey Milovidov).
• Support reinterpret of an Array of fixed-size elements as a String, the inverse of the existing reinterpret of a String/FixedString as an Array. #109383 (Alexey Milovidov).
• In the Web UI, do not display the per-column color-coding toggles when the result has no more than a single row. #109385 (Alexey Milovidov).
• Allow CAST between QBit types that differ in the element type and/or the stride, as long as the dimension stays the same (for example CAST(x AS QBit(Float64, N)) from a QBit(Float32, N)). Stride-only changes are lossless; element-type changes follow the corresponding Array conversion semantics (for example Float32 to BFloat16 may lose precision, and accurateCast / accurateCastOrNull reject rows that are not exactly representable). #109387 (Alexey Milovidov).
• Functions quantizeBFloat16ToInt8 and dequantizeInt8ToBFloat16 now also accept Array and QBit arguments, applying the Lloyd-Max codec to the whole vector (returning Array/QBit of the corresponding element type), in addition to the existing scalar overloads. #109398 (Alexey Milovidov).
• Web UI: the Documentation link now carries the current user name in the URL. #109419 (Alexey Milovidov).
• Support arraySum, arrayAvg, and arrayProduct for arrays of BFloat16. #109420 (Alexey Milovidov).
• The MySQL-style format specifier %f in parseDateTime / parseDateTime64 (and their OrZero / OrNull variants) now accepts between 1 and 6 fractional digits, interpreted as left-aligned microseconds like MySQL’s STR_TO_DATE, instead of requiring exactly 6. Also fixed misaligned PrettyCompact tables in the built-in function documentation examples. #109421 (Alexey Milovidov).
• Web UI: query tabs are now persistent — each tab keeps its rendered result (including images and charts) and its running query when you switch away and back, and long queries continue running in the background. Tab titles show a spinner, progress bar, and completion check-mark for their query. #109425 (Alexey Milovidov).
• Web UI: correctly display an error when a query fails after some of its result has already been streamed, instead of showing a client-side SyntaxError: Unexpected end of JSON input (or, with http_write_exception_in_output_format, marking the failed query as successful). #109430 (Alexey Milovidov).
• Added pinned columns to the Web UI: a column can be pinned from its header so that it stays visible (stuck to the edge) while the results table is scrolled horizontally. #109439 (Alexey Milovidov).
• In the advanced dashboard (/dashboard), charts for specific metrics can now be added directly by the metric name as it appears in the source code or documentation, which makes it easier to add multiple metrics during debugging sessions. #109449 (Mikhail Artemenko).
• Web UI: fixed the connection settings drop-down closing when selecting text with the mouse and dragging past its border. #109451 (Alexey Milovidov).
• Web UI: hovering over a result table column header now shows the full column name and type as a tooltip, so truncated headers can be read in full. #109463 (Alexey Milovidov). Fixed column header titles being clipped with an ellipsis even when the column was wide enough to show the full title. #110397 (Alexey Milovidov).
• Support SSD cache dictionaries (SSD_CACHE layout) and the FileLog table engine on macOS builds. #109493 (Raúl Marín).
• Web UI: do not restore tabs with empty queries when the page is opened. #109529 (Alexey Milovidov).
• Fix unreadable as-you-type autocompletion hints and bright colors in the interactive client on terminals whose TERM does not contain 256 (e.g. Ghostty, kitty, Alacritty, foot): bright colors are now emitted as aixterm bright color codes unconditionally instead of falling back to bold + dark color, which modern terminals render as a dark, hard-to-read color on dark backgrounds. #109622 (Alasdair Brown).
• Enable jemalloc per-CPU arenas and allocation profiling on macOS builds. #109684 (Raúl Marín).
• Allow altering some authentication settings of a OneLake data lake catalog database with ALTER DATABASE ... MODIFY SETTING. #110019 (alesapin).
• The Hive engine now reads ORC file metadata (min/max indexes, row counts) with ClickHouse’s native ORC reader instead of the Apache Arrow ORC adapter. #110086 (Alexey Milovidov).
• Reduced the binary size by ~10.5 MB by executing comparison and arithmetic operations on rarely used mixed type pairs (Decimal vs integer of a different width, and pairs involving Int128/UInt128/Int256/UInt256) via a conversion to a common type instead of a dedicated compiled kernel for every combination of types. Same-type pairs, commonly used pairs, and the memory-bound plus/minus/multiply keep their dedicated kernels; results, result types and exceptional cases are unchanged. #110131 (Alexey Milovidov).
• A subquery on the right side of IN whose single column is an array one dimension deeper than the left argument is now interpreted as the set of the array’s elements (like an array literal or an array-returning function), instead of failing with a confusing type-mismatch error. For example, x IN (SELECT groupArray(x) FROM ...) now works. #110169 (Alexey Milovidov).
• Reading from a SQLite table engine or sqlite table function no longer busy-spins a full CPU core when the SQLite database is locked by another connection. The read now idles while waiting for the lock and stays cancellable. #110248 (Groene AI).
• Round elapsed time, rate, and ratio values in log and exception messages to three digits, so numbers like 1.345844286 sec. are no longer printed at full precision. #110277 (Alexey Milovidov).
• In the Web UI (Play), BFloat16 is now recognized as a floating point number and its columns are right-aligned and colorized accordingly. #110433 (Alexey Milovidov).
• Support the SETTINGS clause for the PostgreSQL table engine and the postgresql table function (for example SETTINGS postgresql_connection_pool_size = 50), bringing feature parity with the MySQL engine. #110614 (Alexey Milovidov).
• SHOW CREATE TABLE (and SHOW CREATE VIEW / SHOW CREATE DICTIONARY) now suggests a similarly-named table in the error message when the requested table does not exist, the same way SELECT queries do. #110633 (Alexey Milovidov).
• Enable the merge_selector_enable_heuristic_to_lower_max_parts_to_merge_at_once setting by default: the merge selector now automatically lowers the maximum number of parts to merge at once based on how full the partition is. See #91163. #110726 (Mikhail Artemenko).
• Accept PostgreSQL cleanup commands RESET, UNLISTEN, and DISCARD as no-ops in the PostgreSQL wire protocol instead of failing them with a syntax error. This improves compatibility with drivers such as Skunk that send RESET ALL and UNLISTEN * during connection setup and cleanup. #110780 (Alexey Milovidov).
• Change the default value of the auto_statistics_types setting from basic, uniq to basic, uniq_v2. #110878 (Han Fei).
• randomHadamardTransform now computes an exact, length-preserving transform for any vector length whose largest odd factor is at most 64 (for example 3584 = 512 * 7, common in embedding models), extending the previous 2^N, 2^k * {12, 20}, and 2^k * 9 families. A full transform of a length that cannot be represented exactly now raises an exception instead of silently zero-padding to a longer vector; pass output_dims to compute a truncated projection of an arbitrary length. #111006 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR exception during backup of a Replicated database that is being dropped and recreated concurrently; such backups now fail cleanly with CANNOT_GET_REPLICATED_DATABASE_SNAPSHOT. #100651 (Alexey Milovidov).
• Surface the real underlying error when a zip archive cannot be unpacked (for example, when the archive is read from S3 and the read buffer refuses a seek). Previously, the actual error was hidden behind the generic Couldn't unpack zip archive: Code = -100 / Couldn't open zip archive message. #105103 (Groene AI).
• Fix a confusing “Maybe you meant X?” hint after a server restart (or DETACH DATABASE/ATTACH DATABASE): dropping an already-dropped table could suggest the just-dropped name as the alternative. #106238 (Groene AI).
• Fixed a logical error (Logical error: 'removed', aborting the server in debug builds) in the background table-drop queue, triggered when the same explicit UUID is reused across several CREATE OR REPLACE TABLE queries, which enqueues more than one dropped table sharing that UUID. #107031 (Groene AI).
• Fixed an Inconsistent AST formatting logical error that could abort the server in debug and sanitizer builds when an aliased lambda was used as the operand of an access operator (tuple element .N or array element []) at a non-first position of an expression list, e.g. SELECT 1, ((p0, p1) -> p0 AS a7).4[3] FROM t. #107092 (Groene AI).
• A cluster table function (urlCluster, fileCluster, s3Cluster, …) nested inside another distributed query, such as clusterAllReplicas(..., urlCluster(...)), is now rejected with a BAD_ARGUMENTS error instead of failing with a logical error (Distributed task iterator is not initialized). #107107 (Groene AI).
• OPTIMIZE ... DRY RUN interrupted by a query timeout (max_execution_time with timeout_overflow_mode = 'break') now returns TIMEOUT_EXCEEDED instead of a logical error about rows_sources (which aborted the server in debug/sanitizer builds). #107114 (Groene AI).
• Stop logging a benign Net Exception: Socket is not connected error during ZooKeeper session finalize on macOS. clickhouse keeper-client no longer prints this spurious error to stderr at exit. #107438 (Groene AI).
• Fixed a signed integer overflow when a refreshable materialized view retries a failed refresh with refresh_retries set to a very large value (near Int64 max). The overflow could abort the server in builds with the undefined behavior sanitizer. #108005 (Groene AI).
• Fix a LOGICAL_ERROR (server abort in debug/sanitizer builds) when a JOIN uses the null-safe comparison operator (<=> / IS NOT DISTINCT FROM) and one of the keys is a scalar subquery, e.g. ... JOIN t2 ON (SELECT x FROM t1) <=> t2.k, with the old analyzer (enable_analyzer = 0). Such a query now returns a regular NOT_FOUND_COLUMN_IN_BLOCK error instead of a logical error. #108123 (Groene AI).
• Vector search queries that select the _distance column now return a proper error instead of failing with a logical error. #108423 (Robert Schulze).
• Fix a std::future_error (The associated promise has been destructed prior to the associated state becoming ready) that could surface, and abort the server in debug and sanitizer builds, when scheduling the final asynchronous S3/Azure multipart-upload completion task failed (for example under thread-pool exhaustion). The real scheduling error is now reported instead. #108730 (Groene AI).
• Avoid excessive server log output and an oversized error message when compiling a very large regular expression (for example a LIKE or match pattern with hundreds of thousands of wildcards); such patterns now fail with a clear CANNOT_COMPILE_REGEXP error. #108821 (Raúl Marín).
• Fixed undefined behaviour when stringifying an out-of-range protocol packet type (e.g. in the Unexpected packet from server / Received ... packet error messages) for a desynced or fuzzed connection. #108885 (Groene AI).
• Fixed a confusing internal error (Method getResultType is not supported for TABLE query tree node) when a table expression was used as the left argument of the IN operator; such queries now produce a clear error message. #109412 (Alexey Milovidov).
• Fixed reading an Iceberg table with a metadata file whose version number is all digits but exceeds the 32-bit integer range (for example v99999999999999999999.metadata.json). Such a name previously produced an opaque std::out_of_range (STD_EXCEPTION) instead of a clear BAD_ARGUMENTS error. #109619 (Groene AI).
• Fixed a possible Logical error: 'ReadBuffer is canceled. Can't read from it.' when a row-based input format (e.g. TSV/CSV) failed to parse input and the underlying read had already been canceled (for example a malformed HTTP chunk or a truncated body). Building the verbose parse diagnostics no longer reads from a canceled buffer. #109708 (Groene AI).
• Row policy filter expressions using arrayJoin (or its unnest alias) are now rejected with a clear error; previously such policies raised a column->size() == num_rows logical error at query time. #109753 (Raúl Marín). #109973 (Raúl Marín).
• Reading a Paimon table whose schema contains an unsupported nested type (for example a ROW field) now reports a clear BAD_ARGUMENTS error naming the unsupported type, instead of a bare DB::Exception. (OK) with error code 0 and no message. #109762 (Groene AI).
• Fixed a LOGICAL_ERROR (std::length_error) when a query with the experimental setting make_distributed_plan = 1 used a very large distributed_plan_default_reader_bucket_count or distributed_plan_default_shuffle_join_bucket_count. Such values are now rejected with INVALID_SETTING_VALUE. #109770 (Groene AI).
• Fix a LOGICAL_ERROR (query tree node does not have valid source node) when a recursive CTE resolved an identifier from an outer scope as a correlated column (with allow_experimental_correlated_subqueries = 1). Such queries now return a clear UNSUPPORTED_METHOD error. #109863 (Groene AI).
• Fix a signed integer overflow when ORDER BY ... WITH FILL skips a very large gap over an Int64/UInt64 column (e.g. a step of 2 across a gap near 2^63). The overflow was undefined behavior under sanitizers and silently wrapped in release builds. #109937 (Groene AI).
• Fix a Bad cast from type DB::FunctionNode to DB::ConstantNode logical error (server abort in debug/sanitizer builds) when running SELECT ... ORDER BY ... WITH FILL against a Distributed table with a low optimize_const_name_size. #109938 (Groene AI).
• Fixed an Invalid number of rows in Chunk logical error when an INTERPOLATE target is an alias of a WITH FILL column with the old analyzer (enable_analyzer = 0). Such queries are now rejected with a clear INVALID_WITH_FILL_EXPRESSION error. #110103 (Groene AI).
• Fix a logical error (server abort in debug/sanitizer builds) when the indexHint/ignore/isZeroOrNull functions are given an argument whose type resolves to Nothing, e.g. inside expressions like indexHint(assumeNotNull(materialize(NULL))). #110192 (Groene AI).
• Fixed a logical error Too large size (...) passed to allocator that could occur when an out-of-range value was set for a read buffer size setting such as max_read_buffer_size or max_read_buffer_size_local_fs. Such values are now clamped to 256 MiB by the settings sanity check, and additionally at the consumption site for INSERT ... FROM INFILE, which reads files inside clickhouse-client where the sanity check does not apply. #110207 (Alexey Milovidov).
• CHECK TABLE now reports the actual corruption for a projection part that failed to load its metadata, instead of a misleading Columns doesn't match ... Expected: 0 columns error. #110262 (Raúl Marín).
• A type mismatch when a value read from a PostgreSQL source (e.g. the postgresql table function or the PostgreSQL table engine) cannot be parsed into the declared column type (for example a text column declared as Int32) is now reported as a proper query error; previously it aborted the server in debug and sanitizer builds. #110264 (Groene AI).
• Support the query profiler, memory and trace profilers, and system.trace_log symbolization on macOS. #109825 (Raúl Marín).
• Update chdig to v26.7.1: compressed stack traces for sharing, query patterns, separate-arena memory support, Perfetto improvements, and compatibility fixes. #110938 (Azat Khuzhin).
• Fix incorrect monotonicity inference for intDiv on an unsigned key divided by a signed constant, which could disable primary-key index pruning for IN/NOT IN predicates that cross the signed boundary of the result type; debug builds also hit a LOGICAL_ERROR Invalid binary search result in MergeTreeSetIndex. #107586 (Groene AI).
• Added the allow_lossy_numeric_supertype setting (disabled by default). When enabled, if, multiIf, coalesce, ifNull, array, and map over numeric arguments that have no lossless common type (for example Decimal and Float64, or Int64 and Float64) resolve to Float64 instead of Variant, so the result can be used with aggregate functions such as sum, avg, min, and max. Relevant aggregate-function error messages now mention the setting by name. Closes #106707. #107236 (Groene AI).
• Start the Prometheus endpoint (for metrics-only configurations) and asynchronous metrics collection before tables are loaded, so metrics are visible during the potentially long metadata loading phase. #108402 (Christoph Wurm).

​

Bug Fix (user-visible misbehavior in an official stable release)

• Fix TRUNCATE TABLE and DROP PARTITION failing on tables with many deduplication blocks, where removing them in a single ZooKeeper request could exceed the default 1 MB jute.maxbuffer limit. #105991 (Clayton McClure). This only affects you if you use Apache ZooKeeper (not recommended) instead of ClickHouse Keeper.
• Fix the usage of qualified column names (database.table.column) in the WHERE clause of mutations, such as DELETE FROM and ALTER TABLE ... UPDATE/DELETE, over MergeTree-family tables; previously such queries failed with a missing-columns error. Closes #71760. #109491 (Mikhail Artemenko).
• Fix the query result cache for PromQL queries. The promql dialect bypassed the non-deterministic-function check and could serve stale results anchored at now(); the Prometheus HTTP API (/api/v1/query, /api/v1/query_range) never stored cache entries at all. #110887 (Nikita Mikhaylov).
• Fix the server failing to start when the TZ environment variable is empty. Closes #68920. #68921 (Ardenwick).
• Fixed the server failing to start when a Backup database engine refers to a backup that has become unavailable (for example, when its files were deleted or the underlying storage is inaccessible). The database is now loaded without tables instead of preventing the whole server from starting. #83188 (Vitaly Orlov).
• When using background inserts into tables with the Distributed engine, delays caused by repeated errors when sending data to remote shards now decrease once the errors are resolved. Previously, the delay would only increase and never reset. #87378 (Andrei Kochemirovskii).
• Fix broken_data_files in system.distribution_queue and the BrokenDistributedFilesToInsert metric always reporting 0 for broken files discovered when scanning the Distributed async-insert queue at startup. #92124 (KG.Xu).
• Fix recovery of logging after a disk-full error. Previously, when the log disk became full, ClickHouse would enter a failed state and continuously spam error messages to syslog (potentially writing 100+ GB/hour), never recovering even after disk space was freed. Now the logging system automatically recovers once disk space becomes available, without requiring a server restart. #93127 (jaehanbyun).
• Forward static S3 credentials to Unity DataLakeCatalog table reads when vended_credentials is disabled, preventing anonymous object-storage access failures. #96910 (kgeg401).
• Fix a crash in the schema parsing code of the DeltaLake table engine. #97112 (Kseniia Sumarokova).
• Fix a security issue where an unauthenticated TCP client could probe table existence and replication status via the interserver port. #99854 (Shaohua Wang).
• Fix a syntax error when an alias follows a subquery in DESCRIBE TABLE. Fixes #100031. #100205 (Yarik Briukhovetskyi).
• Fixed an exception in correlated subqueries when outer columns become Nullable under group_by_use_nulls with ROLLUP/CUBE. #100365 (Alexey Milovidov).
• Fixed a server abort (in sanitizer builds) and silent data loss during table startup when a rolled-back transactional MergeTree part intersects a committed part. #100992 (Tuan Pham Anh).
• Fixed indexOfAssumeSorted returning incorrect results for Array(LowCardinality(String)) columns in MergeTree tables. #101771 (Yash ).
• Fix incorrect primary-key pruning when a table’s sorting key wraps a Date column in toDateTime (for example, ORDER BY toDateTime(date_column)) and a query filters on the original column with a comparison like WHERE date_column >= '...'. toDateTime(Date) overflows for Date values beyond the DateTime range (after 2106-02-07), so the stored key is non-monotonic; ClickHouse no longer uses primary key pruning for this key/predicate combination because doing so could drop granules that contain matching rows. Closes #101744. #101814 (Nihal Z. Miaji).
• Fixed intDivOrNull, moduloOrNull and positiveModuloOrNull returning 0 instead of NULL, and intDivOrNull and intDivOrZero raising an exception instead of returning NULL / 0, when the division leads to a floating-point exception (division by zero or INT_MIN / -1), including for mixed signed/unsigned arguments. #101976 (Yarik Briukhovetskyi).
• Fix LOGICAL_ERROR exceptions when reading Iceberg or DeltaLake tables in corner cases such as concurrent Iceberg metadata updates or reads through the merge table function over DeltaLake tables. #102033 (Groene AI).
• Fix outbound HTTP requests (e.g. the aiGenerate function, the url table function, S3) failing with No route to host on hosts that advertise both IPv4 and IPv6 addresses when only one address family is routable. The HTTP connection pool now falls back to the next resolved address when the first one fails with a network error, instead of propagating the error on the very first request. #103786 (Alexey Milovidov).
• Fix a server abort during RESTORE of backups containing tables with cyclic dependencies. #103824 (Konstantin Bogdanov).
• Schema inference for Arrow time32/time64 columns (e.g. SELECT * FROM file(..., 'Arrow')) now infers Time64 instead of DateTime64; anything downstream relying on the old inferred type will observe the new type after an upgrade. Also, Time/Time64 values can now be exported to Arrow with the appropriate Apache Arrow time type selected by precision. Fixes #104038. #104316 (/bin/cat).
• Fixed a race between ALTER TABLE ... RENAME COLUMN and a concurrent OPTIMIZE TABLE ... FINAL (or any background merge) in MergeTree that could silently replace the renamed column’s data with default values. #104822 (Groene AI).
• Fixed a Block structure mismatch logical error during a concurrent INSERT while an ALTER TABLE ... RENAME COLUMN is in flight, in an Atomic database with lazy_load_tables = 1 after DETACH DATABASE/ATTACH DATABASE. #104852 (Groene AI).
• Fix groupConcat when the parametric and two-argument spellings are mixed, e.g. groupConcat(',', 2)(x, '/'): the row-limit parameter was silently dropped, so every row was returned instead of the requested number; the delimiter from the second argument now correctly overrides the parameter. #104882 (Yarik Briukhovetskyi).
• Fix a LOGICAL_ERROR (Expected one block from input stream) thrown by KILL QUERY/KILL MUTATION/KILL PART_MOVE_TO_SHARD/KILL TRANSACTION when their WHERE clause contains a per-row subquery, or when max_block_size is small enough that the internal SELECT over the relevant system.* table emits more than one block. #104927 (Groene AI).
• Fixed CAST from smaller to larger interval units returning wrong results (for example, CAST(toIntervalSecond(60) AS IntervalMinute) returned 0 instead of 1). Closes #104986. #105058 (Yarik Briukhovetskyi).
• Fix a logical error (Assertion 'row < chunk.getNumRows()' failed) in LIMIT ... WITH TIES queries running with the read-in-order pipeline (optimize_read_in_order = 1 and read_in_order_use_virtual_row_per_block = 1). #105102 (Groene AI).
• Fix a leak in refreshable materialized views (MATERIALIZED VIEW ... REFRESH ...): the temporary inner table that a refresh rotates out via EXCHANGE TABLES could not be dropped if it was larger than max_table_size_to_drop, so every subsequent refresh created yet another temporary inner table, growing the view’s data directory until the disk was exhausted. The refresh task’s internal drop now bypasses max_table_size_to_drop and max_partition_size_to_drop; those safety nets still apply to user-issued DROP TABLE of the view itself. Closes #104900. #105106 (Groene AI).
• Fix an unexpected DATA_TYPE_CANNOT_BE_USED_IN_KEY error on ALTER queries that do not change the sorting key (changing settings, comments, codecs, adding a non-key column, etc.) for MergeTree tables that have a SimpleAggregateFunction (or another type allowed only with allow_suspicious_primary_key = 1) in the sorting key. #105111 (Groene AI).
• Fixed a use-after-free when querying system.clusters replica-state columns (such as is_active, unsynced_after_recovery, recovery_time) while a Replicated database is being dropped or detached. Also, benign Keeper exceptions swallowed in this code path (the affected database is treated as transiently unavailable and skipped) are now logged at the Information level instead of Error, so they no longer reach clients at the default send_logs_level = 'warning'. #105149 (Groene AI).
• Fixed an issue where running multiple statements in clickhouse-local --ignore-error would echo every subsequent statement inside the error message for each lexical error or unmatched-parenthesis error. #105480 (Groene AI).
• Fix a rare server crash when refreshable materialized views are dropped or replaced concurrently with their refresh scheduling. #105588 (Groene AI).
• Reject ALTER TABLE ... DELETE and ALTER TABLE ... UPDATE on Iceberg tables whose data file format is not Parquet with a clear NOT_IMPLEMENTED error instead of crashing the server or silently corrupting the table. #105893 (Groene AI).
• Fix wrong results from a view defined with EXCEPT or INTERSECT whose operand is a UNION chain, after DETACH/ATTACH or a server restart. The formatter was missing parentheses around UNION children of INTERSECT/EXCEPT, so the stored SQL was reparsed with reversed precedence. #105935 (Groene AI).
• Fix JSON data misdetected as TSKV during format auto-detection. Closes #100797. #106009 (Pavel Kruglov).
• Forbid creating a minmax skip index directly on JSON columns — it could later fail inserts with a NO_COMMON_TYPE exception when mixed-type arrays were inserted. Create the index on typed subcolumns instead (e.g. INDEX idx json.field TYPE minmax). #106094 (linhaojie).
• Fixed a logical error (Parsed partition value ... doesn't match partition value for an existing part with the same partition ID) thrown by OPTIMIZE TABLE ... PARTITION ... and other queries that resolve a partition value, on tables with a Time-typed partition key. #106202 (Groene AI).
• Fix a data part being incorrectly marked as having a broken projection after a lightweight delete leaves no projection part behind — either when the projection is rebuilt with zero output rows or when it is dropped (lightweight_mutation_projection_mode = 'rebuild'/'drop'). The broken state previously disabled projection optimization for queries on that part. #106273 (Shaohua Wang).
• Fixed removeDirectory on plain object storage disks (such as s3_plain) removing all files inside a non-empty directory as if the removal were recursive; removing a non-empty directory now fails with CANNOT_RMDIR, and recursive removal handles the contents explicitly. #106281 (RinChanNOW).
• Fixed a logical error (Part ... doesn't exist) that could abort the server during recovery of a quorum INSERT from a Keeper hardware error when the quorum was concurrently marked as failed. #106424 (Alexey Milovidov).
• Fix a LOGICAL_ERROR exception about duplicate column names when applying row policies. #106438 (János Benjamin Antal).
• Fix CREATE OR REPLACE TABLE and REPLACE TABLE leaving a stranded _tmp_replace_* table on disk and replacing the original table with an empty one when the existing target exceeds max_table_size_to_drop. The size check now runs before the swap, so a violation aborts cleanly with the user-visible table name in the error. #106782 (Groene AI).
• Fix JSON_QUERY/JSON_VALUE/JSON_EXISTS returning Dynamic instead of String/UInt8 on Dynamic arguments. Closes #106461. #106877 (Pavel Kruglov).
• Fixed wrong results when querying a ReplacingMergeTree table with FINAL and a filter on a text index while query_plan_optimize_lazy_final was enabled. The lazy FINAL optimization built reading steps that did not reproduce the direct read from the text index, so the filter dropped all matching rows. #106894 (Jimmy Aguilar Mena).
• Apply query_masking_rules to messages appended to exceptions, so URL-encoded credentials in (in file/uri ...) suffixes of errors from the jdbc/odbc table functions are not leaked when masking rules are configured. #106916 (Gaurav Dubey).
• Fixed parts being marked as broken and detached on any part reload (server restart, DETACH/ATTACH) for tables with a LowCardinality(Nullable(...)) column in the partition key. Since 26.5, the per-part minmax index file was not written when such a column’s minimum and maximum were NULL, while the part consistency check still required the file. Parts written by affected versions lack the minmax index file and still need to be re-attached manually. Closes #106837. #106945 (Pedro Ferreira).
• Fix a LOGICAL_ERROR (Unexpected expression in JOIN ON section. Expected boolean (UInt8), got 'Nothing') when a non-equi JOIN ... ON predicate references a Nothing-typed column, such as one produced by ARRAY JOIN []. #106981 (Groene AI).
• Fix a rare LOGICAL_ERROR (Attempt to release query context that does not exist) and the accompanying server crash when reading MergeTree tables through a filesystem cache disk created with enable_filesystem_query_cache_limit = 1. #107028 (Groene AI).
• Fix a server crash when moving an empty part to a plain_rewritable disk (for example, with ALTER TABLE ... MOVE PARTITION ... TO DISK for an empty part kept by remove_empty_parts = 0). #107040 (Groene AI).
• Fix incorrect row order in ORDER BY queries over UNION ALL with optimize_read_in_order and read_in_order_use_virtual_row enabled. Closes #106879. #107053 (Vladimir Cherkasov).
• Fixed a LOGICAL_ERROR (Unsupported argument types) in the geohashesInBox function when its coordinate arguments mixed constant and non-constant values, or when a coordinate was a BFloat16. Mixed const/non-const Float32 arguments now work, and BFloat16 arguments are rejected with a clear error. #107063 (Groene AI).
• Fixed wrong results (duplicate rows) for SELECT DISTINCT over a GROUP BY with WITH CUBE, WITH ROLLUP, or GROUPING SETS on the same keys. The query_plan_remove_redundant_distinct optimization no longer removes the DISTINCT in these cases, because the grouping modifiers emit extra rows with the key columns defaulted, which can collide with real group values. #107072 (Groene AI).
• Fix a wrong result in an OUTER JOIN when the WHERE filter is a conjunction that contains a constant term (for example p AND 1, or WHERE p QUALIFY NULL where the QUALIFY NULL is merged into WHERE). The constant conjunct could be pushed into the non-preserved side of the join and removed from the post-join filter, so the join produced non-matched rows with the side columns defaulted, and those rows incorrectly passed the filter. #107084 (Groene AI).
• Fixed executable_pool user-defined functions configured with <lifetime> not picking up changes to the underlying script. Previously, only SYSTEM RELOAD FUNCTIONS would re-read an edited script; periodic <lifetime> reloads kept executing the old version. #107087 (Groene AI).
• Fix a LOGICAL_ERROR exception (Cannot find input column ... on its position in inputs of expression actions DAG) for a correlated scalar subquery over a source projecting the same column identifier twice (e.g. SELECT number, *) when correlated_subqueries_default_join_kind = 'left'. #107112 (Groene AI).
• Fixed a data race on the server-global trace collector between a worker thread starting its profiler and server shutdown. #107307 (Groene AI).
• Incremental backups no longer store S3 credentials in the <base_backup> locator of the .backup metadata file. Backups created with use_same_s3_credentials_for_base_backup = 1, or with explicit base backup credentials matching this backup locator, store a non-secret marker and are restored without extra restore-time settings; for backups created with different explicit base backup credentials or extra base authentication arguments, pass them to RESTORE with the base_backup setting. Backups created by older versions with embedded credentials remain restorable. #107357 (Pablo Marcos).
• Fixed the CSVWithNames and CSVWithNamesAndTypes header having fewer columns than the data when output_format_csv_serialize_tuple_into_separate_columns is enabled (the default). The header (and the types row) now flattens Tuple columns into their leaf fields with dotted names (e.g. t.a, t.b), so the header column count matches the data. A new setting output_format_csv_header_serialize_tuple_into_separate_columns (default 1) controls this and can be set to 0 to restore the previous single-name header. #107371 (Groene AI).
• Fixed reading Iceberg v3 tables whose Parquet data files contain reserved row-lineage columns (such as _row_id); the native Parquet reader no longer raises ICEBERG_SPECIFICATION_VIOLATION for reserved field IDs that are not part of the table schema. #107377 (Greg Maher).
• Fixed a rare server abort during shutdown (std::future_error: The associated promise has been destructed prior to the associated state becoming ready) caused by a scheduled task being dropped from a thread pool queue before it ran. #107383 (Groene AI).
• Fix the MaterializedPostgreSQL table and database engines so that a single table or database can be replicated from a non-default PostgreSQL schema (materialized_postgresql_schema), including the case where tables with the same name exist in several schemas of the same database (previously they would share a publication and replication slot and cross-talk). #107425 (Alexey Milovidov).
• Fix MaterializedPostgreSQL stopping replication of an entire database (with LOGICAL_ERROR: Columns number mismatch) when a single replicated table’s structure changed while the server was down. Now only the affected table is skipped and can be recovered with DETACH/ATTACH. #107427 (Alexey Milovidov).
• BACKUP of a MaterializedPostgreSQL database no longer hangs forever with Table ... were created or changed its definition during scanning, and now actually backs up the table data (delegated to the underlying ReplacingMergeTree), which can be restored as a standalone ReplacingMergeTree. #107433 (Alexey Milovidov).
• Fixed a regression where setting compatibility = '26.6' (which implicitly enables the hive partition strategy) silently accepted {_partition_id} in S3/object storage table paths instead of raising BAD_ARGUMENTS. #107437 (Lefteris).
• Fixed a memory leak in the bundled mongo-c-driver that could occur when reading from MongoDB (for example, via a MongoDB dictionary or the mongodb table function) if a retryable read error was followed by a failed retry server selection. #107448 (Groene AI).
• Fix an exception when inserting into an Iceberg table whose metadata was created by an external engine and omits the optional snapshots, metadata-log, or snapshot-log arrays. Such inserts now succeed instead of failing. #107473 (Shaohua Wang).
• Fixed DELETE FROM (lightweight delete) requiring the ALTER UPDATE privilege in addition to ALTER DELETE. A user granted only ALTER DELETE can now run DELETE FROM, as documented. #107491 (Shaohua Wang).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK when reading Hive-partitioned files with use_hive_partitioning = 1 and a WHERE/PREWHERE clause that filters a real (non-virtual) column while a Hive partition column is also selected. #107505 (Groene AI).
• Fix a LOGICAL_ERROR (Different order of columns in UNION subquery) when the old analyzer (enable_analyzer = 0) reads a subset of columns from a subquery whose UNION children are INTERSECT/EXCEPT set operations. #107511 (Groene AI).
• Fix NULL values being silently converted to empty strings when inserting Arrow/ORC data into a LowCardinality(Nullable(...)) column. This was a regression introduced in 26.5. #107532 (Jimmy Aguilar Mena).
• Fix a logical error (Stream ... variant_discr ... is not found) that could occur when merging or reading a MergeTree part produced by a mutation of a table with a Dynamic column. #107562 (Alexey Milovidov).
• Fixed a logical error (WhichDataType(const_type).isArray()) during primary-key analysis when pointInPolygon is called with a constant polygon argument of a wrapper type such as Variant or Dynamic (for example, pointInPolygon((x, y), if(c, [(0, 0), ...], NULL))). #107589 (Groene AI).
• Fixed a logical error (Last stored last_written_position in meta file ... is bigger than current last_written_pos) in the FileLog engine that could happen when a watched file was deleted and recreated reusing the same inode. #107617 (Groene AI).
• Fix a rare server abort (std::future_error: The associated promise has been destructed prior to the associated state becoming ready) during ZooKeeper client request processing, when an asynchronous Keeper request was dropped while being sent, for example under memory pressure. #107647 (Groene AI).
• Fixed RESTORE for ReplicatedMergeTree tables so duplicate-content parts from a backup are preserved instead of being silently deduplicated. #107652 (Pablo Marcos).
• Fix runtime join filter producing wrong results for JSON columns. Closes #107646. #107663 (Pavel Kruglov).
• Fixed incorrect results from distributed queries selecting ALIAS columns that share a common subexpression: columns could be misaligned and values returned under the wrong column. #107675 (Vladimir Cherkasov).
• Quotas keyed by normalized_query_hash now account all resources (read_rows, read_bytes, result_rows, result_bytes, execution_time, written_bytes, errors) per query pattern, like the query-count counters, instead of accounting them against a single shared per-user bucket. #107681 (Alexey Milovidov).
• Fixed a Block structure mismatch in UnionStep stream logical error (server abort on debug/sanitizer builds, Code: 49 on release builds) that occurred when sibling branches of a UNION/INTERSECT/EXCEPT differed only in their WHERE predicate and one branch’s predicate was constant-folded to a Const column. #107719 (Groene AI).
• Fix logical errors (Bad cast) caused by inconsistent stripping of LowCardinality nested inside Variant and Dynamic columns, for example in concat, format, and primary key analysis. Closes #107598. #107773 (Pavel Kruglov).
• Fixed a severe work-distribution skew with parallel replicas when the cluster contains inactive replicas (for example, stale entries left after autoscaling). Such replicas are no longer counted by the reading coordinator, so work is balanced across the online replicas instead of piling onto a single one. #107805 (Alexey Milovidov).
• Fix the Not-ready Set is passed as the second argument exception that could occur when building an IN subquery set during primary key analysis failed silently (for example, a subquery timeout with overflow_mode = 'break'), leaving the set permanently unbuilt for the query pipeline. #107924 (Alexey Milovidov).
• Fix a server crash in predicate pushdown (enable_optimize_predicate_expression, legacy analyzer) when a UNION subquery appears inside a JOIN ... ON condition. #107930 (Groene AI).
• Fix spurious CHECKSUM_DOESNT_MATCH and Parquet read errors (and, for some formats, silently wrong results) when an S3 or S3-compatible (e.g. GCS) object is overwritten in place while being read. Such reads now fail with a clear, retryable error instead of returning data stitched from two object versions; controlled by the new setting s3_validate_etag_on_read (on by default). #107934 (Shaohua Wang).
• Fixed a SELECT failing with NOT_FOUND_COLUMN_IN_BLOCK on a table that has a set data-skipping index when a row policy filters it using an always-true condition combined with a check on a non-indexed column. #107971 (Shaohua Wang).
• Fixed executable user-defined function command parameter parsing so placeholders with empty or invalid names are not accepted as parameters. #107983 (Goutam Adwant).
• Fix performance regression for Map subcolumns with PREWHERE. Closes #107912. Caused by #99200. #107988 (Pavel Kruglov).
• Fix an out-of-bounds write when reading a Variant column from a malformed Native block whose discriminators reference a variant index that does not exist. #107991 (uwezkhan).
• Fix wrong query results from incorrect primary key and partition pruning when toStartOfDay, toStartOfISOYear, toDaysSinceYearZero, toRelativeSecondNum, toRelativeMinuteNum, toRelativeHourNum, toRelativeWeekNum, toRelativeDayNum, toMonthNumSinceEpoch, or toYearNumSinceEpoch is applied to a Date32 key column containing dates outside the range the function can represent (in particular dates before 1970-01-01). These functions report themselves as monotonic to the primary index, but previously wrapped around for such arguments, so the index could prune granules that actually contained matching rows; they now saturate at the bounds of their result type and stay monotonic over the whole Date32 range. #108018 (Nihal Z. Miaji).
• Fix accurateCastOrNull of a Tuple whose element is Dynamic/Variant: a genuine source NULL was treated as a conversion failure. For a Nullable target element a source NULL now stays an element NULL (matching a plain Tuple(Nullable(...)) source); for a non-Nullable target element a source NULL now nulls the whole tuple instead of producing the element default; parse/overflow failures still null the whole tuple. #108023 (Groene AI). Also fix a logical error when casting a Dynamic or Variant column nested inside a Tuple to a non-Nullable element type with accurateCastOrNull or accurateCastOrDefault. #108061 (Alexey Milovidov).
• Fixed numericIndexedVectorPointwiseMultiply returning an empty result when multiplying by an all-ones vector that has a different BSI (bit-sliced index) configuration. #108027 (Alexey Milovidov).
• Fixed toTime64 and CAST(... AS Time64) not clamping out-of-range values to the Time64 range in saturate and ignore overflow modes, which could produce values that display identically but compare as different. #108028 (Alexey Milovidov).
• getClientHTTPHeader is now correctly treated as non-deterministic, so its result is no longer incorrectly reused by the query result cache. #108029 (Alexey Milovidov).
• clickhouse-client --port 9440 again automatically enables a secure (TLS) connection; this was broken by a refactoring that stopped passing the port to the secure-connection check. #108032 (Alexey Milovidov).
• A transient error while refreshing data parts of a read-only table no longer permanently stops the background refresh task. #108034 (Alexey Milovidov).
• Fixed memory/CPU/mutation overload warnings in system.warnings lagging one asynchronous-metrics cycle behind (and being absent on the first cycle). #108035 (Alexey Milovidov).
• The mongodb table function now accepts oid_columns passed as a named argument (e.g. oid_columns='_id'), instead of rejecting it with BAD_ARGUMENTS. #108039 (Alexey Milovidov).
• Fix a Bad cast from type DB::ColumnNullable to DB::ColumnVector<...> exception (logical error) when a qualified asterisk (t.*) selects a JOIN USING key and that join is nested below a PASTE/CROSS/comma join or an outer ON join, with join_use_nulls = 0. #108043 (Groene AI).
• Fixed incorrect handling of zero-width assertions (\b, ^, $) in the regexp “match all” functions extractAll, extractAllGroupsVertical, extractAllGroupsHorizontal, countMatches and splitByRegexp. For example, extractAll('new york is the greatest', '\b(\w)') now correctly returns the first letter of each word instead of every letter. #108047 (Alexey Milovidov).
• Fixed CREATE TABLE ... AS SELECT on Atomic databases leaving an empty table behind when the query fails — for example when the user has no access to a table referenced from a subquery. Previously a retry reported that the table already exists instead of the original error. The table is now created via a temporary table and becomes visible only after it has been fully populated. #108048 (Alexey Milovidov).
• Allow querying a range_hashed or complex_key_range_hashed dictionary that uses a DateTime64, Decimal or floating-point range with a matching argument to dictGet/dictHas. Previously such queries failed with must be convertible to Int64, and open-ended intervals of Decimal/DateTime64 ranges incorrectly returned the default value. #108052 (Alexey Milovidov).
• Fixed an exception (CANNOT_PARSE_TEXT) when loading a dictionary with a composite key whose key columns are not the first columns in the dictionary definition, with a local ClickHouse source using an explicit query. Source columns are now matched to the dictionary structure by name instead of by position. #108053 (Alexey Milovidov).
• Fixed the values table function failing with ARGUMENT_OUT_OF_BOUND when a decimal literal that is not exactly representable in a narrow floating-point column (such as 0.1 for a Float32 column) is used, e.g. SELECT * FROM values('x Float32', 0.1). Such values are now accepted and converted to the nearest representable value, consistent with CAST and INSERT ... VALUES. #108055 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR (No available columns) when executing SELECT count() (or other trivial queries) on a table where the user is granted SELECT access only on an ALIAS column. #108056 (Alexey Milovidov).
• Fixed mapFilter, mapSort (and its variants) and mapConcat dropping LowCardinality from the key and value types of a Map. Previously mapFilter over a Map(LowCardinality(String), String) column returned Map(String, String), which could corrupt the metadata of a table created via CREATE TABLE ... AS SELECT and make CHECK TABLE fail. #108057 (Alexey Milovidov).
• Fix incorrect results when using hasToken on text indexes with a tokenizer other than splitByNonAlpha. #108066 (Robert Schulze).
• Fixed a freeze of SYSTEM RELOAD DICTIONARIES (and RELOAD DICTIONARY) when many MySQL dictionaries defined in XML shared a single connection pool via share_connection. Such pools now honor connection_pool_size and connection_wait_timeout (default 5 seconds) from the configuration, matching dictionaries defined through named collections. #108083 (Alexey Milovidov).
• Allow EXISTS <dictionary> for a user that has only the SHOW DICTIONARIES privilege on the dictionary (previously it required SHOW TABLES). #108084 (Alexey Milovidov).
• Fixed CREATE TABLE dst AS src SETTINGS ... (and similar variants with ORDER BY/PARTITION BY but without an explicit ENGINE) silently dropping the source table’s engine, storage clauses (such as table TTL and keys) and settings. The engine and the non-overridden storage clauses are now inherited from the source table, and the specified settings are merged on top of the source table’s settings. #108085 (Alexey Milovidov).
• Fixed INSERT into MergeTree tables failing with filesystem error: in rename: Permission denied on filesystems backed by Windows (WSL, CIFS/SMB, Docker Desktop bind mounts), caused by the part writer leaving file descriptors open across the part-directory rename. #108089 (Alexey Milovidov).
• Fixed a crash (null pointer dereference) when querying a Hive table without a WHERE clause, e.g. SELECT * FROM hive_table. #108094 (Alexey Milovidov).
• Fix a LOGICAL_ERROR (“Unexpected return type from if”) when reading a column under apply_mutations_on_fly = 1 after an ALTER UPDATE col = ... WHERE <cond> with a non-constant or false condition followed by ALTER MODIFY COLUMN col <new type>. #108128 (Groene AI).
• Fix toStartOfInterval and dateTrunc returning a rounded value instead of the start of the containing interval when the input has finer precision than the interval unit, e.g. toStartOfInterval(toDateTime64('2023-10-09 10:11:12.000999', 6), INTERVAL 1 millisecond) returned 10:11:12.001 instead of 10:11:12.000. The overload with an explicit origin was affected too. Closes #103535. #108186 (Yarik Briukhovetskyi).
• Hide sensitive information in the view_query column of system.query_views_log. #108214 (Valerii Mordovskii).
• Fixed a wedge on a MergeTree table attached with a legacy hypothesis skip index (a removed index type kept only for ATTACH compatibility). Previously a lightweight DELETE, OPTIMIZE, a plain INSERT, or a filtered SELECT on such a table failed with ILLEGAL_INDEX (“Index of type ‘hypothesis’ is no longer supported”), and the DELETE mutation retried forever. The dead index is now treated as inert: it is carried forward untouched during merge/mutation, skipped on insert and query planning, and can still be dropped with ALTER TABLE ... DROP INDEX. #108217 (Groene AI).
• Fix the setting type_json_allow_duplicated_key_with_literal_and_nested_object not working with typed paths in JSON. #108218 (Pavel Kruglov).
• Fix a LOGICAL_ERROR (Different list of shards in child plans) when running a query with make_distributed_plan = 1 over a MergeTree table that has a normal/aggregate projection. #108256 (Groene AI).
• Fixed replaceRegexpOne and replaceRegexpAll so . matches newline characters by default, consistently with other regular expression functions. #108265 (linjiayu1025-collab).
• Fix a logical error when casting an Array(Dynamic) or Array(Variant) to QBit with accurateCastOrNull, e.g. accurateCastOrNull(CAST(range(114), 'Array(Dynamic)'), 'QBit(Float32, 114)'). #108288 (Groene AI).
• index_granularity_bytes is now honored for AggregateFunction state columns. Previously the granule byte cap was ignored for such columns (for example uniqExact states in AggregatingMergeTree tables and aggregating projections), producing granules far larger than the configured limit and increasing read amplification and query-time memory. #108297 (Groene AI).
• Fix lightweight UPDATE/DELETE conditions being evaluated twice, which could lead to incorrect behavior for non-deterministic conditions and extra work for deterministic ones. Closes #86032. #108323 (ofeliacode).
• Fix an exception in a predefined HTTP handler when a header whose headers_regexp regular expression allows an empty value is absent from the request. #108324 (Vitaly Baranov).
• Fix arrayPartialSort with a non-constant limit argument. #108327 (Vitaly Baranov).
• Fixed a cancelled or KILLed INSERT continuing to run for a long time while building a skip index (for example an unbounded set(0) index on a high-cardinality column). The index build now stops promptly when the query is cancelled. #108351 (Shaohua Wang).
• Fixed a SELECT from the primes table function not responding to cancellation: with a large step (or limit) the query could keep running for a long time after KILL QUERY or a timeout. It now stops promptly. #108353 (Shaohua Wang).
• Fixed parsing of data-type names whose name contains the substring INT (for example a function-like name such as quantileInterpolatedWeighted used in a data-type position). Such names were mistaken for MySQL integer types and had their first (...) argument group silently consumed as a display-width modifier, which could break the query formatting round-trip. #108354 (Groene AI).
• Fixed an inconsistency where a malformed ARRAY JOIN followed by a comma and a parenthesized table list was misparsed as a cross join instead of being rejected. #108365 (Raúl Marín).
• Hide secret arguments of functions such as encrypt, decrypt, and HMAC in EXPLAIN actions, EXPLAIN header, and EXPLAIN PIPELINE output when format_display_secrets_in_show_and_select is disabled (the default). #108386 (Raúl Marín).
• Fix CREATE OR REPLACE of a refreshable materialized view with a TO target: it was incorrectly rejected because the target table is already owned by the view being replaced. #108392 (Nikolay Degterinsky).
• Fixed a server crash (Received signal 4, illegal instruction) that could be triggered by a malformed or desynchronized native TCP protocol stream. The client-supplied initial address is now validated to be a numeric host:port before it is parsed, instead of letting a non-numeric port reach the trapped getservbyname libc function. #108410 (Groene AI).
• Fixed a server crash during shutdown when shutting down a database throws an exception (for example when a table’s flush hits a ZooKeeper timeout). The rest of the shutdown sequence is no longer skipped, so system logs are flushed and their threads are joined before the server exits. #108417 (Groene AI).
• Fixed regular expression functions (match, extract, extractAll, replaceRegexpOne, replaceRegexpAll, etc.) silently returning wrong results when the pattern contained a NUL (\0) byte. The NUL is now treated as an ordinary literal byte, consistent with RE2. #108427 (Alexey Milovidov).
• Fixed clickhouse-local returning an empty reply to HTTP OPTIONS requests when no http_options_response is configured, which made the web UI (/play) show the connection as broken even though queries worked. #108428 (Alexey Milovidov).
• Fix a LOGICAL_ERROR (“Invalid action query tree node …”) exception when a distributed query both deduplicated structurally-identical duplicate-ALIAS columns and referenced a table function with a key = value named-collection argument (for example oss(s3_conn, filename = '...')). #108435 (Groene AI).
• Fixed a server crash (LOGICAL_ERROR: 'Unexpected exception in refresh scheduling') that could happen on startup when a coordinated refreshable materialized view in a Replicated database was attached on a Keeper that does not support the MULTI_READ feature flag (for example after a Keeper downgrade). The view is now stopped gracefully instead of aborting the server. #108441 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK error (e.g. Column _part not found) when selecting virtual columns from a MergeTree table under parallel replicas, including via SELECT * with asterisk_include_virtual_columns = 1. #108451 (Groene AI).
• SYSTEM RESET DDL WORKER now requires the new SYSTEM RESET DDL WORKER privilege. Previously any authenticated user (including readonly ones) could run it and repeatedly reset the DDL worker state, blocking ON CLUSTER DDL. #108460 (Groene AI).
• Restrict the model path of catboostEvaluate to the user_files directory, like the file table function and the dictionary sources. Previously the function accepted an arbitrary filesystem path with no containment check, which allowed probing the existence of and triggering reads of files outside user_files. Models must now be located inside user_files. #108463 (Groene AI).
• The hasColumnInTable function now requires the SHOW COLUMNS privilege on the target table, the same grant required by DESCRIBE and SHOW CREATE TABLE. Previously any user could call hasColumnInTable to probe column names and test table and database existence without any access check. #108464 (Groene AI).
• Fixed excessive memory allocation when deserializing crafted aggregate function states for mannWhitneyUTest, rankCorr, largestTriangleThreeBuckets, quantileGK, sequenceMatch/sequenceCount and groupArrayIntersect. A malformed state could declare a huge element count and make the server try to allocate tens of gigabytes from a few bytes of input; such states are now rejected with TOO_LARGE_ARRAY_SIZE. #108465 (Groene AI).
• Validate the redirect target of an S3 301 Moved Permanently response against remote_url_allow_hosts (RemoteHostFilter). Previously the AWS SDK 301 handling followed the attacker-supplied endpoint from the response (Location header or <Endpoint> element) without the host check, so a malicious or compromised S3-compatible server could redirect ClickHouse to an arbitrary host (SSRF). The 307 redirect path already performed this check; both paths are now consistent. #108466 (Groene AI).
• Fixed a credential leak where the legacy storage_aws_access_key_id, storage_aws_secret_access_key, storage_catalog_credential, and storage_auth_header settings of the DataLakeCatalog database engine were shown in plaintext in system.databases.engine_full and SHOW CREATE DATABASE. They are now redacted as [HIDDEN]. #108470 (Groene AI).
• Fixed ssl_certificate user identification so a single * wildcard matches exactly one name component (RFC 6125 6.4.3). Previously a wildcard in a CN or DNS: SAN subject (for example *.corp.example.com) also matched multi-label names such as evil.deep.corp.example.com, letting a holder of a certificate for a deeper subdomain authenticate as the wildcard user. URI: SAN matching is unchanged. #108472 (Groene AI).
• Fixed SQL injection in the MySQL wire protocol: the SHOW TABLE STATUS LIKE argument and the SET <mysql_setting> value sent over the MySQL interface (port 9004) are now parsed and re-quoted instead of being concatenated verbatim into the rewritten query. Also fixed KILL QUERY <id> over the MySQL interface silently ignoring multi-digit connection ids. #108474 (Groene AI).
• Sensitive values passed as HTTP query-string parameters (for example the param_* query-parameter binding such as param_secret_key/param_aws_secret_access_key, the password parameter, or S3-style signature parameters) are no longer written verbatim to system.text_log or OpenTelemetry spans. The Request URI log line now redacts the values of parameters whose names look sensitive, replacing them with [HIDDEN]. #108475 (Groene AI).
• Enforce the http_forbid_headers configuration during schema inference for the url table function and URL storage engine. Previously a forbidden header was still sent over the network while inferring the schema (for example with the Regexp format or DESCRIBE), and its response body could be reflected back in a parse error, leaking secrets. The header filter is now checked before any network access on the inference path, matching the existing behavior of urlCluster. #108476 (Groene AI).
• Fixed wrong results when using SELECT ... SAMPLE ... together with the query condition cache (setting use_query_condition_cache, enabled by default). #108488 (Groene AI).
• Fixed a server crash (stack overflow) caused by deeply nested expressions such as [[[ ... ]]] or array(array( ... )) when max_parser_depth is set to a large value. #108493 (Raúl Marín).
• Fixed the RabbitMQ table engine hanging indefinitely on DROP TABLE or server shutdown when the broker closes the AMQP connection due to missed heartbeats without sending a TCP RST. All blocking event loop calls now have a 30-second timeout so shutdown always completes. #108497 (Aly Ayman).
• Secondary queries executed as part of internal queries will now be logged as internal queries. #108506 (Miсhael Stetsyuk).
• The ArrowFlight table engine and the arrowFlight table function now honor the <remote_url_allow_hosts> allow-list. Previously, the connection host and port were not validated, so a query could reach hosts not present in the allow-list. #108507 (Groene AI).
• The MySQL wire protocol commands COM_FIELD_LIST (mysql_list_fields) and COM_INIT_DB (USE database) now enforce the same access control as their SQL equivalents (SHOW COLUMNS/DESCRIBE and USE). Previously, they could disclose column names of tables the user only had partial column grants on, and switch the current database without the SHOW DATABASES privilege. #108508 (Groene AI).
• Match http_forbid_headers case-insensitively. HTTP header names are case-insensitive, so forbidding Authorization now also blocks authorization, AUTHORIZATION, and other case variants. Configured header_regexp patterns are now matched case-insensitively without needing an explicit (?i) flag. #108509 (Groene AI).
• Fixed a server crash (segmentation fault) that could occur when a distributed query referenced a not-yet-materialized MATERIALIZED CTE as an external table and the remote source was created lazily. #108547 (Groene AI).
• Fixed wrong results when the query condition cache (use_query_condition_cache = 1) reused a skip-index-derived mark exclusion for a query that ran a different set of skip indexes, for example with use_skip_indexes = 0, ignore_data_skipping_indices, or a different use_skip_indexes_for_disjunctions mode. Skip-index-derived cache entries are now keyed by the effective set of skip indexes that ran. #108548 (Groene AI).
• Fixed several bugs in changeYear, changeMonth, changeDay, changeHour, changeMinute and changeSecond with DateTime64 arguments: the result column was created with the hardcoded default scale 3 instead of the argument’s scale, which led to a LOGICAL_ERROR (writeSlice expects same column types) when the result was passed to arrayPushBack/arrayPushFront/arrayConcat; nanosecond-precision (scale 9) inputs threw DECIMAL_OVERFLOW; and pre-epoch sub-second inputs returned the wrong calendar second. Also fixed timeSlots over DateTime64 ignoring the scale of the optional Size argument in its declared return type and returning wrong timestamps for the largest scale in Size. #108551 (Groene AI). #108681 (Takumi Hara). #108994 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK error when a compound predicate is aliased in GROUP BY and referenced again, with enable_identifier_resolve_cache enabled (the default). #108553 (Groene AI).
• Fix a logical error (Equal values are not contiguous within the range assumed to be sorted) when running DISTINCT over a STREAM read (SELECT DISTINCT ... FROM table STREAM). Read-in-order optimizations are no longer applied to STREAM reads, which produce rows in commit order rather than sorting-key order. #108568 (Groene AI).
• Fixed projections returning a column’s type default (e.g. 0) instead of its declared DEFAULT value (e.g. -1) in two cases: when reading a column that was added with ALTER TABLE ... ADD COLUMN ... DEFAULT after the projection was created (reads from the base table were already correct), and after the column’s TTL expired on a wide part. #108569 (Shaohua Wang). #109013 (Shaohua Wang).
• Fixed a ReadBuffer is canceled. Can't read from it. error that could occur when an upload to S3 (for example a backup) is retried after a transient read error. #108573 (Groene AI).
• Fix ATTACH PARTITION ... FROM rejecting tables whose primary keys are equivalent but declared differently (one explicit PRIMARY KEY, the other implicit from ORDER BY). #108590 (Jordi Orihuela).
• Fixed direct read from the text index failing when multiple text indexes are partially materialized. Fixes #108530. #108607 (Anton Popov).
• Fix a metadata disclosure where DESCRIBE loop('db', 'table') and DESCRIBE loop(<inner table function>) bypassed the SHOW COLUMNS / source access check, letting an unprivileged user read a table’s column schema. #108624 (Groene AI).
• Fixed the _etag virtual column for the S3Queue and AzureQueue table engines: it was declared but never populated, so SELECT _etag always returned an empty string. It now returns the object ETag, like the S3 engine and the s3 table function. #108625 (Groene AI).
• Fixed reading Arrow, ArrowStream and Arrow-based ORC files whose timestamp column carries a fixed numeric UTC offset (e.g. +05:30, -08:00, 00:00) or the non-IANA marker fixed as its timezone, which previously failed with Cannot load time zone ... (BAD_ARGUMENTS). #108633 (Groene AI).
• Fixed an exception (LOGICAL_ERROR: Join is supported only for pipelines with one output port or NOT_IMPLEMENTED: MergeJoinAlgorithm is not implemented for strictness Semi) when join_algorithm = 'full_sorting_merge' was used for a SEMI/ANTI join, including the joins produced by decorrelating an EXISTS correlated subquery. full_sorting_merge now declines strictness/kind combinations it cannot execute, so the planner falls back to another enabled algorithm or reports that the join cannot be executed. #108658 (Groene AI).
• Fixed a heap buffer overflow when building a text index with positions = 1 over many distinct short tokens. #108659 (Groene AI).
• A query reading from or writing to S3 that is cancelled (e.g. by KILL QUERY) while an S3 request is in flight is now reported as cancelled instead of with a misleading network/S3 error. In particular, a cancelled backup or restore to S3 now reports BACKUP_CANCELLED/RESTORE_CANCELLED in system.backups. #108673 (Julia Kartseva).
• Fixed a startup failure where a DataLakeCatalog database created by an older version (25.12 or earlier) with a malformed auth_header could not be attached after upgrading to 26.2 or later, preventing the server from starting. The auth_header is now validated only on CREATE, and on ATTACH the catalog is built lazily on first use instead of during startup, so a single misconfigured or unreachable catalog database no longer blocks server startup. #108674 (Groene AI).
• Fixed a logical error (Sort order of blocks violated) during a merge, and incorrect reads, when a non-nullable LowCardinality element of a Nullable(Tuple(...)) column is used as a subcolumn (for example as a sort key). #108679 (Groene AI).
• Fix a possible crash (use-after-free) when SYSTEM FLUSH DISTRIBUTED runs concurrently with DROP TABLE of the same Distributed table. #108684 (Groene AI).
• Fix a crash in Base64 decoding of the Azure account key, reachable via azureBlobStorage (and related Azure table functions and storages) when the key is not valid Base64: a byte >= 0x80 caused an out-of-bounds read, and any invalid byte caused undefined behaviour. Invalid keys are now rejected cleanly. #108716 (Groene AI).
• Fix NUMBER_OF_COLUMNS_DOESNT_MATCH error when a subquery on a Distributed table reads two or more ALIAS columns that expand to the same expression (for example a1 String ALIAS toString(x), a2 String ALIAS toString(x)) and the subquery feeds an outer query, e.g. SELECT count() FROM (SELECT a1, a2 FROM dist GROUP BY a1, a2). #108725 (Groene AI).
• Fixed CREATE OR REPLACE MATERIALIZED VIEW ... POPULATE leaving the new view unsubscribed from its source table, which silently dropped every row inserted after the replace. #108728 (Alexey Milovidov).
• Fix an exception (Cannot find sharding key column, and in debug and sanitizer builds a server abort) when a Distributed table’s sharding key is an expression the analyzer const-folds, for example if(1, toInt32(id), toInt32(id) + 1), with optimize_skip_unused_shards = 1. #108737 (Groene AI).
• Fixed a logical error (Bad cast from type DB::ColumnNullable to DB::ColumnString) and possible wrong results when using group_by_use_nulls with a LowCardinality constant grouping key in GROUPING SETS/ROLLUP/CUBE. #108771 (Alexey Milovidov).
• Fixed partition and primary key pruning being silently disabled when a LowCardinality(FixedString) (or LowCardinality(Nullable(FixedString))) key column is wrapped in a function in the key, for example PARTITION BY sipHash64(k) % N with WHERE k = 'literal'. Such queries scanned all partitions instead of pruning them. #108777 (Groene AI).
• Fixed a logical error (a server abort in debug and sanitizer builds) when an INSERT ... SELECT through an Alias table, or into a TimeSeries table, is cancelled without an exception, for example with timeout_overflow_mode = 'break'. #108783 (Groene AI). #108796 (Shaohua Wang).
• Fix a Bad cast from type DB::ColumnSparse to DB::ColumnString logical error (a server abort in debug and sanitizer builds) when groupConcat is applied to a Tuple whose String element is stored sparse. #108790 (Groene AI).
• Fix TYPE_MISMATCH in mapFilter, mapSort, mapReverseSort, mapPartialSort and mapConcat when a Map value contains a nested Map with LowCardinality, e.g. mapFilter((k, v) -> 1, map('a'::LowCardinality(String), map('x'::LowCardinality(String), 'y'))). #108798 (Groene AI).
• Fixed insert_quorum = 'auto' not rejecting inserts up front when fewer than a majority of replicas were alive. Such inserts now fail immediately with TOO_FEW_LIVE_REPLICAS instead of writing a local part and later timing out with UNKNOWN_STATUS_OF_INSERT. #108800 (Gagan Dhakrey).
• Fixed a bug where the WITH TOTALS row of a JOIN could contain default values (such as 0) instead of the constant values coming from a constant subquery on one side of the join. The wrong result appeared only for some join orderings or with the query_plan_join_swap_table setting. #108807 (Vladimir Cherkasov).
• Fixed a segmentation fault (null-pointer dereference) that could occur when server shutdown ran concurrently with the shutdown of an S3Queue/AzureQueue table. #108810 (Miсhael Stetsyuk).
• Fixed a startup abort (Cannot allocate ThreadStack, EINVAL) on glibc builds running on CPUs with a large signal-stack size (for example Intel Sapphire Rapids / Granite Rapids with an AMX-aware kernel), where the signal alt-stack size was not rounded up to a multiple of the page size. #108813 (Groene AI).
• Fix distributed queries occasionally failing with UNEXPECTED_PACKET_FROM_SERVER (expected TablesStatusResponse, got ProfileInfo) when a connection that a previous cancelled query left out of sync was reused from the pool. #108854 (Alexey Milovidov).
• Fixed a logical error (an exception in release builds, a server abort in debug and sanitizer builds) that could happen for a function call with more than one LowCardinality argument over a distributed table, for example concatAssumeInjective((SELECT toLowCardinality('p')), s) used as a GROUP BY key with the remote table function. The messages were Default functions implementation for LowCardinality is supported only with a single LowCardinality argument or Expected the argument ... to have N rows, but it has M. #108871 (Groene AI).
• Fix reinterpret(x, 'Decimal128(scale)') (and Decimal32/Decimal64/Decimal256/DateTime64 targets) producing a result column whose internal scale was the source scale instead of the requested target scale when the source and target had the same physical type. The values were correct, but the column object was structurally inconsistent with its declared type. #108878 (Groene AI).
• Fixed a LOGICAL_ERROR (New empty part is about to materialize but the directory already exist) that could abort the server in debug and sanitizer builds when a DROP/DETACH/MOVE/REPLACE PARTITION on a MergeTree table ran after a previous such operation was interrupted (for example a rolled-back transaction or a crash) and left a stale tmp_empty_<part> directory behind. The stale directory is now reclaimed instead of failing. #108879 (Groene AI).
• Fixed ClickHouse Keeper returning incorrect dataLength = 0 in the Stat of Create2 responses, and the ClickHouse ZooKeeper client not deserializing the Stat of Create2 responses. #108909 (unintended).
• Fixed clickhouse-local rejecting an empty option value written adjacent to = (for example --format_csv_null_representation=''); it now behaves the same as --format_csv_null_representation "" and SET format_csv_null_representation = ''. #108910 (Vismay).
• Fixed a LOGICAL_ERROR (Inconsistent AST formatting) that aborted the server in debug and sanitizer builds when formatting a Tuple data type that mixes named and unnamed elements (for example Tuple(a UInt8, UInt16)). #108915 (Groene AI).
• Fix a segmentation fault when merging uniqExact aggregate states with GROUPING SETS, ROLLUP or CUBE and max_threads > 1. #108928 (Raúl Marín).
• Fixed possible wrong results or out-of-bounds reads when an INSERT is rolled back after a mid-batch error (for example in Buffer tables, asynchronous inserts, or the Kafka/RabbitMQ/FileLog engines) while lazy column replication (enable_lazy_columns_replication) is in effect. #108935 (Groene AI).
• join_any_take_last_row is now respected by all supported hash-based join paths, including joins that use automatic spilling to disk. #108936 (János Benjamin Antal).
• Fixed a bug in the analyzer where FINAL on one table of a JOIN (e.g. FROM t1 FINAL JOIN t2) was incorrectly applied to the other joined tables as well, which could make such queries slower. #108979 (Vladimir Cherkasov).
• Fix a LOGICAL_ERROR (“Cannot find __grouping_set column in header of MergingAggregatedTransform with grouping sets”, or “Chunk info was not set for chunk in MergingAggregatedTransform”) that could occur, with the analyzer disabled, for a UNION ALL/INTERSECT/EXCEPT where one branch uses GROUP BY GROUPING SETS with parallel replicas and another branch uses FINAL. #109003 (Groene AI).
• Fixed a LOGICAL_ERROR in the automatic parallel replicas planner that could occur when automatic_parallel_replicas_mode is enabled together with parallel_replicas_min_number_of_rows_per_replica greater than 0. #109011 (Groene AI).
• Fixed an excessive memory allocation during schema inference of the MsgPack format: a corrupted input whose array/map/string/binary header declares a huge element count no longer drives a single multi-gigabyte allocation and is rejected as malformed input. This fixes an out-of-memory condition in release builds (and an allocation-size-too-big abort under sanitizers). #109019 (Groene AI).
• Fixed use_client_time_zone being ignored for DateTime/DateTime64 string literals interpreted on the server (asynchronous INSERT, SELECT literals). The client now propagates its local time zone as session_timezone when use_client_time_zone is enabled, so server-side parsing matches the synchronous INSERT path. #109051 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK error when query_plan_optimize_lazy_final is enabled together with a WHERE filter partially pushed to PREWHERE (via optimize_move_to_prewhere_if_final). #109073 (Groene AI).
• Fixed an error when using a qualified asterisk like b.* in a query where b is a table (or subquery/CTE) alias that shares its name with a non-compound column of another table. Such queries no longer fail and correctly select all columns of b. #109078 (Vladimir Cherkasov).
• Fix a Block structure mismatch in JoinStep logical error (a server abort in debug and sanitizer builds) when a correlated subquery is decorrelated into a join and the source relation carries a column name more than once, e.g. WITH t AS (SELECT number, * FROM numbers(3)) SELECT *, (SELECT t.number WHERE t.number >= 0) FROM t with correlated_subqueries_default_join_kind = 'right'. #109114 (Groene AI).
• Fixed a hang where a query reading through the filesystem cache could keep waiting on an in-progress download after being cancelled with KILL QUERY, and where dropping or SYSTEM STOP VIEW-ing a refreshable materialized view whose refresh was stuck in such a wait would block. #109116 (Murphy).
• Fixed the parallel_view_processing setting being ignored when inserting into a table with materialized views — since 25.10, views always ran in parallel regardless of the setting. #109170 (Antonio Andelic).
• Fixed a parser inconsistency where an INSERT column list accepted a qualified column matcher written as t.* LIKE '<pattern>' / t.* ILIKE '<pattern>' but rejected its canonical t.COLUMNS('<regexp>') form, breaking the query format round-trip (and aborting the server in debug and sanitizer builds). #109176 (Groene AI).
• Fixed non-monotonic ProfileEvents increments: the NetworkReceiveBytes and AIOWriteBytes counters could be incremented by a negative amount (wrapping around to a huge value) on a socket timeout or an AIO failure, and MemoryOvercommitWaitTimeMicroseconds was measured with a non-monotonic clock. #109180 (Azat Khuzhin).
• Fix schema inference for the Arrow, ArrowStream, Avro formats and the legacy ORC and Parquet readers returning Nullable(Tuple) for nullable struct columns while the Nullable(Tuple) type is not allowed (allow_experimental_nullable_tuple_type is disabled). DESCRIBE returned a type that CREATE TABLE rejects, so creating a table or inserting data using the inferred schema failed with the error Nullable Tuple type is not allowed. #109185 (Nihal Z. Miaji).
• Fixed a wrong result of correlated EXISTS and scalar subqueries when the correlation appears only in the subquery projection together with a non-correlated WHERE clause (the subquery could evaluate to false / NULL for every row). Fixes #105760. #109186 (Dmitry Novik).
• Fixed CREATE USER ... HOST LIKE 'a', 'b' (and the equivalent ALTER USER) silently keeping only the first pattern. All specified HOST LIKE patterns are now stored in the user’s host allow-list. #109187 (Groene AI).
• Fixed a text index defined on mapValues(map) or mapKeys(map) being silently not used when a table was queried through a Distributed engine table with the analyzer. The index was used for the local table and via the cluster/remote table functions, but a query through a Distributed engine table skipped it (and failed with INDEX_NOT_USED under force_data_skipping_indices). #109188 (Groene AI).
• Fixed a false LOGICAL_ERROR (There was an error on <host>: Cannot obtain error message (probably it's a bug)) that could abort the server in debug and sanitizer builds when running a DDL query on a Replicated database and its per-host finished status node in Keeper had already been cleaned up. #109192 (Groene AI).
• Fix SAMPLE ratios with an exponent whose magnitude overflows Int32 (e.g. SAMPLE 1e-3000000000) being silently treated as SAMPLE 1. #109197 (Raúl Marín).
• Fix the access check for SYSTEM PREWARM PRIMARY INDEX CACHE ... ON CLUSTER, which incorrectly required the SYSTEM PREWARM MARK CACHE privilege instead of SYSTEM PREWARM PRIMARY INDEX CACHE. #109198 (Raúl Marín).
• Fixed ALTER queries failing for MergeTree tables created with a custom disk setting (SETTINGS disk = disk(...)). Fixes #63019. #109199 (Mikhail Artemenko).
• Fixed a server crash when states of the aggregate functions cramersV, cramersVBiasCorrected, contingency or theilsU from a window context (OVER ()) and from a plain aggregation are combined through chained set operations (EXCEPT, INTERSECT) or nested UNION ALL and then read, for example via the -Merge combinator. #109200 (Groene AI).
• Fix a wrong result for ANY JOIN with a constant ON condition (e.g. t1 ANY INNER JOIN t2 ON 1): the query returned the full Cartesian product instead of the ANY join result. #109207 (Vladimir Cherkasov).
• Fixed a server crash on asynchronous insert with deduplication when optimize_on_insert makes the inserted block empty (for example, rows summing to zero in SummingMergeTree). #109229 (Den Kalantaevskii).
• Fix SYSTEM DROP REPLICA and SYSTEM DROP DATABASE REPLICA ... FROM ZKPATH to reject empty or root-only ZooKeeper paths at parse time with BAD_ARGUMENTS, route auxiliary Keeper paths to the named Keeper, and canonicalize Keeper paths so the local self-protection checks cannot mis-target or bypass the intended replica. #109230 (Groene AI).
• Fix S3 settings priority so a URL-scoped <s3> endpoint block takes precedence over the top-level <s3> defaults. #109251 (Bharat Nallan).
• Fixed a bug where column DEFAULT values were not applied for INSERT INTO TABLE FUNCTION (for example remote or file) with inline VALUES data when the server parses the inline data itself (send_table_structure_on_insert_with_inline_data = 0). An explicit NULL inserted into a non-Nullable column with a DEFAULT became 0 instead of the declared default. It now behaves like a plain table INSERT and the HTTP protocol. #109258 (Groene AI).
• Fix CREATE TABLE ... AS SELECT from s3Cluster (and other cluster table functions such as fileCluster/urlCluster) failing with NOT_FOUND_COLUMN_IN_BLOCK inside a Replicated database. #109266 (Groene AI).
• Fixed SHOW TABLES and SELECT ... FROM system.tables returning different results for data lake catalog databases: tables that ClickHouse cannot read are now consistently hidden from both. #109273 (Smita Kulkarni).
• Fixed a bug where merging a MergeTree table that has projections while the enable_block_number_column or enable_block_offset_column setting is enabled produced projection parts containing a spurious _block_number/_block_offset column. The merged projection part then no longer matched the projection definition and the insert-produced projection parts, so CHECK TABLE and OPTIMIZE ... DRY RUN reported it as corrupted (CORRUPTED_DATA). #109284 (Groene AI).
• Allow reading an Avro array of two-field {key, value} records into a ClickHouse Map column. This is the encoding Iceberg and Spark produce for a MAP<K, V> with a non-string key (Avro native maps only support string keys). Previously such a file could be read as Array(Tuple(key, value)) but not as Map(K, V), which failed with Type Map(...) is not compatible with Avro array. #109289 (Groene AI).
• Fixed a server crash on CREATE HYPOTHETICAL INDEX ... TYPE set (and ngrambf_v1/tokenbf_v1) when the required index argument is omitted. Such statements are now rejected with a clear error. #109294 (Groene AI).
• Fixed a LOGICAL_ERROR (Current component is empty) that could be raised by mutation operations such as KILL MUTATION on a MergeTree table that stores its metadata in Keeper, when the enforce_keeper_component_tracking server setting is enabled. #109297 (Groene AI).
• Fixed reading ORC timestamps beyond ~year 2262 into a DateTime64 column with a scale coarser than 9. The native ORC reader used to convert every timestamp through a fixed DateTime64(9) intermediate, which overflows Int64 at nanosecond scale and rejected such values with VALUE_IS_OUT_OF_RANGE_OF_DATA_TYPE, even when the requested DateTime64 scale (for example DateTime64(6), as produced by Iceberg) can represent the value. Timestamps are now read directly at the requested scale, and date_time_overflow_behavior is honored when even the target scale cannot hold the value. #109302 (Groene AI).
• clickhouse-local acting as a server (after SYSTEM START LISTEN HTTP) now serves HTTP connections using the TabSeparated default output format, matching clickhouse-server, instead of the interactive PrettyCompact default. This fixes connecting to clickhouse-local over HTTP with drivers such as clickhouse-connect. The interactive terminal session still renders results as PrettyCompact. #109362 (Alexey Milovidov).
• Fixed a transient Code: 499 ... InvalidPart (S3_ERROR) failure on S3 multipart uploads. MinIO can briefly report a just-uploaded part as missing on CompleteMultipartUpload; this error is now retried like the already-handled NoSuchKey, bounded by s3_max_unexpected_write_error_retries. #109364 (Groene AI).
• Fix UNKNOWN_IDENTIFIER error on ALTER TABLE ... DROP COLUMN when another column has a DEFAULT or MATERIALIZED expression that defines and references an inline alias. #109374 (Alexey Milovidov).
• Fix a data race (and resulting rare crash) in the parallel merge of uniqExact aggregate states with two-level sets, which could occur with GROUPING SETS, ROLLUP and CUBE. #109389 (Groene AI).
• Fixed an error (NOT_FOUND_COLUMN_IN_BLOCK, or std::bad_function_call in older versions) when running CREATE TABLE ... ON CLUSTER ... AS SELECT reading a Distributed table on a cluster with two or more shards. #109407 (Alexey Milovidov).
• Fixed merges for tables with TTL ... GROUP BY (rollup): parts that were already fully aggregated could be rescheduled for merging again and again indefinitely, and tables with MATERIALIZED columns or enabled persistent virtual columns were handled incorrectly — such columns are now aggregated with any. Closes #105647. #109410 (Mikhail Artemenko). #109532 (Mikhail Artemenko).
• Fixed IN with a bare array column on the right argument silently returning a wrong (always-false) result or throwing an exception, so that x IN arr behaves like has(arr, x). #109416 (Alexey Milovidov).
• Fix UNKNOWN_IDENTIFIER errors on ALTER TABLE operations (RENAME/ADD/MODIFY COLUMN, and MATERIALIZE INDEX on a table with freshly inserted parts) for MergeTree tables with an implicit min-max index over the persistent virtual columns _block_number/_block_offset (enabled by add_minmax_index_for_block_number_column/add_minmax_index_for_block_offset_column). #109428 (Groene AI). #110236 (Groene AI).
• Fix a LOGICAL_ERROR in arrayFold over a non-const Array(LowCardinality(T)) argument (for example arrayFold((acc, x) -> acc + x, materialize([1, 2, 3]::Array(LowCardinality(Int64))), toInt64(0))), which failed with Arguments of 'plus' have incorrect data types (a server abort in debug and sanitizer builds). #109462 (Groene AI).
• Fixed CHECK TABLE on a ReplicatedMergeTree table reporting a healthy part as broken when a transient, retryable ZooKeeper error (e.g. connection loss) occurred during the check; such errors are now surfaced as a retryable query error instead. #109465 (Alexey Milovidov).
• Fix segfaults in groupArrayLastMerge and the largestTriangleThreeBuckets aggregate function: deserialization of aggregate function states is now validated, so a broken state does not lead to an out-of-bounds memory access. #109485 (Miсhael Stetsyuk). #109492 (Miсhael Stetsyuk).
• Fixed the global max_rows_in_join / max_bytes_in_join limit not being enforced for the parallel_hash join algorithm, where a query with join_overflow_mode = 'throw' could silently succeed instead of raising SET_SIZE_LIMIT_EXCEEDED. #109488 (Hechem Selmi).
• Fixed reading Paimon tables partitioned by a BIGINT column whose value does not fit into Int32. Such partition values were truncated (for example 9223372036854775807 became -1), which produced a wrong partition path and failed with a filesystem error. #109510 (Groene AI).
• Fix a crash when reading Iceberg tables with equality delete files. If a column is nullable in the equality delete file but non-nullable in the table schema (or vice versa), the values read from the delete file were inserted into a column of a different type through an unchecked cast (a column type confusion), corrupting the column and crashing the server. #109551 (Miсhael Stetsyuk).
• Fixed the MySQL connect_timeout, read and write timeouts not taking effect: a connection attempt to an unresponsive MySQL server could hang for more than two minutes even with connect_timeout = 1, because the sampling query profiler’s periodic signals kept resetting the MySQL client’s internal poll deadline. #109592 (Shaohua Wang).
• Fixed a LOGICAL_ERROR (creation_csn is not set while removal_csn is set to 1) that could be thrown when a non-transactional TRUNCATE ran concurrently with an uncommitted transaction that had inserted into the same table. Such a TRUNCATE no longer removes parts created by not-yet-committed transactions. #109598 (Tuan Pham Anh).
• Fix TYPE_MISMATCH errors when reading Parquet struct columns whose physical nullability differs from the requested ClickHouse type: a non-nullable Parquet group can now be read as Nullable(Tuple(...)), and the native Parquet reader can read a physically nullable struct (a Parquet OPTIONAL group) as Nullable(Tuple(...)). #109615 (Groene AI). #109898 (Groene AI).
• Fixed silent data loss when an INSERT ... VALUES in a multi-query stream is followed by a trailing SQL comment (for example VALUES (1) -- comment) under the server-side inline insert parsing path (send_table_structure_on_insert_with_inline_data = 0). The trailing comment was scanned as row data past the terminating ;, causing the following queries in the stream to be silently skipped. #109643 (Groene AI).
• Fix a segfault when a query arrives in a narrow window during server shutdown after an error at startup. #109675 (Miсhael Stetsyuk).
• Fixed stale skip indices (text, bloom_filter, etc.) and projections after materializing lightweight updates with ALTER TABLE ... APPLY PATCHES. Previously the index and projection files that depend on a patch-updated column were left unchanged, so queries using them could return wrong results (e.g. hasToken missing an updated row, or a projection returning stale aggregates once the spent patch part was removed). #109709 (Groene AI).
• Fix system.tables silently skipping databases for users with per-database grants when the query reads only the name/database columns. Introduced in 26.2. #109723 (Samay Sharma).
• Fix a LOGICAL_ERROR when writing a LowCardinality(Time) column to the Arrow format with output_format_arrow_low_cardinality_as_dictionary = 1. #109730 (Groene AI).
• Fix ObjectStorageQueue/AzureQueue (and the azureBlobStorage table function) silently skipping objects when the object storage returns an empty listing page together with a continuation token. Azure Blob Storage does this (e.g. when a listing crosses an internal partition boundary), and the async list iterator was treating the empty page as the end of the listing and dropping the token. It now follows the continuation token. #109761 (Gal Ben Moshe).
• Fixed partition pruning returning no rows for Iceberg tables partitioned by a timestamp (DateTime64) column. #109764 (Den Kalantaevskii).
• Fix an ILLEGAL_COLUMN exception in the conv function for FixedString arguments; they are now correctly cast to String. Closes #109670. #109771 (hp).
• The getClientHTTPHeader function now treats header names as case-insensitive, according to RFC 9110; in particular, the authorization header is now filtered out regardless of case. Closes #103957. #109791 (Mikhail f. Shiryaev).
• Fix max_execution_time (with timeout_overflow_mode = 'throw') sometimes never cancelling a query: when the internal timeout watcher was already waiting for a query with a later deadline, a query with an earlier deadline registered afterwards could be missed entirely, letting it run long past its time limit. #109792 (Shaohua Wang).
• Fixed the file name of gzip-compressed Iceberg metadata files. ClickHouse wrote them as v{N}.gzip.metadata.json (the HTTP Content-Encoding token), while the Iceberg spec expects the gz extension v{N}.gz.metadata.json. As a result Spark and other Hadoop-catalog readers could not find the metadata written by ClickHouse. ClickHouse now writes v{N}.gz.metadata.json and still reads the legacy gzip name for backward compatibility. #109812 (Groene AI).
• Fixed NOT_IMPLEMENTED error (“Method getDataAt is not supported for Nullable(String)”) that could be thrown by a join with enable_join_runtime_filters = 1 (on by default since 26.2) when the build-side join key was LowCardinality(Nullable(...)) with NULL values and the runtime filter fell back to its bloom filter. #109824 (Groene AI).
• External database engines (e.g. PostgreSQL) no longer push down range comparisons (>=, >, <=, <) on UUID columns. ClickHouse and external databases order UUIDs differently, so pushing these down silently dropped rows from the result; such predicates are now evaluated by ClickHouse instead. #109833 (Vismay).
• Fixed a server crash when a lambda expression was passed where a higher-order function expects a concrete value (for example the accumulator of arrayFold, as in arrayFold(lambda, arr, another_lambda)). Such queries are now rejected with ILLEGAL_TYPE_OF_ARGUMENT instead of crashing when enable_analyzer = 0. #109840 (Groene AI).
• Fixed the icebergLocal table function and IcebergLocal engine declaring azure instead of local as their object storage type, which made them fail the disk type check (Disk type doesn't match) when used with a local disk via SETTINGS disk = '...'. #109872 (Pedro Ferreira).
• Fix reading Iceberg tables partitioned by the same source column more than once (for example PARTITIONED BY (hours(ts), ts)). Such tables previously failed with Cannot add column ...: column with this name already exists (ILLEGAL_COLUMN). #109895 (Groene AI).
• Fix wrong query results caused by the primary key index incorrectly pruning granules for tables with a reversed (DESC) key column when parts have no final mark (non-adaptive granularity, index_granularity_bytes = 0). #109901 (Nihal Z. Miaji).
• Fixed generateRandomStructure occasionally producing an invalid structure string with two data types concatenated (for example Decimal32(7)IPv4) when type nesting exceeded the internal depth limit, which made the result unparseable. #109928 (Groene AI).
• Fixed NOT_IMPLEMENTED error (Method getDataAt is not supported for Nullable(String) in case if value is NULL) when a text index is built on Array(LowCardinality(Nullable(String))) (including Nested fields stored that way) and an indexed array contains a NULL element. NULL array elements are now skipped during index construction, matching Array(Nullable(String)). #110055 (Groene AI).
• Fix wrong results when a minmax skip index is built on a LowCardinality(Nullable(...)) column: WHERE/PREWHERE/HAVING ... IS NULL predicates pushed to storage previously pruned every granule and returned 0 rows even though the column contained NULL values. #110061 (Groene AI).
• Fix distinct_overflow_mode = 'break': DISTINCT now returns the partial result accumulated up to the limit and stops reading the source, as documented. Previously the chunk that crossed max_rows_in_distinct / max_bytes_in_distinct was discarded (truncated or empty results) and the query kept reading and inserting into the hash set to the end of the input, which could end in MEMORY_LIMIT_EXCEEDED. #110075 (Sergey Kuznetsov).
• Fixed a logical error ChunkInfoRowNumbers does not exist on OPTIMIZE TABLE of an Iceberg table containing a data file in a non-Parquet format (e.g. ORC) newer than all position delete files. #110107 (Alexey Milovidov).
• Fix max_bytes_in_distinct and max_bytes_in_set: string keys are stored in an arena that was not counted by the limit checks, so DISTINCT / IN over string keys could hold memory exceeding the byte limit by the whole key payload (unbounded in the key length). The limits now account for the arena, matching their documentation; queries with string keys close to a byte limit may now trip it earlier (correctly). #110120 (Sergey Kuznetsov).
• Fix a LOGICAL_ERROR (Bad cast from type ColumnLowCardinality to ColumnString) when identity (or a scalar subquery result) wraps a value containing a nested LowCardinality and the query uses WITH TOTALS/WITH ROLLUP. #110138 (Groene AI).
• Fixed reading Npy files with zero-sized inner dimensions: the number of materialized rows and the optimized and non-optimized count results now agree. #110146 (Yanjun Qiu).
• Fixed a quadratic-time blowup (and unresponsiveness to max_execution_time) in aggregation in order when grouping by multiple keys whose sort order is only a prefix of the grouping key. #110159 (Alexey Milovidov).
• Fix a logical error (Unexpected number of columns in result sample block) when a filter containing a correlated subquery (for example exists((SELECT ...))) is optimized with convert_query_to_cnf or optimize_and_compare_chain. #110187 (Alexey Milovidov).
• Fixed an exception (UNION mode UNION_DEFAULT must be normalized) when a DELETE or ALTER UPDATE mutation used a set operation (UNION/UNION ALL/UNION DISTINCT/EXCEPT/INTERSECT) inside a subquery in its WHERE condition or in an UPDATE assignment. #110196 (Alexey Milovidov).
• Fixed a ReadBuffer is canceled. Can't read from it. logical error (server abort in debug/sanitizer builds) that could occur while reading a zip archive (e.g. during RESTORE from a zip backup) after a prior read from the same archive failed mid-stream. #110197 (Groene AI).
• Fixed slow server shutdown and unresponsive KILL MUTATION when a mutation with x IN (subquery) was building the subquery set during primary-key analysis; the set build is now cancelled promptly. #110198 (Alexey Milovidov).
• Keeper now rejects the coordination settings max_requests_batch_size and max_requests_append_size being set to 0 instead of getting stuck in an infinite append-entries loop in a multi-node setup. #110200 (Alexey Milovidov).
• Fixed a NOT_FOUND_COLUMN_IN_BLOCK error when using GROUP BY ALL over a tuple expression together with ORDER BY. #110206 (Alexey Milovidov).
• Fixed a NOT_FOUND_COLUMN_IN_BLOCK error on INSERT into a table that has a CHECK constraint referencing a subcolumn (such as x.null of a Nullable column or arr.size0 of an Array). #110208 (Alexey Milovidov).
• Fix reading Iceberg tables whose default sort order references a column that needs quoting (e.g. @timestamp). Such tables were unreadable because the synthesized storage ORDER BY was built from the raw column name and failed to parse with SYNTAX_ERROR. #110233 (Groene AI).
• Fixed system.tables for a DataLakeCatalog (Iceberg/Glue) database aborting the whole query, or silently dropping tables, when a single table’s metadata is unresolvable. Such a table now stays listed by name with default/NULL values for the columns that need the opened storage object (engine, total_rows, etc.), regardless of database_datalake_require_metadata_access. Direct access to the broken table still reports the error. #110242 (Groene AI).
• Fixed a possible crash (heap-buffer-overflow) when a quantileTDigest-family aggregate-function state column was used as a GROUP BY key and serialized concurrently by several threads. #110263 (Groene AI).
• Fixed INSERT into a Microsoft Fabric / OneLake DataLakeCatalog table failing with IncorrectEndpointError (HTTP 400) by routing ADLS Gen2 (DFS) writes to the .dfs endpoint host instead of the .blob host. Note: with remote_url_allow_hosts, both the .blob (read) and .dfs (write) Fabric hosts must be allow-listed for INSERT. #110290 (Mohammad Lareb Zafar).
• Fixed a server abort when parsing certain PRQL queries (SET dialect = 'prql'). A panic inside the prqlc compiler was aborting the process instead of being reported as a query error. #110316 (Groene AI).
• Fix silently wrong query results when a projection part written before the projection’s column set changed (for example, after upgrading across versions that materialize an ALIAS column’s source differently, or after ALTER TABLE ... MODIFY COLUMN re-points an ALIAS column used by a projection) was read through the projection or merged: the missing column was filled with default values instead of the real data. Such parts are now read from the base table, and merges rebuild the projection from the base data. #110328 (Shaohua Wang).
• Fixed ORDER BY ... WITH FILL not respecting max_execution_time and being slow to cancel when generating a large fill range (especially with INTERPOLATE). #110332 (Alexey Milovidov).
• Fix a logical error (a server abort in debug/sanitizer builds) when a multi-command ALTER such as UPDATE ..., DELETE ... was issued on an Iceberg table. Such mutations are now rejected with a clear NOT_IMPLEMENTED error. #110347 (Groene AI).
• Fixed a parser bug where COMMENT was incorrectly consumed as an implicit alias when a SELECT query ended directly after a bare table identifier (e.g. ... FROM t COMMENT 'x'), causing a misleading syntax error instead of the comment being applied to the view/table. #110372 (Aditya Kumar).
• Fixed a server crash (native stack overflow) when a deeply nested Array/Tuple/Map/Object literal is copied or destroyed, for example a query with a very deeply nested literal at a raised max_parser_depth. #110393 (Raúl Marín).
• Fixed two cases where vector search could return fewer rows than expected: vector_search_index_fetch_multiplier values below 1.0 could truncate the computed fetch count to zero and produce an empty result (such values are now rejected), and the vector search optimization is now skipped for LIMIT ... WITH TIES queries, because the optimization bounds the search to exactly N candidates and dropped rows tied with the N-th row. #110452 (Tamish Mhatre). #110453 (Tamish Mhatre).
• Fix the logical error Expected CommonSubplanReferenceStep to reference CommonSubplanStep, the error Subplan cannot be used to build pipeline, and the logical error Trying to extract chunk from ChunkBuffer before all inputs are finished for IN (subquery) where the subquery contains a correlated subquery and the set is built during index analysis. #110491 (Alexey Milovidov).
• Fix substitution of query parameters inside the definitions of a named WINDOW clause. Previously, parameters there were silently kept unsubstituted, and an unset Identifier parameter could lead to the exception Logical error: '!part.empty()' during EXPLAIN SYNTAX. #110506 (Alexey Milovidov).
• Fixed server startup failure for MergeTree tables on object storage disks when a leftover txn_version.txt.tmp file has broken disk-level metadata. #110519 (Alexey Milovidov).
• Fix a logical error in the filesystem cache (Expected file ... not to exist) that could occur when a background cache download failed with a non-filesystem error (for example while running OPTIMIZE on an Iceberg table), leaving an empty orphan cache file behind. #110549 (Groene AI).
• Fixed reading a corrupted Native-format stream whose Dynamic type count or JSON/Object path count is close to SIZE_MAX: such malformed input is now rejected with a clear INCORRECT_DATA error instead of an uncaught std::length_error. #110590 (Alexey Milovidov).
• Fixed a server crash in the MaterializedPostgreSQL database engine that could happen when a table was detached (or its structure changed) while it was still queued for synchronization. #110596 (Alexey Milovidov).
• Fixed SET param_<name> in the client (and --param_<name> flags) misbehaving when the parameter name matches a builtin setting name such as limit, offset, max_threads, or log_comment. Depending on the colliding setting, the query failed with CANNOT_PARSE_QUOTED_STRING, the value was silently normalized (e.g. max_threads=0 became auto(N)), or a setting alias name was lost. Such a parameter also broke every subsequent SET param_* in the same session. #110597 (Raúl Marín).
• Fix reading subcolumns of a column that has a DEFAULT expression and is not materialized in a part (e.g. after ALTER TABLE ADD COLUMN or during ALTER TABLE MATERIALIZE COLUMN): the subcolumn was filled with type default values instead of the evaluated DEFAULT expression, and the transposed vector distance functions (cosineDistanceTransposed and others) on such QBit columns failed with the SIZES_OF_ARRAYS_DONT_MATCH error. This closes #110634. #110636 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR Invalid partition key size: 0 that could fail an INSERT into a partitioned MergeTree table with non_replicated_deduplication_window enabled, when an async insert batch was only partially deduplicated (some insert_deduplication_token values were duplicates and some were new). #110651 (Groene AI).
• A background asynchronous insert flush now stops promptly when killed with KILL QUERY; previously a flush of a large buffered payload kept parsing to the end and ignored the cancellation. #110652 (Shaohua Wang).
• Fixed wrong results from GROUP BY with optimize_aggregation_in_order and DISTINCT with optimize_distinct_in_order placed over a partial_merge JOIN. The read-in-order optimization was propagated through the partial-merge join, which re-sorts its left input by the join key and therefore does not preserve the left stream’s original order, so rows were grouped incorrectly. #110671 (Groene AI).
• Fixed a signed integer overflow in toStartOfInterval with an extreme MONTH, WEEK, or DAY interval count on Date/Date32/DateTime64 values. #110688 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK error for queries with a subquery in FROM when parallel replicas run in custom_key_sampling / custom_key_range mode. Such queries (e.g. SELECT * FROM (SELECT id, k, v FROM t WHERE id < 20) ORDER BY k) previously failed with Not found column __table2.id in block .... #110690 (Groene AI).
• Fixed readWKT rejecting WKT strings with leading whitespace (e.g. readWKT(' POINT(1 2)')), which are accepted by the typed readWKTPoint/readWKTPolygon/… readers and by the WKT grammar. #110706 (Groene AI).
• Fix a logical error (Column ... already added for reading, a server abort in debug/sanitizer builds) when the same text-index predicate is used in both PREWHERE and WHERE of a query over a Merge table on top of a Distributed table. #110710 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK error for the transposed distance functions over QBit columns (such as cosineDistanceTransposed) called with a scalar-subquery or constant reference vector when parallel replicas are enabled. This closes #110719. #110729 (Alexey Milovidov).
• Fix a rare LOGICAL_ERROR (Unexpected number of parts to remove from parts_queue; a server abort in debug/sanitizer builds) on ReplicatedMergeTree when two partition operations (for example two MOVE PARTITION/REPLACE PARTITION queries, or one racing a background DROP_RANGE) cancelled the background part check over overlapping ranges at the same time. #110738 (Groene AI).
• Fixed a Hash is not set for serialization logical error that could occur when a column with a Quantized(...) codec was wrapped inside another serialization, for example in a Merge table over sources whose same-named column has different types (merged into a Variant). #110776 (Groene AI).
• Fix a server hang when multiplying a median/quantile aggregate function state by a huge integer constant (e.g. medianState(x) * 18446744073709551615). The query became unresponsive to cancellation and could run indefinitely. #110779 (Groene AI).
• Fix a server crash (integer division by zero) when a quota with a zero-length interval (for example CREATE QUOTA q FOR INTERVAL 0 SECOND MAX queries = 1000) was consumed. A non-positive quota interval duration is now rejected at quota creation. #110846 (Pedro Ferreira).
• Fixed throwIf(notLike(col, pattern)) (and other throwIf over a function) throwing unconditionally when col is LowCardinality. The default LowCardinality implementation ran throwIf on the whole dictionary, which always holds the reserved default value even when no row references it, so a non-zero value in that unused slot made throwIf throw for data that does not satisfy the condition. #110864 (Groene AI).
• Propagate a query’s active roles (SET ROLE or the role parameter) to remote nodes on parallel-replica and Distributed reads over a cluster secured with an interserver secret; previously the remote nodes fell back to the user’s default roles, evaluating row policies inconsistently with the initiator. #110867 (Nikolay Degterinsky).
• Fixed Iceberg tables producing duplicate field IDs after ALTER TABLE ... ADD COLUMN when the initial schema contains nested fields (Tuple/Array/Map). last-column-id now records the maximum assigned field ID including nested children, so a subsequently added column no longer reuses a nested field ID and reads no longer fail with ICEBERG_SPECIFICATION_VIOLATION (Duplicate field id). #110884 (Groene AI).
• Fixed a crash in FIPS builds when using an Ed25519 SSH key (CREATE USER ... IDENTIFIED WITH ssh_key ... TYPE 'ssh-ed25519', or such a key in users.xml). Ed25519 is not FIPS-approved; such keys are now rejected with a clear LIBSSH_ERROR error. #110891 (Konstantin Bogdanov).
• Fix the reduced-precision results of the transposed distance functions over QBit columns (cosineDistanceTransposed, L2DistanceTransposed, dotProductTransposed). Values truncated to precision bit planes were reconstructed with the dropped bits zero-filled, which biased every reconstructed value towards zero and was degenerate at low precision: for QBit(BFloat16) at precision 1 only the sign bit survives, so every value reconstructed to ±0.0 and the distance was the same constant for every row, carrying no ranking information. Truncated values are now reconstructed to a bounded midpoint of the remaining precision cell (raw Int8 codes to their cell centre; exact zeros and infinities are preserved), so a smaller precision trades accuracy for speed without wildly biased results. Full-precision results are unchanged. This closes #110898. #110911 (Alexey Milovidov).
• Fix a server crash (use-after-free) during graceful shutdown after Iceberg tables were queried from sessions that also used temporary tables. #110914 (Miсhael Stetsyuk).
• Fixed a LOGICAL_ERROR: Cannot sum Bools (a server abort in debug/sanitizer builds) when merging sumMap/sumMapWithOverflow/sumMapFiltered aggregate states with Bool values that were serialized in the old (version 0) state format. The same fix also corrects two silent wrong-result cases with version-0 Bool states: map keys of type Bool were not deduplicated across states, and zero-value compaction dropped the wrong entries. #110922 (Groene AI).
• Fixed a crash when an aggregate function with the -Tuple combinator was used with RESPECT NULLS and produced an intermediate state (-State, distributed aggregation, WITH ROLLUP/WITH CUBE). The combined function was named after the wrong state variant, so its serialized AggregateFunction type re-resolved to an incompatible in-memory layout on a round-trip. #110930 (Groene AI).
• Fix PREWHERE (and row-level policies) on an Iceberg column that was renamed by schema evolution silently returning 0 rows for old data files. Such queries now return the correct rows, matching WHERE. #110941 (Groene AI).
• Fix a LOGICAL_ERROR (LazyMaterializingTransform: Number of rows in lazy chunk N does not match number of offsets M) that could happen for vector search queries with vector_search_with_rescoring = 1 combined with lazy materialization when several distances tie. #111003 (Groene AI).
• Fixed reading Arrow/ArrowStream data with empty nested Array/Map columns produced by Apache Arrow Java < 19.0.0 (bundled with Apache Spark), which were previously rejected with an INCORRECT_DATA error about the offsets buffer being too small. #111101 (Raúl Marín).
• Fix stale reads on plain_rewritable disks with the page cache enabled. #111105 (Mikhail Artemenko).
• Fix CANNOT_CONVERT_TYPE error for constants of Variant type (including Geometry) in distributed queries and queries with parallel replicas. #111136 (Nikita Fomichev).
• Fix a bug where reading large files from HDFS could fail or read incorrectly when a single read request exceeded the hdfsPread 32-bit size limit (INT_MAX). #98470 (Arup Chauhan).
• Fix incorrect use of the text index for equality comparisons with an empty needle. #108340 (Robert Schulze).
• Fixed a hang in SYSTEM SYNC MERGES with the Manual merge selector when a scheduled merge could not be placed into a full background pool. #108770 (Alexey Milovidov).
• Fix ATTACH TABLE ... AS REPLICATED discarding committed data and resurrecting pre-mutation rows when a stale txn_version.txt.tmp file was left on a data part. #108772 (Alexey Milovidov).
• Fix a possible race condition in ALTER TABLE ... MOVE PARTITION TO TABLE with non-replicated MergeTree tables that could produce intersecting data parts, because the commit of moved parts and block number allocation were not under the same lock. #108922 (Mikhail Artemenko).
• Fix reading Iceberg tables written by Databricks UniForm, which writes a degenerate placeholder schema (with an empty field list) into manifest files. #108945 (Konstantin Vedernikov).
• Fix a use-after-free when reading a non-partitioned Hive table with multiple threads. #109164 (Alexey Milovidov).
• Fix the console log level not being reverted after server startup when logger.startup_console_log_level is set: the level was captured from the wrong configuration key, so the console stayed at the elevated startup level for the lifetime of the server. Closes #103472. #109858 (Garrett Thomas).
• Fix snapshot-isolation anomalies in plain_rewritable disk metadata transactions. Fixes #92055. #110948 (Mikhail Artemenko).
• Fix an abort (front() called on an empty vector) in clickhouse-client --login when the host was not passed as an explicit --host argument, for example when it came from --connection, the configuration file, or the CLICKHOUSE_HOST environment variable. Closes #103603. #110279 (Christoph Wurm).
• Fix a logical error (!part.empty()) when a lambda argument is a backtick-quoted identifier whose name contains a dot (for example __table1.). Such a query previously produced a handled exception in release builds and aborted the server in debug and sanitizer builds while formatting an error message. #108735 (Groene AI).
• Fix a LOGICAL_ERROR (No set is registered for key) in ALTER TABLE ... DROP COLUMN, ALTER TABLE ... DELETE/UPDATE mutations, and lightweight DELETE FROM, on tables that have an ALIAS column whose expression uses an IN operator (including through another ALIAS column). #111039 (Pedro Ferreira).
• Fixed silent data loss with async inserts and deduplication (async_insert=1, async_insert_deduplicate=1). When several async-insert entries with distinct insert_deduplication_token values were coalesced into one flush that wrote to disjoint partitions, each token was registered in the deduplication log of every partition the flush touched, not only the partition its own rows landed in. A later insert reusing one of those tokens in a partition it never wrote to was then silently deduplicated away. Tokens are now registered only against the partition their rows actually landed in. Closes #111031. #111049 (Groene AI).
• Fixed a LOGICAL_ERROR exception (ScatterExchangeStep should have one source shard, got 8) and duplicated rows when querying a Merge table over Distributed tables or views with the experimental setting make_distributed_plan = 1. Closes #107946. #108401 (Groene AI).
• Serialize a Nullable(Tuple(...)) column as a single CSV field instead of flattening it into separate columns. Previously a non-null value was written as several CSV fields while NULL was a single \N, giving a row-dependent column count that broke CSVWithNames/CSVWithNamesAndTypes headers and round-trips. #108959 (Groene AI).
• Fix a logical error (KeyCondition uses PREWHERE output) in the Parquet v3 reader when a format-level key condition references a column that is not read from the Parquet file. #109267 (Groene AI).
• Fix wrong results for table-function reads with FINAL or SAMPLE under serialize_query_plan = 1. Also fix the aggregation hash-table statistics cache key for table functions so different table functions no longer share the same preallocation entry. #109847 (Azat Khuzhin).
• Fixed cleanup of a BACKUP ... TO Memory(...) that fails before finalization: the failed backup was left registered, so its name could not be reused. #109947 (Julia Kartseva).
• Fix extreme interval arithmetic for WITH FILL STEP and the add* / subtract* date/time interval functions: out-of-range deltas no longer rely on signed overflow, and WITH FILL with a huge YEAR step no longer stalls by running the calendar backward. #110158 (Groene AI).
• Fix 411 Length Required errors from Azure services: the Poco-based Azure HTTP transport now sets Content-Length from the request body for SDK clients that do not set it themselves, such as Azure Key Vault. #110299 (Konstantin Bogdanov).
• The lazy FINAL optimization (query_plan_optimize_lazy_final, disabled by default) is no longer applied when the query reads in sorting-key order, because its replacement plan does not preserve that order and could return incorrectly ordered results. It is also no longer applied when a LIMIT smaller than the number of rows to read applies directly to the reading step, where the mandatory set-building pass defeats early termination. #110576 (Nikolai Kochetov).
• Fixed readWKT and readWKTPoint returning uninitialized (scalar) or stale prior-row (vectorized) coordinates for POINT EMPTY. POINT EMPTY is now rejected with CANNOT_PARSE_TEXT, since ClickHouse Point is a fixed Tuple(Float64, Float64) with no empty representation. #110692 (Groene AI).
• Fix wrong timezone handling when materializing right-side columns in hash joins. Closes #111033. #111074 (Yarik Briukhovetskyi).
• Fixed wrong count() results and dropped rows when a String (or narrower FixedString) key or minmax skip index is filtered by a comparison with a wider FixedString constant, e.g. toFixedString('abc', 257) = string_col. Primary-key and minmax pruning built the range from the NUL-padded constant and wrongly skipped matching granules. #111106 (Groene AI).
• Fixed asynchronous Native-format inserts so one buffered entry that becomes incompatible after ALTER ... MODIFY COLUMN no longer causes the whole batch to fail. Closes #111064. #111108 (Mikhail f. Shiryaev).
• Fix CREATE OR REPLACE of a dictionary with an object of another kind: it failed with CANNOT_DETACH_DICTIONARY_AS_TABLE after the replace was already committed, leaving an orphan _tmp_replace_* table. #111142 (Nikolay Degterinsky).
• Fixed INCOMPATIBLE_TYPE_OF_JOIN for an ANY join on a Join engine table when a WHERE filter on a right-side column allowed the query planner to rewrite the join to SEMI or ANTI. A Join engine table has a fixed declared strictness that cannot be changed, so the conversion is now declined for such tables. #111362 (Groene AI).
• Fix the format table function returning no columns when the input data parses to zero rows (for example an empty JSON document or a GeoJSON FeatureCollection with no features). It now returns an empty result with the correct columns instead of throwing NOT_FOUND_COLUMN_IN_BLOCK. Closes #111390. #111428 (Groene AI).

​

Build/Testing/Packaging Improvement

• Sanitizer builds now capture the sanitizer report into the global buffer sanitizer_report, so that the core dump analyzer can read it from the core dump. #109333 (Miсhael Stetsyuk).
• Build musl from source instead of using vendored binaries. #97452 (Konstantin Bogdanov).
• Use SIMD memcmp/memcpy/memmove/memset/bcmp/memmem implementations from LLVM-libc. #107566 (Konstantin Bogdanov). Fix silent memmove corruption in -O0 debug builds (-DDEBUG_O_LEVEL=0) on machines with AVX-512. #110904 (Konstantin Bogdanov).
• Remove usages of the curl library across the codebase and bundled third-party libraries. #108296 (Konstantin Bogdanov).
• Updated mariadb-connector-c to 3.4.9. #108328 (Konstantin Bogdanov).
• Update libssh to 0.12.0. #108329 (Konstantin Bogdanov).
• Fix a hang on transaction COMMIT on macOS. #108375 (Raúl Marín).
• Added a COMPRESS_DEBUG_SECTIONS CMake option that compresses DWARF debug sections in object files (-gz), auto-enabled where the compiler supports it, to shrink Debug build directories; final binaries are left uncompressed, so runtime stack-trace symbolization is unaffected. #109306 (Murphy). The bundled LLVM toolchain is now built with zlib, so its clang supports -gz=zlib, and the missing clang++-21 symlink was added. #109596 (Murphy).
• Debug builds now compile the vendored Rust crates with reduced debug info (debug = 1, cargo’s limited level) and no incremental cache (REDUCE_RUST_DEBUG_INFO, on by default), shrinking Rust build artifacts (~12 → ~7 GB from scratch) and preventing unbounded cargo cache growth across rebuilds. Set -DREDUCE_RUST_DEBUG_INFO=OFF to restore full DWARF and incremental compilation. #109471 (Murphy).
• Support the streaming distributed query exchange and STREAM reads on macOS. #109769 (Raúl Marín).
• Bump libarchive from 3.8.7 to 3.8.8. #109874 (Robert Schulze).
• Bump c-ares from 1.34.6 to 1.34.8. #109905 (Robert Schulze).
• Update krb5 to fix CVE-2026-40355 and CVE-2026-40356. #109910 (Robert Schulze).
• Removed the libpng third-party dependency. FORMAT PNG now uses a small built-in PNG encoder (backed by the bundled zlib) instead. #110051 (Alexey Milovidov).
• Track raw C allocations (not going through operator new) in the memory tracker on macOS by wrapping the jemalloc malloc zone, so memory limits are enforced for them. #110370 (Raúl Marín).
• The -distroless Docker images for clickhouse-server and clickhouse-keeper are now built from the shell-free production stage; previously they were built from the debug stage and included /busybox/sh. Closes #105677. #105678 (ashishch432).
• Update abseil-cpp to LTS 20260526.0. #108991 (Konstantin Bogdanov).

​

ClickHouse release 26.6, 2026-06-25. Presentation, Video

​

Backward Incompatible Change

• The change removes allow_experimental_query_deduplication feature. The feature was experimental for a long time, has no tests and not going to be supported (initial tests were unstable and has been removed in #49579). #99398 (Igor Nikonov).
• The XRay-based SYSTEM INSTRUMENT feature now uses arguments instead of parameters: the system.instrumentation.parameters column was renamed to arguments, and the SYSTEM INSTRUMENT ADD ... HANDLER PARAMETERS ... syntax was renamed to SYSTEM INSTRUMENT ADD ... HANDLER ARGUMENTS .... Queries or automation that use the old column or syntax must be updated to use the new names; the old names are no longer supported. #103854 (Pablo Marcos).
• The setting show_data_lake_catalogs_in_system_tables has been renamed to show_remote_databases_in_system_tables and broadened: when its value is 0 (the default), MySQL and PostgreSQL databases are also hidden from system.tables, system.columns, and system.completions, in addition to data lake catalogs. The old setting name is kept as an alias. #104416 (Pablo Marcos).
• The default x86 build now targets x86-64-v3 (AVX2) instead of x86-64-v2 (SSE4.2). This requires a CPU with AVX2 support (Intel Haswell or later, AMD Excavator or later). If your CPU does not support AVX2, use the amd64compat build which targets plain x86-64. #105019 (Raúl Marín).
• Reject nested Dynamic/Variant in min/max aggregates and minmax indexes. Previously only top-level Dynamic/Variant were checked. Closes #104747. #105468 (Pavel Kruglov).
• icebergHash and icebergBucket now reject Int128, UInt128, Int256, UInt256, and Decimal256 arguments with an explicit error. Previously they silently truncated wider values and produced colliding hashes. The Iceberg spec only defines hashing for 32/64-bit integers and decimals with precision up to 38; cast to Int64 or Decimal128 to keep the previous behaviour for values that fit. #105866 (Raúl Marín).
• Added granular control over query echoing in the CLI. The --echo option now accepts an optional boolean value and works in both interactive and batch mode. New options --echo-formatted and --echo-query-id control whether echoed queries are formatted and whether the query_id is printed, respectively. The --hilite/--highlight option now also controls highlighting of echoed queries. As a side effect, --echo is now a boolean-value option, so a positional query directly after a bare --echo is treated as its value; use --echo --query "..." or --echo=false instead. #106191 (Alexey Milovidov).
• Removed the experimental KQL (Kusto) functions array_sort_asc and array_sort_desc, and their SQL backends kql_array_sort_asc and kql_array_sort_desc. These functions were experimental, implemented with low quality, and the source of correctness and parser bugs. Queries using these names now return UNKNOWN_FUNCTION. #108101 (Groene AI).
• ALTER TABLE ... REPLACE PARTITION ... FROM ... no longer silently drops the destination partition’s data when the source table has no parts in the requested partition. Previously such a request removed the destination partition and wrote nothing in its place — a data-loss footgun. It is now rejected with BAD_ARGUMENTS by default. To restore the previous silent-clear behavior, set the new setting allow_replace_partition_from_empty_source = 1, or set compatibility to 26.5 or lower; to explicitly drop destination data, use ALTER TABLE ... DROP PARTITION. #104939 (Groene AI).
• The default of the server setting insert_deduplication_version changes from compatible_double_hashes to new_unified_hash. Insert deduplication now works on the whole inserted block (per insert) rather than per part/partition: a retry of the same insert is still deduplicated, but two different inserts that produce an identical part are no longer cross-deduplicated, and reordered inserts of the same rows are no longer deduplicated. Set insert_deduplication_version = compatible_double_hashes to restore the previous behavior. Instances upgrading directly from a release that defaulted to old_separate_hashes should first run with compatible_double_hashes until the longest relevant deduplication window has elapsed before relying on new_unified_hash. #107886 (Sema Checherinda).

​

New Feature

• Added support for continuous queries over MergeTree tables, implemented via a series of snapshot reads. This is a first step toward general support for streaming queries. #105114 (Mikhail Artemenko).
• Added hypothetical (what-if) skip indexes. Use CREATE HYPOTHETICAL INDEX ... ON t (expr) TYPE ... to define a session-scoped virtual skip index, then EXPLAIN WHATIF SELECT ... to estimate its skip ratio and cost without materializing it. Defined indexes are visible in the new system.hypothetical_indexes table. #104608 (Yarik Briukhovetskyi).
• Add an embedded /schema web UI to clickhouse-server that visualizes the dependency graph between tables, materialized views, refreshable materialized views, dictionaries, distributed tables and views. #105276 (Nikita Mikhaylov).
• Added the PNG output format support, allowing query output results to be directly rendered as PNG images. #74691 (Maksim Dergousov).
• Added detection of the AI coding agent (Claude Code, Cursor, Codex, Gemini CLI, Goose, etc.) that invoked clickhouse-client or clickhouse-local, based on environment variables. The detected agent is reported in the new client_agent column of system.query_log, system.query_thread_log and system.processes. #106619 (Alexey Milovidov).
• Added a new system table system.documentation that collects the embedded reference documentation of the uniform components of the system (functions, table engines, data types, settings, formats, and others) into a single table, with the documentation rendered as Markdown. #107463 (Alexey Milovidov).
• Add the GeoJSON input format for reading GeoJSON FeatureCollection documents, producing one row per feature with id (String), geometry (Geometry), and properties (Nullable(JSON)) columns. Point, LineString, MultiLineString, Polygon, and MultiPolygon geometries are read into ClickHouse’s native Geometry type. GeometryCollection and MultiPoint, which the Geometry type cannot represent, raise an exception by default, or can be stored as NULL via the input_format_geojson_unsupported_geometry_handling setting. Contributes to #91533. #98124 (Mark Needham).
• Added functions for serving Mapbox Vector Tiles directly from SQL: MVTEncodeGeom projects a geometry into the pixel space of a slippy-map tile and clips it, MVTEncode aggregates the projected geometries of a group into the binary bytes of a single-layer tile, and MVTBoundingBox / MVTBoundingBoxMercator return the bounding box of a tile for restricting rows to it. Point, line and polygon geometry is supported. Also available under the PostGIS aliases ST_AsMVTGeom and ST_AsMVT. #106107 (Saarthak Gupta).
• clickhouse-local now supports SYSTEM START LISTEN and SYSTEM STOP LISTEN queries, allowing it to be transformed into a server accepting TCP and HTTP connections. #101143 (Alexey Milovidov).
• Added an interactive help command to clickhouse-client and clickhouse-local. Typing help <name> (also /help, man, /man) shows the documentation of a function, table engine, data type, format, setting, or other component, rendered from Markdown in the terminal with syntax-highlighted SQL. When the word is ambiguous all matches are shown, and when it is not found, similar names are suggested. #108042 (Alexey Milovidov).
• Support writing into Azure Data Lake Storage Gen2. #105406 (Konstantin Vedernikov).
• Adding RowBinaryWithNamesAndTypesAndDefaults format to get better support for schema evolution. #105736 (Mark Zitnik).
• Implement ADD ENUM VALUES in ALTER TABLE queries to simplify appending new values to an existing Enum type without the need to specify all current Enum values again. #93830 (Ilya Golshtein).
• Refreshable materialized view now supports REFRESH DEPENDS ON to trigger refreshes on another RMV’s refreshes instead of time-based schedule. (REFRESH EVERY ... DEPENDS ON already existed, but couldn’t be used reliably for this use case.) See CREATE MATERIALIZED VIEW documentation. #104440 (Michael Kolupaev).
• Add support for selecting columns by name pattern with * LIKE '<pattern>' and * ILIKE '<pattern>', including qualified forms such as table.* LIKE '<pattern>' and table.* ILIKE '<pattern>'. LIKE matches column names case-sensitively, ILIKE case-insensitively. #104569 (Yue Ni).
• Added ESCAPE syntax that defines a custom escape character in a LIKE pattern, allowing you to treat % and _ as literal characters instead of wildcards. #99774 (Ilya Yatsishin).
• Add materialize_projections_on_insert and materialize_projections_on_merge MergeTree table settings. When materialize_projections_on_insert = 0, INSERTs skip building projection parts, which improves insert throughput for tables with many projections. When materialize_projections_on_merge = 1, a merge rebuilds a projection that is missing from all of its source parts, so projections can be built during merges instead of on insert. Merges still combine only parts that share the same set of projections. #100993 (Christoph Wurm).
• Add a new immutable MergeTree setting allow_tuple_element_aggregation, disabled by default. When enabled, SummingMergeTree, AggregatingMergeTree and CoalescingMergeTree recursively flatten Tuple columns and aggregate each leaf element independently during merges, exactly as if it were a top-level column — SummingMergeTree sums it, AggregatingMergeTree merges its aggregate-function state, and CoalescingMergeTree keeps its last non-NULL value. The setting must be specified at table creation time and is silently ignored by engines that do not support it. #98039 (johnjing).
• New functions quantizeBFloat16ToInt8 and dequantizeInt8ToBFloat16: a scalar codec that compresses embedding components to 8-bit using a 256-level Gaussian Lloyd-Max quantizer, from which Int4/Int2/binary codes can be extracted by bit-truncation. #108102 (Alexey Milovidov).
• Add functions arrayTopK and arrayBottomK: - arrayTopK(k, array) returns the K largest elements in descending order - arrayBottomK(k, array) returns the K smallest elements in ascending order. #104563 (Vitaly Baranov).
• Added a new system table system.iceberg_files exposing per-file metadata for Iceberg tables, with one row per data or delete file in each table’s current snapshot. This closes #98777. #104415 (Asya Shneerson).
• Added system.constraints table that provides information about all CHECK and ASSUME constraints across all tables, including constraint name, type, and expression. #105337 (Pedro Ferreira).
• Added a new system.dictionary_layouts table that lists the available dictionary layouts together with embedded documentation (description, syntax, examples, introduced_in, related). #106182 (Alexey Milovidov).
• Added a new system.dictionary_sources table that lists the available dictionary sources together with embedded documentation (description, syntax, examples, introduced_in, related). #106184 (Alexey Milovidov).
• Added a new system.data_skipping_index_types table that lists the available data skipping index types together with embedded documentation (description, syntax, examples, introduced_in, related). #106186 (Alexey Milovidov).
• Added a new system.disk_types table that lists the available disk types together with embedded documentation (description, syntax, examples, introduced_in, related). #106187 (Alexey Milovidov).
• Implemented SYSTEM RESTART DISK <name>: it now reloads a disk’s in-memory metadata and re-scans the data parts of readonly-replica tables located on it. This lets a readonly replica of a table on shared plain_rewritable storage observe data written by another server on demand, without waiting for refresh_parts_interval or restarting the server. #106645 (Jordi Villar).
• PostgreSQL-style expr OP SOME(array) / expr OP ALL(array) (non-subquery right-hand side) is now supported and rewritten to has / NOT has for =/<>, or to arrayExists / arrayAll lambdas for other comparison operators. ANY is not accepted for the array form because any is also an aggregate function; use SOME instead. The subquery form of ANY/SOME/ALL continues to be lowered to IN / NOT IN. #105129 (Alexey Milovidov).
• Added two new load_balancing strategies, hostname_longest_common_prefix and hostname_longest_common_suffix, which prefer the replica whose hostname shares the longest common prefix (respectively, suffix) with the initiator’s hostname. They are useful when the data center is encoded as a prefix or suffix of hostnames whose numeric segments have variable length, where the existing nearest_hostname and hostname_levenshtein_distance strategies pick the wrong replica. #107360 (Denny [DBA at Innervate]).
• Added TLS client certificate information (subjects, serial number, issuer, and validity period) to the system.session_log table to improve observability of certificate-based authentication. #107679 (Alexey Milovidov).
• Add an optional external_id credential for S3 role-based access. #106941 (Elian Gidoni).
• Added a new S3Queue setting after_processing_move_preserve_path. When enabled together with after_processing='move' and after_processing_move_prefix, processed objects are moved while preserving their full source path under the destination prefix instead of being flattened to just the file name. #105354 (Asya Shneerson).
• Add message_queue_disable_insertion server setting to disable insertion from message queue engines (Kafka, RabbitMQ, NATS) into attached materialized views. Useful for read-only replica scenarios. #104911 (JIaQi Tang).
• Added LOCALTIME and LOCALTIMESTAMP (SQL-standard / PostgreSQL syntax). LOCALTIMESTAMP is an alias for now() (returns DateTime); LOCALTIME returns the current time of day as a Time value. #106139 (Thomas Cabral).
• Added date_part('unit', expr) as syntactic sugar for EXTRACT(unit FROM expr). Standard interval kinds and the PostgreSQL extras (epoch, dow, doy, isodow, isoyear, century, decade, millennium) are all supported. #105127 (Alexey Milovidov).
• PostgreSQL-compatible EXTRACT(TIMEZONE_HOUR FROM dt) and EXTRACT(TIMEZONE_MINUTE FROM dt) for the hour and minute parts of a timezone offset, and EXTRACT(<unit> FROM INTERVAL n <unit>) / date_part('<unit>', INTERVAL n <unit>) for extracting the value out of an interval. #106227 (Vinayak Joshi).
• Added min_by and max_by aggregate aliases for argMin and argMax. #105712 (Joey Yu). These aliases are stupid, real engineers use the original ClickHouse names.
• Added REGEXP_SUBSTR as a case-insensitive alias of regexpExtract for Oracle/MySQL/Snowflake compatibility. #105122 (Alexey Milovidov).
• Added SESSION_USER as a case-insensitive alias of currentUser() for PostgreSQL / SQL-standard compatibility. #106081 (Takumi Hara).
• Support a trailing NULL / NOT NULL modifier in ALTER TABLE ... ADD/MODIFY COLUMN, mirroring CREATE TABLE. #106150 (Takumi Hara).
• You can now use h3PolygonToCellsWithContainment function from h3 to use their new algorithm which supports center-based, fully-contained, and overlapping containment modes. #104455 (Youssef Kadry).
• Adds IPV4_PREFIX_BITS and IPV6_PREFIX_BITS flags when Keying done by IP_ADDRESS or FORWARDED_IP_ADDRESS. #89270 (Aditya Chopra).
• formatReadableTimeDelta now accepts an INTERVAL expression of type other than Month and Year as input. #64315 (Francisco J. Jurado Moreno).
• User can now specify an optional precision argument to the formatReadableSize, formatReadableDecimalSize and formatReadableQuantity functions, controlling the number of digits after the decimal point. Default is 2, preserving the prior behavior. #104648 (Antonio Filipovic).
• Add output_format_float_precision setting to control the number of decimal digits in floating-point text output. #99721 (phulv94).
• Added a setting output_format_always_write_decimal_point_in_float_and_decimal to always print a decimal point for floating-point and Decimal numbers in text formats, even when the value is a whole number. For example, output 1. instead of 1. Disabled by default. #62614 (Ilya Yatsishin).
• Add support for new HTTP handler configuration elements <url_prefix> and <full_url_prefix> (match all paths under a base path, useful e.g. for Prometheus handlers), as well as <url_regexp>, <full_url_regexp>, and <headers_regexp>. The old <url>regex:...</url> form is now obsolete. #107492 (Vitaly Baranov).
• Support basic statistics, a compact per-column statistic that stores numeric min/max, average string length, and NULL counts where applicable. #106048 (Han Fei).
• Support TTL nodes in ClickHouse Keeper, opt-in via the create_ttl feature flag. #100397 (Konstantin Vedernikov).
• Introduce a memory reservation feature for workloads. See the workload scheduling documentation. #82414 (Sergei Trifonov).

​

Experimental Feature

• Multi-stage distributed query execution: the planner splits the query plan into stages connected by scatter / broadcast / gather / shuffle exchanges and dispatches the plan fragments to worker nodes. The data between stages is streamed via TCP or passed via temporary files in shared object storage, the path supports distributed shuffle and broadcast hash joins, shuffle aggregation, and distributed sort. The feature is experimental and is disabled by default. #106020 (Alexander Gololobov). The experimental distributed query plan engine (make_distributed_plan) can now use a different task-dispatch and streaming-exchange port per worker, configured per replica in <remote_servers> with stateless_worker_port and streaming_exchange_port. When unset, the previous server-level ports (stateless_worker_client.port and distributed_query.streaming_exchange_port) are used. This makes it possible to run several workers on one host. #107885 (Alexander Gololobov).
• Added the experimental dphyp join reordering algorithm for inner joins as an option for the query_plan_optimize_join_order_algorithm setting, and the query_plan_optimize_join_order_max_searched_plans setting, which bounds the join-order search and falls back to the next algorithm in the chain when the bound is exceeded; set it to 0 to keep the previous unbounded search behavior. #98798 (Alexander Gololobov).
• Add aiEmbed function for using LLM APIs to generate embeddings from within ClickHouse. #102922 (George Larionov).
• Added an experimental Linux OOM canary: a sacrificial child process that the Linux OOM killer targets before the main server process. When enabled with oom_canary_enable, it sheds memory pressure by purging allocator caches, cancelling queries, cancelling merges, and recording an event in system.crash_log. The OOM response requires cgroup v2 memory.events OOM-kill evidence. #101942 (Peng).
• The web terminal interface at /webterminal is now a production feature and is enabled by default. It is controlled by the new enable_webterminal server setting; the former allow_experimental_webterminal setting is deprecated but still honored for backward compatibility. To disable the endpoint, set enable_webterminal to false. #106255 (Alexey Milovidov).
• PromQL HTTP endpoints now support specifying the database and table via URL query parameters (database=, table=, or table=db.table). The previous implicit default.prometheus table fallback is removed: if neither the handler config nor the URL specifies a table, the handler now returns The time series table name is not set. To keep the old behavior, configure <table>default.prometheus</table> or pass table=... in the URL. #107553 (Vitaly Baranov).
• Added support for the PromQL histogram_quantile function in prometheusQuery and prometheusQueryRange table functions. This enables computing quantiles over classic Prometheus histogram buckets identified by the le label. #103477 (Joe Smith).
• Added lazy posting list apply mode for the text index. When enabled via SET allow_experimental_text_index_lazy_apply = 1 and SET text_index_posting_list_apply_mode = 'lazy', posting lists are decoded on demand at packed-block granularity using a cursor-based approach instead of being fully materialized into Roaring Bitmaps, reducing memory usage and CPU time for selective text index queries. #100035 (Peng).
• Added support for the AssemblyScript ABI for WebAssembly UDFs. You can try the AssemblyScript ABI without installing the toolchain locally by using the browser-based playground, which compiles AssemblyScript code and loads the resulting WebAssembly module into local ClickHouse. #104606 (Vladimir Cherkasov).
• Allow for attaching Prometheus handlers on the main HTTP port with an optional prefix. #104975 (James Cunningham).
• Added an experimental MergeTree setting packed_skip_index_max_bytes that bundles small skip-index substreams into a single skp_idx.packed archive per part, reducing inode pressure when many skip indices are defined on a table. The decision is per substream at write time: substreams whose serialized size stays under the threshold go into the archive, anything larger keeps the standalone skp_idx_<name>.idx2 / .mrk2 layout. A single part can mix layouts. Full-text indices are not supported and are always per-file. Default is 0 (packing disabled). #105321 (Raúl Marín).
• Support Buffers serialization for WebAssembly UDFs using ABI BUFFERED_V1, and add webassembly_udf_enable_fuel as a persisted WASM UDF function setting. #105574 (Antonio Andelic).
• Add an experimental use_reader_executor setting (default off) that routes reads through a new pipeline-based ReaderExecutor instead of the legacy chain of read buffers. #106570 (Sema Checherinda). Added observability metrics for the experimental use_reader_executor read path, including a modeled read-cost KPI (ReaderExecutorModeledCostMsPerRequestedMiB). #106968 (Sema Checherinda). Introduced the ChainedBuffers buffer behind the experimental use_reader_executor read path, with a ReaderExecutorChainedBufferBytes metric. #107210 (Sema Checherinda).
• Descending sort order in MergeTree sorting keys (e.g. ORDER BY (time DESC, key)) is now always supported and no longer requires the experimental setting allow_experimental_reverse_key, which became obsolete. #106440 (Nikita Mikhaylov).
• Nullable(Tuple(...)) is now Beta. Disabled by default, set enable_nullable_tuple_type = 1 to use it. #107754 (Nihal Z. Miaji).

​

Performance Improvement

• Various changes to Keeper that make it around 2x faster overall (better batching, pipelining messages to leader, pipelining log appends). #101757 (Michael Kolupaev).
• Reduce per-query overhead for simple SELECT queries (parsing, analysis and planning). For example, SELECT count() FROM hits from a single connection is roughly 50% faster. #104513 (Raúl Marín).
• Optimize primary key index analysis for long and high-cardinality primary keys. For a long primary key, the run time of index analysis now mainly depends on the complexity of the query’s filter (the key columns it actually uses), not on the length of the primary key — so extending the sorting key has negligible extra overhead on index analysis for queries that filter on only a few of its columns. For a high-cardinality primary key, where ClickHouse keeps only a selective prefix of the key columns in memory and does not load the trailing ones, index analysis now works on just that in-memory prefix instead of the whole key. The optimization is enabled by default and can be turned off with the new setting use_lightweight_primary_key_index_analysis. Closes #65103. #91836 (Nihal Z. Miaji).
• The default x86 build now targets x86-64-v3 (AVX2) instead of x86-64-v2 (SSE4.2). This requires a CPU with AVX2 support (Intel Haswell or later, AMD Excavator or later). If your CPU does not support AVX2, use the amd64compat build which targets plain x86-64. #105019 (Raúl Marín).
• Use ipnsort and driftsort (stable sorting) — C++ ports of the Rust standard library’s sort implementations — for the general-purpose comparison sorts. Speeds up ORDER BY on non-numeric columns and stable sorts, and removes a worst case on reverse-sorted input. #106650 (Alexey Milovidov).
• New GROUP BY optimization for high cardinality evenly distributed keys that scatters rows across threads by hashing the grouping key, so each thread aggregates a disjoint subset of keys without a merge phase. Set enable_sharding_aggregator = 1 to enable it. #104233 (Nihal Z. Miaji).
• Enable hash table prefetching for string key GROUP BY queries, improving high-cardinality string aggregation performance by ~8%. #101007 (Le Zhang).
• Reduced peak memory usage when merging partial two-level aggregation results with large aggregate states (e.g. groupArray), by freeing each bucket’s source states incrementally during the merge instead of keeping them all alive until the merge completes. #102330 (Yuri Fedoseev).
• Memory representation of Enum DataType was optimized. Memory footprint of tables that have Enum type in it can improve up to 10x times for Enum component. Operation time to retrieve name by value(number) is similar or faster. Search by name is slightly slower, but this access direction is less common. #95668 (Ilya Yatsishin).
• Add identifier resolution caching to prevent duplicate identifier resolution during query analysis. #88043 (Max Justus Spransy).
• Improve performance of query analysis for queries over tables with many columns: avoid computing column node hashes (which include the whole source table expression) when not needed. Analyzing nested SELECT * subqueries over a table with ~1200 columns is now about 50 times faster. #106957 (Dmitry Novik).
• Sped up query analysis for queries with many or deeply nested function calls by removing a redundant query-tree hash from the function resolution cache. #107516 (Dmitry Novik).
• Fixed long login and query-startup stalls with replicated access storage when many access entities (row policies, roles, quotas, settings profiles) change at once. Each per-entity cache now recomputes once per notification batch instead of once per changed entity, removing quadratic work that could hold the access lock for minutes. #107672 (Azat Khuzhin).
• Share the hash join’s FixedHashMap as the join runtime filter on the probe side. When the build-side hash table is (or can be converted to) a FixedHashMap, it is published as the runtime filter and replaces the Set/BloomFilter that BuildRuntimeFilterStep would otherwise install. Controlled by the new setting join_runtime_filter_from_fixed_hash_table (default true). #105640 (Xiaozhe Yu).
• Lazy application of selector and replication indexes in case of JOIN followed by a selective LIMIT or TopN or another JOIN. To control the number of payload columns for enabling lazy selector indexes use the setting query_plan_min_columns_for_join_lazy_indexing (0 means the optimization is disabled). To control the LIMIT for which the optimization is applied use the setting query_plan_max_limit_for_join_lazy_indexing. #106566 (Hechem Selmi).
• DP (dynamic programming) JOIN reordering is now allowed with parallel replicas. #105889 (Nikita Taranov).
• Improved query plan when JOIN uses runtime filters (default-on enable_join_runtime_filters): the join-reorder cost model now sees through WindowTransform (and other row-preserving plan steps) on the right subtree and uses the underlying row count and per-column NDV instead of falling back to no statistics. #107229 (UnamedRus).
• De-virtualize the emission of non-joined rows in RIGHT and FULL join by batching them. #105679 (Hechem Selmi).
• Support packed keys32 and keys64 methods in HashJoin. #107202 (Nikita Taranov).
• Support packed keys32 and keys64 methods in Set. #107564 (Nikita Taranov).
• More optimal handling of nullable columns in aggregation. #106015 (Nikita Taranov).
• Improve performance of LIMIT BY queries. #103349 (Nihal Z. Miaji).
• Speed up ORDER BY ... LIMIT BY queries by running LIMIT BY inside each parallel sorted stream during Sort when LIMIT BY’s columns are a prefix of the ORDER BY. This reduces the number of rows flowing through the final sort merge and any downstream pipeline steps. This optimization is controlled by the new setting query_plan_push_limit_by_into_sort (enabled by default). #104000 (Nihal Z. Miaji).
• Speed up SELECT ... LIMIT N BY <cols> queries when <cols> are a prefix of the table’s sorting key, or become one after WHERE col = const fixes leading columns. With this enabled the MergeTree reads data in primary-key order and LIMIT BY first filters in streaming mode with O(1) memory per sorted stream which filters out most of the data, then finally running normal LIMIT BY on reduced data to get the final result. Controlled by the new setting optimize_limit_by_in_order (enabled by default). #105135 (Nihal Z. Miaji).
• Speed up LIMIT BY queries on partitioned MergeTree tables by running LIMIT BY inside each partition’s stream in parallel, instead of merging all streams into one before applying the limit. This applies when the partition expression is a deterministic function of the LIMIT BY columns, so no LIMIT BY group can span two partitions. Controlled by the new setting allow_limit_by_partitions_independently (enabled by default). #105126 (Nihal Z. Miaji).
• Speed up LIMIT BY queries by removing redundant key expressions: a key that is a deterministic function of the other keys is dropped (e.g. LIMIT 5 BY x, f(x) becomes LIMIT 5 BY x), and an injective function of a key is replaced by its argument (e.g. LIMIT 5 BY toString(x) becomes LIMIT 5 BY x). This evaluates fewer and cheaper expressions per row. Controlled by the new settings optimize_limit_by_function_keys and optimize_injective_functions_in_limit_by, both enabled by default. #106818 (Nihal Z. Miaji).
• Speed up sharded aggregation, grace_hash join bucketing, and parallel window partitioning by replacing per-chunk column hashing with a kernel that uses hardware CRC32C. #106538 (Harikrishnan Prabakaran).
• Optimize L1Distance, L2Distance, L2SquaredDistance, LinfDistance, and cosineDistance for array arguments with multi-target auto-vectorization on x86-64-v4/v3. For non-const array arguments, all of these except LpDistance, which uses std::pow with a runtime exponent, can now use compiler-vectorized kernels. The existing AVX-512 and Sapphire Rapids BFloat16 const-argument fast paths for L2Distance and cosineDistance are preserved. #101310 (Peng).
• Improved performance of the {Norm}Distance and dotProduct functions. #106007 (Nikita Taranov).
• Improved performance of the arrayNorm function. #106211 (Nikita Taranov).
• The performance of dotProduct improved in some cases. #106210 (Nikita Taranov).
• Improved performance of the AVX-512 const-left paths for L2Distance and cosineDistance on Array(Float32), Array(Float64), and Array(BFloat16) when the array length is exactly divisible by the vectorized processing width. #104625 (Sergey Kuznetsov).
• Enabled AVX-512 codegen for amd and better codegen for arrayDistance. #106505 (Nikita Taranov).
• Improved performance of the L2DistanceTransposed and cosineDistanceTransposed functions for the QBit data type. #106701 (Raufs Dunamalijevs).
• Improve performance of the encrypt, decrypt, tryDecrypt, aes_encrypt_mysql, and aes_decrypt_mysql functions by up to an order of magnitude, recovering the performance lost in the BoringSSL to OpenSSL 3.x migration (24.4). #107339 (Konstantin Bogdanov).
• Improve performance of the encrypt, decrypt, and halfMD5 functions by avoiding implicit per-row OpenSSL provider lookups in OpenSSL 3.x. #99105 (Konstantin Bogdanov).
• Add SIMD implementation of SHA1 that hashes multiple inputs in parallel using AVX-512, improving throughput. #105459 (Joanna Hulboj).
• Added an AArch64 ASIMD/NEON backend for the multi-buffer MD5 implementation. #105563 (Venkata Vineel ).
• Squash source blocks before calculating projection during MATERIALIZE PROJECTION to reduce the number of temporary projection parts and merge overhead. ~3.4x speedup on a 50M-row table. #100047 (Amos Bird).
• Reduced INSERT latency for MergeTree tables with skip indices on ORDER BY columns by avoiding redundant column permutations during part writing (~17% on traces-shaped workloads). #101101 (Amos Bird).
• Enabled Arrow’s SIMD code paths (NEON on ARM, SSE4.2/AVX2/AVX512 on x86) and used them to speed up decoding of BYTE_STREAM_SPLIT-encoded Parquet FLOAT/DOUBLE columns. #106376 (Raúl Marín).
• Improved insertion performance for LowCardinality columns with bloom_filter indexes. #106410 (Michael Jarrett).
• Parallelize blob copies in DiskObjectStorageTransaction::copyFile using a per-disk thread pool. #105089 (Asya Shneerson).
• Avoid reading file contents when using the One input format with file-like table functions such as file and s3. #105157 (Yue Ni).
• Speed up clickhouse startup on macOS by about 3x by not exporting dynamic symbols, which the dynamic linker would otherwise process on every launch. #107768 (Raúl Marín).
• Speed up filesystem cache loading on startup by avoiding a redundant directory open per cache key. #107414 (Alexey Milovidov).
• Speed up filesystem cache loading on startup by letting directory-listing threads help load cache metadata once listing is finished, instead of leaving half the threads idle. #107416 (Alexey Milovidov).
• Reduced clickhouse process startup overhead by updating the bundled WasmEdge library, which no longer generates a random hash secret in a global initializer at every startup. #107869 (Raúl Marín).
• The MergeTree primary key and skip indexes can now prune granules for filters where ifNull or coalesce wraps a condition, such as ifNull(key = 0, 0) or coalesce(key = 0, 0). Such predicates — often emitted by query generators to turn a possibly-NULL comparison into a definite boolean — were previously opaque to index analysis and could not skip granules. This extends the existing allow_key_condition_coalesce_rewrite setting (enabled by default). Closes #106264. #106272 (Andy Zhao).
• Skip index evaluation on DateTime64 types should consume lesser CPU in 26.6. #107707 (Shankar Iyer).
• Improved performance of text index analysis in multi-token searches by optimizing handling of rare tokens. #98226 (Anton Popov).
• Use a cache for the posting-list compressed segments of the text index. #106299 (Anton Popov).
• Improved the performance of generic exclusion search for queries that select large consecutive ranges of the table. #93813 (Michael Jarrett).
• S3 clients with the same endpoint and bucket share a cache, avoiding duplicate region discovery. Closes #92482. #96802 (Andrey Zvonov).
• If a query contains a has(<constant array>, <expr>), the analyzer will transform the node to use a faster in implementation if it is feasible to do so. #97341 (Shankar Iyer).
• When a new job is scheduled in ThreadPool, wake the most recently idle worker thread (LIFO) instead of an arbitrary one. This reduces memory fragmentation from per-thread allocator caches. The order in which jobs are processed is not changed. #100177 (Alexey Milovidov).
• Improved performance of approximate runtime filters and bloom filter indexes. #100201 (Christoph Viebig).
• Allow ASOF JOIN to use the parallel_hash join algorithm, parallelizing the build across distinct equality-key values. Previously ASOF was unconditionally opted out of parallel_hash. #105375 (Greg Maher).
• Avoid reading unbounded inputs for INTERSECT and EXCEPT queries when an empty input makes the result known to be empty. #105393 (Yue Ni).
• Add constinit for various MemoryTracker blockers. #105490 (Azat Khuzhin).
• Improve performance of bitmapContains for non-UInt64 groupBitmap states by avoiding repeated rb_max calls during range checks. #105960 (Yue Ni).
• Faster discontinuous queries for LowCardinality columns backed by a single dictionary. #103662 (Ivan Babrou).
• Added a fast path for the HashJoin probe when the block has a single row. #106008 (Nikita Taranov).
• Predicates from queries like SELECT ... FROM view(SELECT ... FROM remote(...)) WHERE ... are now pushed into the query sent to the remote shard. The optimization allow_push_predicate_ast_for_distributed_subqueries lacked a case for table functions. #105986 (Nikolai Kochetov).
• Turn on enable_join_transitive_predicates by default. #103724 (Alexander Gololobov).
• Added support for functions multiSearchAny, multiSearchAnyUTF8, and multiMatchAny in text indexes. Also improved text index analysis for the function match: now patterns with alternative groups can skip more granules. #106279 (Anton Popov).
• Allow filters introduced after the initial PREWHERE selection (predicate pushdown, runtime filters, or explicit PREWHERE plus a WHERE set by the planner) to be merged into the existing PREWHERE on a second optimizer pass instead of staying as a separate Filter step above the MergeTree read. #105445 (Yarik Briukhovetskyi).
• The QueryConditionCache now records individually filtered-out granules even within read batches that partially pass PREWHERE, reducing the number of marks re-read by subsequent queries with the same condition. #105335 (Han Fei).
• Added Keeper setting nuraft_use_bg_thread_for_snapshot_io, enabled by default, to make NuRaft read snapshot objects from a background thread instead of Raft worker threads. #106285 (Antonio Andelic).
• Reduce peak memory usage when applying received snapshots in ClickHouse Keeper with KeeperMemoryStorage. #105851 (Antonio Andelic).
• Route LLVM/JIT allocations to a dedicated jemalloc arena. Reduces RSS fragmentation caused by interleaving long-lived JIT bookkeeping with short-lived query allocations. Adds new asynchronous metrics jemalloc.jit_arena.active_bytes, jemalloc.jit_arena.dirty_bytes, CompiledExpressionCacheBytesMax, and CompiledExpressionCacheCountMax for observability. SYSTEM DROP COMPILED EXPRESSION CACHE now also drops the underlying CHJIT instances and purges the JIT arena. #104113 (Raúl Marín).
• Fix excessive memory reservation when reading sparse dictionary-encoded Nullable(String) columns from Parquet with the native V3 reader. #102805 (Francisco).
• Fixed CREATE TABLE ... CLONE AS (and REPLACE / ATTACH PARTITION ... FROM) on encrypted disks copying data instead of hardlinking it. #106731 (Nikita Mikhaylov).
• Improve cardinality estimation in the query plan optimizer: a column produced by a deterministic single-argument function (e.g. toYear(date)) now inherits its argument’s number of distinct values as an upper bound instead of being left without statistics, leading to more accurate join reordering. #107757 (Alexander Gololobov).
• Reduced CPU overhead of asynchronous logging under high log rates by notifying the log consumer only on the empty-to-non-empty queue transition instead of on every message. #107352 (Nikita Mikhaylov).
• Remove duplicate calculations during query execution in more scenarios. #106113 (Yarik Briukhovetskyi).
• Parallelize the processing of a recursive CTE’s result: a GROUP BY or other operation over a large WITH RECURSIVE result is no longer limited to a single thread. #107694 (Alexey Milovidov).
• Improved the performance of case-insensitive substring search (positionCaseInsensitiveUTF8, ILIKE, multiSearchAnyCaseInsensitiveUTF8, and related functions) by adding a NEON SIMD path on ARM and a wider AVX2 kernel on x86. #107882 (Raúl Marín).
• Fixed a ~16% throughput regression of the FPC floating-point codec on ARM that was introduced when its predictor tables started using VectorWithMemoryTracking. #108182 (Groene AI).

​

Improvement

• The background MemoryWorker now periodically updates the server’s hard memory limit based on the current memory usage and the amount of memory the kernel reports as available, so ClickHouse leaves room for other processes running on the same host. The formula is (resident memory + system MemAvailable) * max_server_memory_usage_to_ram_ratio. The same max_server_memory_usage_to_ram_ratio server setting controls both the startup cap and the dynamic adjustment; set it to 0 to disable both. To keep only the static startup/reload cap (the behavior of previous versions), set the new server setting memory_worker_dynamic_hard_limit to 0. #104964 (Alexey Milovidov).
• Support C# PostgreSQL client in PostgreSQL protocol. This closes #18611. #80785 (Konstantin Vedernikov).
• Mutations now use the analyzer. #98884 (Nikolai Kochetov).
• Added granular control over query echoing in the CLI. The --echo option now accepts an optional boolean value and works in both interactive and batch mode. New options --echo-formatted and --echo-query-id control whether echoed queries are formatted and whether the query_id is printed, respectively. The --hilite/--highlight option now also controls highlighting of echoed queries. As a side effect, --echo is now a boolean-value option, so a positional query directly after a bare --echo is treated as its value; use --echo --query "..." or --echo=false instead. #106191 (Alexey Milovidov).
• Support schema evolution together with concurrent consistency checks for data lake catalogs. #106102 (Konstantin Vedernikov).
• Allow cache disk for datalake table engines. #102017 (RinChanNOW).
• Fix DataLakeCatalog databases with catalog_type = 'onelake' reading table data by using the OneLake Blob endpoint (.blob.fabric.microsoft.com) by default. Set onelake_use_blob_endpoint = false to keep the previous DFS endpoint (.dfs.fabric.microsoft.com) behavior. #106843 (Konstantin Vedernikov).
• Update default of S3(Azure)Queue setting persistent_processing_node_ttl_seconds from 1 hour to 6 hours (which stands for cleanup time due to service unclean restart etc). The previous default was fine for processing nodes ttl, but it is also used for bucket lock ttl, which can be hold for longer time, therefore 1 hour was not enough. #106838 (Kseniia Sumarokova).
• Decreased the default size of the Iceberg metadata cache from 1 GB to 128 MB. #106492 (Konstantin Vedernikov).
• Propagate the S3Queue streaming task query_id to dependent table inserts. #106494 (Christoph Wurm).
• Populate used_storages in system.query_log with the storage engine name when querying tables via DataLakeCatalog. #100706 (Melvyn Peignon).
• You can check which ‘operation’ and what ‘summary’ Iceberg’s snapshots have. Useful for testing and debugging. #106246 (Den Kalantaevskii).
• Complete the final S3 multipart upload request asynchronously via the task tracker, so it can overlap with the last part upload instead of running serially in finalize. #105487 (Asya Shneerson).
• Added four new MergeTree table settings to control default parameters of text indexes: text_index_dictionary_block_size, text_index_dictionary_block_frontcoding_compression, text_index_posting_list_block_size, and text_index_posting_list_codec. These settings allow tuning text index behavior at the table level without specifying parameters in every index definition. Explicit per-index arguments still take precedence. #100626 (Anton Popov).
• Allow projections to override additional MergeTree settings (compression, part format, serialization) beyond just index_granularity, with an allow-list of allowed settings and validation. #101170 (Amos Bird).
• Add current_projection, current_projection_progress, projections_completed, and projections_remaining columns to system.merges to expose projection merge progress. #102611 (Amos Bird).
• Fixed PREWHERE with IN subquery on primary key columns not using primary key index for granule pruning, causing full table scans instead of reading only relevant granules. #102570 (Nikita Mikhaylov).
• Function h3PolygonToCells now enforces the maximum array size across all polygons of a MultiPolygon, validates the underlying H3 library return codes, and rejects MultiLineString arguments instead of silently returning an empty result. #106399 (Raúl Marín).
• Add a new keeper-only system.keeper_snapshots table with information about local ClickHouse Keeper snapshots. #105571 (Miсhael Stetsyuk).
• Add a new keeper-only system.keeper_changelogs table with information about local ClickHouse Keeper changelog (Raft log) files. #105617 (Miсhael Stetsyuk).
• Add a keeper-only system.keeper_cluster table. Contains one row per Raft cluster member as seen by the current Keeper. #105646 (Miсhael Stetsyuk).
• Add more Keeper profile (server-side + client-side) events for watches. Closes #97703. #105336 (Konstantin Vedernikov).
• Added Keeper coordination_settings for NuRaft uncommitted log entry admission limiting and append-entries backward-probe throttling. #106108 (Antonio Andelic).
• Add enable_compression setting for the mysql table function, the MySQL table engine, the MySQL database engine, dictionary SOURCE(MYSQL), and named collections. When enabled, ClickHouse negotiates MySQL protocol-level compression for all data transferred over the connection. #103229 (Bernard Lim).
• Reduced cancellation latency for queries running over the PostgreSQL wire protocol: KILL QUERY now interrupts output serialization within a single chunk instead of waiting for the entire chunk to be sent to the client. #106535 (Roman Vasin).
• Reduced cancellation latency for queries running over the MySQL wire protocol: KILL QUERY now interrupts output serialization within a single chunk instead of waiting for the entire chunk to be sent to the client. #107228 (Roman Vasin).
• Fixed reading of the auth_response length in the MySQL handshake, where a length byte >= 128 was interpreted as a multi-gigabyte value because it was read as a signed char. #107384 (uwezkhan).
• MaterializedPostgreSQL now maps PostgreSQL numeric(p, 0) columns with precision greater than 76 (e.g. numeric(78, 0) used for 256-bit integers) to ClickHouse Int256 instead of failing with “Precision too big”. Values that do not fit into Int256 are rejected with a clear error. #107431 (Alexey Milovidov).
• Send an exception to the client when TCP connection setup fails. #107317 (Nikolai Kochetov).
• Users can now insert Avro Fixed fields for 8-bit/16-bit/32-bit/64-bit integer variants. #98139 (Patrick Pichler).
• The AvroConfluent format now retries the Confluent Schema Registry HTTP client on transient failures (transport timeouts, connection refused, DNS errors, HTTP 5xx/408/429) with exponential backoff, instead of aborting the INSERT on the first network glitch. New settings format_avro_schema_registry_max_retries (default 5) and format_avro_schema_registry_retry_initial_backoff_ms (default 100) control the policy. Schema-validation errors (HTTP 409, malformed Avro JSON) remain fatal. #106661 (Groene AI).
• Record privileges in system.query_log.used_privileges for all granted access checks, including those that go through the non-throwing isGranted path (checkAccessWithFilter). Previously only privileges checked via throwing entry points (checkAccess / checkGrantOption) were recorded, which made READ ON FILE / READ ON S3 / READ ON AZURE / READ ON URL invisible in the audit log for DESCRIBE, CREATE TABLE AS, and similar queries that use file / s3 / azure / url table functions. Access enforcement is unchanged. #104693 (Alexey Bakharew).
• Added asynchronous metrics TotalUncompressedBytesOfMergeTreeTables and TotalUncompressedBytesOfMergeTreeTablesSystem, reporting the total uncompressed size of data stored in MergeTree-family tables. #106364 (Alexey Milovidov).
• Adds the GlobalMemoryLimitExceeded profile event so operators can monitor when the server-wide memory limit is hit. #106466 (Sacheendra Talluri).
• Populate the ExecutableUserDefinedFunction* ProfileEvents counters for executable_pool UDFs. Per-call invocation count, wall and pool-wait time, child-process CPU and peak memory, and stdin/stdout bytes are now visible in system.query_log.ProfileEvents. #105010 (Xu Jia) #105618 (Ilya Andreev).
• Added asynchronous metrics ExecutableUserDefinedFunctionMemoryResidentBytes and ExecutableUserDefinedFunctionProcesses, reporting the resident memory (VmRSS) and number of live processes of executable and executable_pool user-defined functions, including descendant processes and idle pool workers. #107300 (Hanzi Jiang).
• Added three Linux-only async metrics — MemoryThreadStacksCount, MemoryThreadStacksVirtual, and MemoryThreadStacksResident — that report pthread stack memory separately from the rest of anonymous memory. Enabled via asynchronous_metrics_enable_heavy_metrics. Requires Linux 5.17+ (prctl(PR_SET_VMA_ANON_NAME)) and readable /proc/self/smaps; otherwise the metrics are absent and the limitation is reported in system.warnings. #106230 (Raúl Marín).
• Expose the MemoryThreadStacks* asynchronous metrics (resident/virtual size and count of thread stacks) on macOS. #107752 (Raúl Marín).
• Table engines now carry embedded documentation, introspectable via the new description, syntax, examples, introduced_in, and related columns of the system.table_engines table. #106177 (Alexey Milovidov).
• Database engines now carry embedded documentation, introspectable via the new description, syntax, examples, introduced_in, and related columns of the system.database_engines table. #106178 (Alexey Milovidov).
• Data types now carry embedded documentation, introspectable via the new description, syntax, examples, introduced_in, and related columns of the system.data_type_families table. #106180 (Alexey Milovidov).
• Input/output formats now carry embedded documentation, introspectable via the new description, examples, introduced_in, and related columns of the system.formats table. #106181 (Alexey Milovidov).
• Aggregate function combinators now carry embedded documentation, introspectable via the new description, syntax, examples, introduced_in, and related columns of the system.aggregate_function_combinators table. #106185 (Alexey Milovidov).
• Added the --chime command-line option to clickhouse-client, clickhouse-local, and the embedded client. When a query finishes (on success or on error) after running for at least N seconds, the client writes the ASCII BEL control character (\x07) to stderr. Terminals decide whether to make a sound or a visual flash based on the user’s preferences. Enabled by default with a 5-second threshold; pass --chime N for a custom threshold or --chime 0 to disable. Closes #92718. #104545 (Groene AI).
• In clickhouse-client, when the terminal does not support bracketed paste, embedded newlines from a pasted multi-line query are now folded into the same edit buffer instead of switching to the continuation prompt for every line. The pasted query stays under a single prompt and the arrow keys navigate across its lines. This is a best-effort improvement that relies on the typical TTY-buffered paste behavior. #104299 (Alexey Milovidov). Now also works with syntax highlighting disabled (--highlight 0). #106665 (Alexey Milovidov).
• Add MySQL-style runtime pager control for clickhouse-client/clickhouse-local. #105706 (Azat Khuzhin).
• clickhouse-client now respects non-UTF-8 terminals: when the terminal character encoding is not UTF-8 (e.g. with LANG=C), the Pretty formats fall back to ASCII grid borders instead of printing Unicode box-drawing characters that would corrupt the terminal. #106213 (Alexey Milovidov).
• Sanitize server-supplied display strings in clickhouse-client’s Hello packet handling (server name, time zone, display name, password-rule patterns and messages) so a hostile server cannot inject bytes that the client renders verbatim, and cap their size to defend against unbounded allocation. #105243 (Raúl Marín).
• The Web UI now displays the result as an image when a query uses an image output format such as FORMAT PNG, instead of showing the raw bytes as a table. #107638 (Alexey Milovidov).
• Add setting output_format_pretty_use_nbsp_for_padding to render table-layout padding in table-style Pretty formats as U+00A0 NO-BREAK SPACE when output_format_pretty_grid_charset is UTF-8. This helps copied Pretty output keep table alignment in tools that collapse regular spaces. The setting is disabled by default, and ASCII charset output keeps regular spaces. #103559 (Ashrith Bandla).
• The universal install script now also installs clickhousectl into ~/.local/bin on Linux and macOS, with a chctl symlink. Set CLICKHOUSE_ONLY=1 to skip. #105399 (Alasdair Brown).
• Fixed a spurious Syntax error: Unterminated quoted string that could occur intermittently when installing ClickHouse via curl https://clickhouse.com/ | sh while a clickhouse binary already existed in the current directory. #106417 (Alexey Milovidov).
• Fix the left panel height on the play.html page when the web terminal is open, so the terminal toggle button at the bottom stays visible. #105087 (Alexey Milovidov).
• Avoid selecting the only query in the Web UI after running it. #105711 (Joey Yu).
• The Web SQL UI now expands the query editor when a query is at most 30% taller than the current editor height. #105713 (Joey Yu).
• In the Web UI, lay out the bottom menu icons (Web Terminal and GitHub) horizontally when the left panel is expanded, giving the list of databases and tables more vertical space. #106427 (Alexey Milovidov).
• In the Web UI (play.html), pressing Escape now deactivates the currently selected result-table cell. #107777 (Alexey Milovidov).
• Improvements to the database panel in the Web SQL UI (Play): a tooltip on the database icon, a button to hide the panel, full-width clickable database/table names, and a fix for the panel width shifting when the scrollbar appears. #108013 (Alexey Milovidov).
• Add a server setting (min_allocation_size_to_throw_on_memory_limit) to allow any allocation to throw MEMORY_LIMIT_EXCEEDED, this should help with preventing OOMs. #105265 (Azat Khuzhin).
• Allow NONE authentication for SSH (prevents asking password for no_password users). #105476 (Azat Khuzhin).
• Reset the supplementary group list in clickhouse su before dropping privileges, matching the behavior of /bin/su and sudo. Previously the dropped process inherited the invoker’s supplementary groups. #105247 (Raúl Marín).
• Add BFloat16 support for numeric predicate (isFinite, isNaN) functions. #105391 (Mohamed Hussain S).
• Reduce memory usage of polygon dictionaries (polygon_index_cell, polygon_index_each) and make system.dictionaries.bytes_allocated correctly reflect the lookup index they build. #105431 (Raúl Marín).
• Fixes Kafka table engine consumers that kept using a short poll interval after partition assignment, so they return to the configured kafka_poll_timeout_ms and avoid excessive empty polls, smaller inserted parts, and extra merge overhead after rebalances. #100431 (sugaf1204).
• Identify columns by position (instead of by name) when removing unused columns in the query plan. This enables unused-column removal when duplicated column names are present. #100586 (János Benjamin Antal).
• ClickHouse throws a proper exception for attempts to use table_readonly for replicated engines, with better exception handling for DDL queries (and DatabaseReplicated) related to this setting. #100950 (alesapin). OPTIMIZE TABLE ... ON CLUSTER and other DDL no longer hang when the target table has table_readonly = 1. The setting now throws a new dedicated error code TABLE_IS_PERMANENTLY_READ_ONLY, which DDLWorker treats as non-retriable (distinct from the transient TABLE_IS_READ_ONLY that arises during temporary ReplicatedMergeTree ZooKeeper disconnects). The table_readonly setting is also now explicitly rejected for ReplicatedMergeTree, both at creation and via ALTER MODIFY SETTING. Follow-up to #97652. #105109 (Alexey Milovidov).
• Fixed refreshable materialized view refreshes that could fail with UNKNOWN_TABLE after the target table had been replaced (for example via EXCHANGE TABLES or drop-and-recreate): the refresh now resolves the target table by name instead of relying on a stale UUID. #102724 (Seva Potapov).
• Added a setting wait_for_part_commit_in_dependent_materialized_views (default false). When a materialized view’s SELECT reads back from its own source table (for example via an INNER JOIN to the source), the cascade could miss the row currently being inserted, because the inserted part is committed only after the dependent views run. Enabling this setting commits the part before the dependent views are triggered, so they observe the in-flight row, at the cost of less insert parallelism. #105943 (Elmi Ahmadov).
• Do not swallow fatal errors (e.g. MEMORY_LIMIT_EXCEEDED) in tryDeserialize. #106808 (Azat Khuzhin).
• Check CREATE TEMPORARY TABLE grant lazily in InverseDictionaryLookupPass. #107098 (Azat Khuzhin).
• Fix an ILLEGAL_TYPE_OF_ARGUMENT error for distributed queries with serialize_query_plan = 1 that contain a lambda with a constant argument (e.g. arrayMap(t -> t.2, ...)). Constant columns of ActionsDAG INPUT nodes are now preserved during query plan serialization. #107124 (Alexey Milovidov).
• Deserialization of the states of the contingency, cramersV, cramersVBiasCorrected, and theilsU aggregate functions now validates that the stored counts form a consistent contingency table and throws a CORRUPTED_DATA exception otherwise. Closes #106899. #107185 (Nihal Z. Miaji).
• Preserve original error codes for exceptions wrapped in arrow::Status. #107267 (Azat Khuzhin).
• Support keyed_by_normalized_query_hash for quotas defined in the static server configuration (users.xml), matching the existing CREATE QUOTA ... KEYED BY normalized_query_hash DDL syntax. #107654 (Alexey Milovidov).
• The compatibility setting no longer applies obsolete settings, so it does not mark them as changed or produce obsolete-setting warnings. #107737 (Maxim Orlovsky).
• Async inserts no longer log the full list of query_ids at trace/debug level, which since 26.2 could blow up text_log on services with heavy async insert traffic. Detailed lines are now at test level. #107852 (Sema Checherinda).
• Filesystem cache improvement: stop invalidated priority entries from pinning KeyMetadata. Follow up to #106387. #107903 (Kseniia Sumarokova).
• FileCache metrics improvement. #106116 (Kseniia Sumarokova).
• Do not print the “Stack trace (when copying this message, always include the lines below):” preamble in exception messages when the stack trace is actually empty. #106524 (Alexey Milovidov).
• Fix a potential overflow in roundUpToMultiple when computing filesystem cache boundary alignment for very large offsets. #92579 (Bharat Nallan).
• EXPLAIN SYNTAX now formats operators as function calls consistently in explain output (for example plus(1, 2) instead of 1 + 2). #94681 (Mohamed Abdelhalim ).
• Improved table name hints in error messages: no longer suggests the exact same name, and includes the database name in the suggestion (e.g., “Maybe you meant other_db.my_table?”). #95116 (Mathuranath Metivier).
• Improved error message for unresolved identifiers in queries without a FROM clause to suggest adding one. #101769 (Yash ).
• Add EXPLAIN PIPELINE option to compact repeated processor chains. #104662 (Yue Ni).
• Fix a heap-buffer-overflow read in syntax-error message construction (UTF8::computeWidthImpl via parseQuery.cpp) that occurred when the parser had backtracked past the first semicolon / end-of-stream token; release builds were splicing bytes from neighboring heap memory into the displayed error, ASan builds aborted. #105086 (Groene AI).
• Make hive partition strategy a default under compatibility setting file_like_engine_default_partition_strategy. #86746 (Kseniia Sumarokova).
• Use a larger (8 MiB) thread stack on macOS to match the Linux default, fixing rare server crashes from stack overflow during JIT compilation. #107033 (Raúl Marín).
• The temporary storage size limit of clickhouse-local is no longer hard-coded to 1 GiB. The default is raised to 1 TiB and can be configured with the max_temporary_data_on_disk_size server setting. #106689 (Alexey Milovidov).
• Add read-checksums command to clickhouse-disks. It reads the checksums.txt file of a MergeTreeDataPart and prints it as a human-readable, tab-separated table. #106901 (Asya Shneerson).
• Added a read-bitmap command to clickhouse-disks. #107834 (murphy-4o).
• Chdig: enable TLS for secure connections (previous leads to either “unknown setting skip_verify” or “connection reset by peer”). #105864 (Azat Khuzhin).
• chdig v26.5.1. #105918 (Azat Khuzhin).
• chdig v26.6.1 - {asynchronous_,}metric_log, query patterns, heatmaps, logs/traces stored on disk (previously it can use >40GiB for 1 hour of traces). #107678 (Azat Khuzhin).
• The server config (programs/server/config.xml) now sets keep_alive_timeout to 30 (formerly 10), aligning with the actual documented and implemented default. #107779 (Dan Checkoway).
• The default for the server setting disk_connections_rcvbuf changed from 0 (kernel TCP autotuning) to 204800 (200 KB), capping the per-socket TCP receive buffer for object-storage (S3/Azure/GCS) disk connections. #107859 (Sema Checherinda).
• The PREWHERE optimizer now groups conjuncts that reference the same column set before estimating selectivity, so that conditions like a > 2500 AND a < 2502 are evaluated together as a combined range rather than as two independent predicates, producing a more accurate selectivity estimate. #106337 (Han Fei).
• Add silk fiber aware secure and plain socket implementations. #107680 (Miсhael Stetsyuk).
• Fixed a garbled error message (swapped arguments) and a typo in the nested function. #108031 (Alexey Milovidov).
• Added a sed command to clickhouse-disks that applies a sed expression to a given path, in-place. #107131 (Asya Shneerson).
• Added the create-time materialized_postgresql_use_extended_date_and_time_types setting for the MaterializedPostgreSQL database engine. By default (enabled), PostgreSQL date/timestamp columns are inferred as Date32/DateTime64; setting it to 0 at CREATE DATABASE time infers the narrower Date/DateTime types. #107428 (Alexey Milovidov).

​

Bug Fix (user-visible misbehavior in an official stable release)

• Make possible to drop detached parts with the tryN suffix. #58957 (János Benjamin Antal).
• Fix a MULTIPLE_EXPRESSIONS_FOR_ALIAS exception for queries with duplicate projection aliases (for example SELECT *, day + 365 AS day) inside nested subqueries when running with parallel replicas. #80310 (Alexey Milovidov).
• Fix a bug with splitMultipartQuery throwing an “Empty query” error for queries that end with comment after semicolon. #85491 (Yarik Briukhovetskyi).
• Fix an exception with ARRAY JOINs: Function writeSlice expects same column types for GenericArraySlice and GenericArraySink which happens when LowCardinality numeric types are used (issue #57243). #91784 (Jimmy Aguilar Mena).
• Fix NOT_FOUND_COLUMN_IN_BLOCK exception when using LIMIT BY with constant columns alongside DISTINCT and ORDER BY with the new analyzer. #93195 (Ashrith Bandla).
• Fix NOT NULL columns being silently created as Nullable when data_type_default_nullable = 1 and the table is created in a Replicated database or via ON CLUSTER. #97572 (xiaohuanlin).
• Fixed HiveCatalog connection stability by adding automatic retry mechanism and reconnection logic for handling TTransportException errors when communicating with Hive Metastore. #98471 (Dmitriy Borisenko).
• Fixed nested function (used internally by ARRAY JOIN) stripping LowCardinality from column types, causing Array(LowCardinality(String)) to become Array(String) in output. Closes #95582. #98974 (Yash ).
• Fix “Distributed task iterator is not initialized” exception when using url, s3, or similar table functions in queries with parallel replicas enabled. #100146 (Alexey Milovidov).
• Fix logical error exception when reading Iceberg tables whose format version was upgraded by an external tool (e.g. Spark). #100407 (Alexey Milovidov).
• Functions like, ilike, notLike, notILike, and match now support constant haystack with non-constant needle (e.g. 'foo' LIKE pattern_column), which previously threw ILLEGAL_COLUMN. #100479 (Yash ).
• Fixed wrong row count returned by a MaterializedView query with query_plan_enable_optimizations = 0 when the view maps an integer column to a Bool column. #100692 (Maksim Moisiuk).
• Fix an exception (LOGICAL_ERROR: 'PREWHERE passed to format that doesn't support it') when reading Iceberg tables containing ORC data files with PREWHERE optimization enabled. #101206 (Groene AI).
• Fix race in RestCatalog. #101216 (Smita Kulkarni).
• Fix NOT_FOUND_COLUMN_IN_BLOCK error when selecting from a VIEW over a table with a normal projection. #101218 (Amos Bird).
• Fix stale metadata in FileLog when a file is deleted and recreated. #101408 (Azat Khuzhin).
• Fix join reordering silently dropping unmatched rows of a RIGHT/LEFT JOIN when it is comma-joined (cross) with another table, e.g. t1 RIGHT JOIN t2 ON t1.c = t2.c, t3. The query previously returned the inner-join result instead of the outer-join one. #101684 (Groene AI).
• Fixed FORMAT clause being consumed by INSERT instead of applying to EXPLAIN output in EXPLAIN INSERT INTO ... SELECT ... FORMAT .... #101772 (Yash ).
• Fixed incorrect compression codec selection for MergeTree parts when the table-level default_compression_codec setting was explicitly configured. Parts written on insert, during merges, and for projections used the server-wide default codec instead of the table-level setting (the empty part produced by a fully-deleting mutation is now covered as well). #101784 (Yash ).
• Reject negative Float64 values (e.g. -100.5) in workload settings like max_bytes_per_second, max_cpus, etc. Previously only negative integers were validated, allowing negative floats to silently create broken scheduler nodes. Closes #101825. #101842 (Groene AI).
• Fix a LOGICAL_ERROR (“Port is not connected”, code 49) that could occur when executing queries involving a VIEW with aggregation inside a JOIN. #102574 (Jimmy Aguilar Mena).
• Fix inconsistent part metadata after mutations of columns with non-default serializations. #102817 (Eduard Karacharov).
• Fix updating metadata before executing mutation resolves #96806. #102882 (Smita Kulkarni).
• Fix SELECT queries being significantly slower when concurrent INSERTs are running. Previously an INSERT pipeline reserved CPU slots up to max_threads at query start even when most slots were never used, starving concurrent SELECTs. CPU slot allocation is now demand-driven: the pipeline only requests slots as it actually pushes parallelizable work. Applies both to concurrency control and to the preemptive CPU scheduler for workloads. New server setting concurrent_threads_lazy_allocation (default true) acts as a rollback lever. #102928 (Seva Potapov).
• Fix NULL propagation when reading subcolumns extracted from Nullable(Tuple(...)) columns. For example, for tup Nullable(Tuple(s Nullable(String))), SELECT tup.s now correctly returns NULL in rows where the outer tuple is NULL instead of garbage values. This covers all element types that can represent NULL: Nullable, Dynamic, Variant and LowCardinality(Nullable(...)). Closes #105356. #102942 (Nihal Z. Miaji).
• Throw INCORRECT_DATA instead of LOGICAL_ERROR when client-supplied data in the Native format is truncated or malformed. #102975 (János Benjamin Antal).
• Added cancellation check to ObjectStorage read operation. Closes #98165. #103016 (Smita Kulkarni).
• Respect input_format_binary_max_type_complexity in decoding aggregate function parameters. Closes #102903. #103026 (Pavel Kruglov).
• Fix nullable inference for geo columns in Arrow-based Parquet reader. Closes #101845. #103032 (Pavel Kruglov).
• Fixed recursiveRemoveLowCardinality erasing custom geometry type names (e.g. LineString vs Ring, MultiLineString vs Polygon), which caused misinterpretation of the geometry type. Closes #103207. #103041 (Joanna Hulboj).
• Fixed input_format_max_block_size_bytes being silently ignored during INSERT parsing when max_insert_block_size_bytes is 0 (the default). The setting now correctly limits the size of blocks produced by row input formats. #103068 (Kirill Kopnev).
• Fixed ClickHouse occasionally producing invalid GSSAPI tokens due to incorrect stripping of trailing null bytes. #103114 (Michael Jarrett).
• Fix analyzer-time constant folding for short-circuit functions (if, multiIf, and, or, etc.) so that statically unreachable branches no longer raise exceptions at analysis time. For example, WITH 0 AS n SELECT multiIf(n = 0, 0, intDiv(100, n)) now correctly returns 0 instead of failing with a division-by-zero error. #103157 (Peng).
• EXPLAIN SYNTAX expands parameterized views. #103263 (Jordi Villar).
• Fix data corruption when writing Parquet (and other trailer-bearing formats such as ORC and Arrow) to HDFS via INSERT INTO FUNCTION hdfs(...). Since 26.1, WriteBufferFromHDFS did not flush its working buffer on finalize(), so the last up to DBMS_DEFAULT_BUFFER_SIZE bytes of every file were silently lost, including the Parquet PAR1 footer. Reading such files returned Not a Parquet file (wrong magic bytes at the end of file). #103268 (Groene AI).
• Fix wrong results and a possible logical error for correlated subqueries when a join size limit (max_rows_in_join / max_bytes_in_join) is set together with join_overflow_mode = 'break'. The join created internally to evaluate a correlated subquery now ignores those user limits, so it can no longer stop early and drop rows. #103322 (Groene AI).
• Fix a bug where ALTER TABLE ... MODIFY SETTING on an EmbeddedRocksDB table could persist an invalid setting value to the table metadata file even when the server rejected the query. On the next server restart the table would fail to attach with CANNOT_PARSE_BOOL (or a similar parsing error), and in databases where load failures are fatal the server would refuse to start. Invalid setting values are now rejected before any metadata is written. #103417 (Groene AI).
• Fixed a server abort when creating a table with an object-storage engine (AzureBlobStorage, DeltaLakeAzure, S3, HDFS) using an unsupported number of arguments. The server now returns a clean NUMBER_OF_ARGUMENTS_DOESNT_MATCH error instead of aborting in debug/sanitizer builds. #103544 (Groene AI).
• Fix a server crash in JSON and Dynamic data type parameter parsing when the abstract syntax tree (AST) was structurally malformed (for example produced by the AST fuzzer): the equals(name, value) parameter expression could be left with fewer than two children, which DataTypeObject and DataTypeDynamic then dereferenced without bounds checking. #103545 (Groene AI).
• Fix logical errors in analyzer with column names collision in lambda used in prewhere. Closes #103584. #103627 (Pavel Kruglov).
• Fix sporadic Logical error: 'Database <name> not found' from DataLakeConfiguration::getCatalog when an Iceberg engine table is loaded inside a regular database during async metadata loading. #103775 (Groene AI).
• Fix MULTIPLE_EXPRESSIONS_FOR_ALIAS errors thrown by remote replicas when running queries that reference projection aliases inside PREWHERE / WHERE / HAVING / QUALIFY (e.g. SELECT x AS a, y AS b, (a AND b) AS c FROM t PREWHERE c) or SELECT * over self-joins with overlapping column names, with parallel replicas and parallel_replicas_local_plan = 0. Closes #74324. #103806 (Groene AI).
• Fix server abort when a query uses nested coalesce/ifNull comparisons (e.g. WHERE coalesce(a, b, coalesce(c, d), e) = const) on a MergeTree table with multiple minmax skip indexes and use_skip_indexes_for_disjunctions = 1. The skip-index rewrite of <op>(coalesce(...), const) is now applied recursively to inner coalesce arguments, so the per-index KeyCondition RPN matches the template’s RPN as the disjunction-tracking code already assumes. #103929 (Groene AI).
• Fix max_rows_to_transfer and max_bytes_to_transfer being silently ignored for GLOBAL IN and GLOBAL JOIN queries under the new analyzer. The settings now raise SET_SIZE_LIMIT_EXCEEDED (or break, depending on transfer_overflow_mode) when the materialised external table exceeds the configured limit, matching the behavior of the old analyzer. Closes #103333. #104119 (Groene AI).
• Fix excessive catalog/S3 metadata reads when an INSERT or DDL statement references a non-existent table in a DataLake catalog database with show_data_lake_catalogs_in_system_tables enabled. The typo-hint suggestion path loaded full per-table Iceberg metadata for the whole catalog, which could exhaust memory on large catalogs. #104124 (DQ).
• Fix a crash (LOGICAL_ERROR “Columns are assumed to be of identical types, but they are different in Nullable” in debug builds, a ColumnString::compareAt null-pointer SIGSEGV in release builds) and a wrong result that could occur in queries using direct join_algorithm with a MergeTree right-side table when the lookup plan reordered columns during optimization. The right-side columns could land in wrong-named slots, so a filter or comparison on the join key saw a column of the wrong type. Surfaced by AST fuzzer as STID 2139-5111 and reported in #107272. #104174 (Groene AI).
• Reject out-of-range integer partition ID for Date instead of silent overflow or LOGICAL_ERROR on ALTER. #104250 (Azat Khuzhin).
• Disable additional_table_filters with parallel replicas and the analyzer without query plan serialization. Previously it could lead to incorrect results. #104296 (Azat Khuzhin).
• Fix dictGetOrNull silently overwriting other columns in the SELECT projection with NULL when called with a Nullable key column whose values are missing in the dictionary. The function was mutating an input-aliased null map in place; it now deep-clones the result column before mutation. Closes #73633. #104327 (Groene AI).
• Fix a broken patch part after ALTER TABLE ... DROP PARTITION ID 'patch-...' followed by DETACH/ATTACH TABLE. Previously the empty covering part was written without partition.dat and source_parts.dat, leading to a broken-on-start_patch-... entry in system.detached_parts after the next attach or server restart. Closes #93132. Closes #102103. #104353 (Groene AI).
• Fixed streaming INSERT with input_format_max_block_wait_ms for the HTTP interface and for INSERT SELECT FROM input, so partial input blocks are flushed before the request finishes. #104534 (Alexey Milovidov).
• Fixed a bug where queries combining arrayJoin with ORDER BY ... LIMIT after a JOIN could silently return zero rows. The query plan optimization that lifts function evaluation above the SortingStep no longer applies when the lifted expression contains arrayJoin, since arrayJoin can change the number of rows. Closes #82279. #104558 (Groene AI).
• Fix a rare server abort during MergeTree family table destruction. The abort could occur when shutdown raised an exception before the background-job scheduler was fully stopped: a background task firing in that narrow window would dispatch a virtual call through a partially-destroyed storage object and hit __cxa_pure_virtual. #104561 (Groene AI).
• Fix inconsistent result in JSON combined subcolumn with a type hint containing null value. #104584 (Pavel Kruglov).
• Fix wrong results or missed projection when an aggregate projection contains multiple sumIf aggregates with different IN (...) conditions. #104765 (Jimmy Aguilar Mena).
• Fix wrong row count from arrayJoin() used inside JOIN ON: the result was multiplied by the array length an extra time. #104785 (Shaohua Wang).
• Fix deltaSumTimestamp returning wrong results for signed integer types crossing zero. Resolves #104750. #104830 (Konstantin Bogdanov).
• Fix block structure mismatch in UnionStep after filter push-down with NULL constants. Closes #104821. #104853 (Pavel Kruglov).
• Fix Logical error: 'Metadata is not initialized' raised by DELETE FROM on a freshly-attached Iceberg, DeltaLake, or Hudi table whose metadata file is corrupted or unloadable. A regular user-facing exception is reported instead, and the server keeps running. Closes #104891. #104917 (Groene AI).
• Fixed the max_memory_usage_soft_limit was not updated on-the-fly when the configuration changed. The soft memory limit was calculated once at startup and cached, so subsequent config reloads had no effect. After this fix, the limit is recalculated and applied immediately on every config reload, no Keeper restart required. #104940 (Kai Zhu).
• Fix JIT symbol resolution on macOS so that sort, expression, and aggregate JIT compilation no longer fail with CANNOT_COMPILE_CODE: Could not find symbol _<name>. #104946 (Alexei Fedotov).
• Fix an exception (Received signal 6 (abort) in vector hardening, observable as Tuple::back() on an empty vector) when executing a query of the form SELECT ... FROM <Distributed table> WHERE/HAVING/GROUP BY ... (sharding_key_column IN tuple()) with optimize_skip_unused_shards_rewrite_in enabled under the new analyzer. #104966 (Alexey Milovidov).
• Fix a Logical error: 'Inconsistent AST formatting' server abort (STID 1941-1bfa) on queries like CREATE TABLE t (c0 Int CODEC(not((not(materialize(1), materialize(2)))), ZSTD)) ENGINE = Memory in debug / sanitiser builds. The formatter no longer emits redundant outer parens around a multi-argument tuple(...) function call that appears inside CODEC / STATISTICS / BACKUP_NAME argument lists (where the operator form (a, b) is disabled), keeping the format-parse-format round-trip stable. #104991 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK thrown by ALTER TABLE ... MATERIALIZE INDEX on parts that were created in 25.8 and contain a skip index over a column that was added with a separate ALTER TABLE ... ADD COLUMN. The mutation now correctly reads every column required by every pre-existing skip index and projection on the part during force-recalculation. Closes #104872. #105039 (Groene AI).
• Fixed two LOGICAL_ERROR: Reading from materialized CTE 'X' before it has been materialized shapes thrown by queries with enable_materialized_cte: (1) a reused materialized CTE filtered by IN (subquery) over another materialized CTE, and (2) a materialized CTE referenced both directly and inside a WHERE ... IN (...) filter that hits a MergeTree primary key through a nested IN-subquery. EXPLAIN on the same queries was affected too because the bugs fired during plan optimization. Closes #101940, closes #102320. #105041 (Dmitry Novik).
• Fix countmin statistics PREWHERE selectivity estimation for Float32 columns. Two bugs caused countmin statistics to silently produce wrong estimates for Float32 columns: ConditionSelectivityEstimator skipped statistics entirely when the column type (Float32) was narrower than the literal type (Float64), falling back to default selectivity and ignoring the sketch; and StatisticsCountMinSketch::estimateEqual hashed the first N bytes of a Field object, but Field stores Float32 as Float64 (NearestFieldType<Float32> = Float64), so the byte pattern used at query time differed from the actual Float32 bytes stored during build. #105047 (Han Fei).
• ATTACH TABLE name <clauses>; queries that supply storage clauses (ORDER BY, PARTITION BY, PRIMARY KEY, SAMPLE BY, TTL, UNIQUE KEY, or engine SETTINGS) without an ENGINE now throw BAD_ARGUMENTS instead of silently re-attaching the table with its stored definition and discarding the user-supplied clauses. Query-level session SETTINGS (such as log_comment) are still applied. Use ATTACH TABLE t; to re-attach with stored metadata, or ALTER TABLE t MODIFY SETTING ... after ATTACH to change settings. #105068 (Groene AI).
• Fixes a use-after-free in runningAccumulate when called on a column whose aggregate function returns its own state (e.g. uniqStateOrDefaultState, sumStateOrDefaultState, uniqStateForEachState). #105085 (János Benjamin Antal).
• Fixed a bug in the Web UI (play.html): running a query while the cursor was past the only trailing ; reported “empty query” instead of running the query. #105107 (Alexey Milovidov).
• Fix LOGICAL_ERROR: 'No user in current context, it's a bug' when a CREATE TABLE ... AS SELECT inside a Replicated database reads from system.current_roles or system.enabled_roles. The query now succeeds and returns an empty result set in this internal DDL worker context. #105150 (Groene AI).
• Fixed ILLEGAL_TYPE_OF_ARGUMENT when merging quantileExactWeightedInterpolated, quantileDD, or quantilePrometheusHistogram aggregate states with their plural quantilesXxxMerge counterparts (and vice versa). The singular and plural variants of these three quantile families share the same internal aggregate state but were not listed in the internal name-mapping table, so cross-function state merge — and the function-fusion optimization for these families — were rejected. #105189 (Groene AI).
• Fix incorrect results of toStartOfWeek, toLastDayOfWeek, toMonday, toStartOfMonth, toLastDayOfMonth, toStartOfQuarter and toStartOfYear for Date32 and DateTime64 arguments whose result falls outside the Date range: instead of overflowing into arbitrary dates, results before 1970-01-01 are now clamped to 1970-01-01 and results after 2149-06-06 are clamped to 2149-06-06. This also fixes wrong query results (incorrectly pruned parts and granules) when such functions were used in WHERE over a Date32 or DateTime64 key containing out-of-range values. #105244 (Yarik Briukhovetskyi).
• Fix a logical error in arrayRemove when the first argument is an array of Variant whose alternatives are all incompatible with the type of the second argument and variant_throw_on_type_mismatch is disabled. The function now treats the comparison as “never equal” and returns the array unchanged instead of triggering a server-side assertTypeEquality failure. #105248 (Groene AI).
• Query can be executed successfully even if FileCache fails to create some directory on disk for caching. #105250 (Den Kalantaevskii).
• Fix server abort with Assertion 'px != 0' failed in ExtremesTransform when a query with extremes = 1 hits an exception (for example MEMORY_LIMIT_EXCEEDED) while the transform is building its extremes columns from the first chunk. The client now receives the original exception instead of crashing the server. #105264 (Groene AI).
• LEFT ANTI JOIN now correctly returns the right-side join key column with default values for unmatched rows, instead of duplicating the left key value. Previously, SELECT a.x AS l, b.x AS r FROM a LEFT ANTI JOIN b ON a.x = b.x could return r = a.x for the unmatched rows when it should return the default (0 for Int32, '' for String, NULL for Nullable). Only the join key column was affected; non-key right columns were already defaulted correctly. The same bug appeared in RIGHT ANTI JOIN whenever the optimizer swapped sides. Closes #99959. #105278 (Groene AI).
• Fix data loss in clickhouse-local when started repeatedly with a non-default --default_database (for example clickhouse-local --path X -- --default_database=mydb): the lookup for the persisted database UUID was hardcoded to metadata/default, so each restart minted a fresh UUID and the previous run’s tables became invisible. Closes #101831. #105284 (Groene AI).
• Fix a logical error Mutation of Memory table produced incomplete output raised when running ALTER TABLE <memory_table> APPLY PATCHES or ALTER TABLE <memory_table> APPLY DELETED MASK. Memory tables do not own patch parts or deletion masks, so both commands are now correctly treated as no-ops. #105286 (Groene AI).
• Fix long-running queries with Unity Catalog by passing a table_id parameter in the query request. #105303 (Konstantin Vedernikov).
• Fix CANNOT_CONVERT_TYPE for Merge over Merge over Distributed with distributed_group_by_no_merge=1. #105330 (Azat Khuzhin).
• Fix clickhouse local and clickhouse client silently ignoring --query/-q when a positional file argument was also given. Both forms now produce BAD_ARGUMENTS consistently. #105334 (Raúl Marín).
• Fix REPLACE TEMPORARY TABLE silently creating a new table when the target didn’t exist. It now throws UNKNOWN_TABLE, matching REPLACE TABLE semantics. Use CREATE OR REPLACE TEMPORARY TABLE to keep the create-or-replace behavior. #105373 (Groene AI).
• Fixed long-running reads (e.g. INSERT ... SELECT) from a DataLakeCatalog database with catalog_type = 'glue' failing with an ExpiredToken error once the STS session credentials captured at table-metadata load time crossed their expiry. The AWS STS credentials are now refreshed mid-query, so such reads continue past the token’s expiry. #105381 (Pratima Patel).
• Fix logical error in negative LIMIT BY in some cases when used with ARRAY JOIN. #105403 (Nihal Z. Miaji).
• Fix inflated read_bytes (and the derived bytes/s shown in system.query_log, progress bar, etc.) when reading Parquet files. The previous implementation reported the row group’s total compressed size on every chunk, so reading K of N columns overcounted by N / K. It is now summed only across the selected columns. Also fixes file-level progress tracking for Iceberg tables, which previously never reported the data file size. #105413 (Groene AI).
• Fixed a possible server segfault in cluster table functions (s3Cluster, urlCluster, fileCluster, …) when the planner produces a SELECT with the recursive_with flag set but no WITH expression. Closes #105370. #105433 (Groene AI).
• Fix wrong result of countDistinct / uniqExact on Nullable columns when count_distinct_optimization = 1 (the NULL group was incorrectly counted). #105439 (Raúl Marín).
• Fixed a heap-buffer-overflow when reading Arrow or ArrowStream files with corrupted intermediate offsets in a binary or string column, and a null-pointer dereference when reading geo-tagged Arrow columns. #105449 (Raúl Marín).
• Validate Dynamic/Variant types in window PARTITION BY, it was not checked before when allow_suspicious_types_in_group_by is disabled. Closes #105028. #105450 (Pavel Kruglov).
• Fix input_format_json_empty_as_default not working in JSONStrings* formats. Closes #104913. #105453 (Pavel Kruglov).
• Fix NULL insert into Variant column via VALUES format during expression parsing. Closes #104909. #105455 (Pavel Kruglov).
• Fix dictGetOrDefault throwing CANNOT_INSERT_NULL_IN_ORDINARY_COLUMN when the default argument is a Nullable expression and short_circuit_function_evaluation is enabled with a mix of found and not-found dictionary keys. #105461 (Raúl Marín).
• Fix silent value truncation in Dynamic columns when the Quoted (Values) text reader falls back to String for incomplete inferred types like Map(Array(Nothing), ...). Previously the fallback wrapped the raw field with naive '...' concatenation, so an inner single quote (always present for such fallbacks) prematurely terminated the quoted string and the rest of the value was lost. #105463 (Groene AI).
• Fix SIGSEGV when reading Dynamic subcolumns from compressed Memory table after ALTER. Closes #104901. #105464 (Pavel Kruglov).
• Fix toFloat64/toUInt32/toString/etc. on Dynamic ignoring cast_keep_nullable. Closes #104789. #105467 (Pavel Kruglov).
• Include skip_first_lines in schema cache key for WithNames formats. Closes #104527. #105469 (Pavel Kruglov).
• Fix a server segfault in uniqStateOrNull / uniqStateOrDefault / uniqOrNullState (and similar combinator chains over uniq) when used with GROUP BY ... WITH ROLLUP, WITH CUBE, or WITH TOTALS and a Nullable argument. #105470 (Groene AI).
• Fix NOT_FOUND_COLUMN_IN_BLOCK exception when combining ORDER BY ... WITH FILL INTERPOLATE and LIMIT N BY with the analyzer enabled. Closes #103474. #105481 (Yakov Olkhovskiy).
• Fix toStartOfMillisecond and toStartOfMicrosecond returning a result off by nearly a second for negative (pre-epoch) DateTime64 values, and fix UndefinedBehaviorSanitizer signed-integer-overflow in toStartOfSecond, toStartOfMillisecond, and toStartOfMicrosecond for DateTime64 inputs near INT64_MIN. #105482 (Groene AI).
• Add a new compatibility setting analyzer_compatibility_prefer_alias_over_subcolumn (disabled by default). When enabled, the new analyzer prefers the alias-prefix interpretation over Tuple-subcolumn / dotted-column matches for multi-part identifiers, restoring the previous interpreter’s behavior. This avoids AMBIGUOUS_IDENTIFIER (and related) errors when a query joins a table whose name matches an inner table of a CTE/subquery that uses SELECT * over a join, where asterisk-renamed columns (e.g. b.id) would otherwise leak the inner table identifiers into the outer scope. #105491 (Vladimir Cherkasov).
• A vector search query executed with the index only optimization now returns the correct value for L2Distance() if it is in the SELECT list. The function was incorrectly returning the L2-squared distance value. Note that the vector search query was returning the correctly ranked results, only applications explicitly retrieving and interpreting the L2Distance() value would have seen the squared value. #105495 (Shankar Iyer).
• Fixed a rare issue in new Kafka table engine where messages could be incorrectly handled during topic partition reassignment, potentially leading to skipped messages after offset rollback. #105500 (János Benjamin Antal).
• Fix an exception (Trying to execute PLACEHOLDER action logical error) when a MATERIALIZED CTE whose body is a correlated subquery is used as the right-hand side of IN. Such a CTE is now rejected at analysis time, consistent with how the same pattern is already rejected when the CTE is referenced directly in FROM. #105518 (Groene AI).
• Fix stdout/stdin compression not applied when format is explicitly specified. Closes #104441. #105528 (Pavel Kruglov).
• Fix SEGFAULT in singleValueOrNull deserialization for JSON type. Closes #103630. #105535 (Pavel Kruglov).
• Fix clickhouse-format --backslash silently dropping INSERT VALUES data. Closes #103533. #105536 (Pavel Kruglov).
• Fix variant_throw_on_type_mismatch/dynamic_throw_on_type_mismatch=false not catching exceptions during function execution. Closes #103484. #105543 (Pavel Kruglov).
• Fix ignoring input_format_try_infer_datetimes during insertion into shared data in JSON. Closes #103221. #105544 (Pavel Kruglov).
• Fix histogram producing wrong results for small unsorted inputs. Closes #103109. #105548 (Pavel Kruglov).
• Fix DateTime wrapping around for out-of-range values in JSONExtract and text deserializations. Closes #103094. #105551 (Pavel Kruglov).
• Fix usage of insertion table in table functions when optimize_trivial_insert_select is enabled. Closes #103083. #105555 (Pavel Kruglov).
• Fix CASE expression returning the ELSE branch instead of the matching THEN when both the expression and a WHEN value were NULL. #105556 (Raúl Marín).
• Fixed a crash in replxx caused by hitting the FD_SETSIZE limitation of select (now replaced with poll). Previously, with SSH or the web terminal enabled, this could crash the server. #105559 (Azat Khuzhin).
• Materialize subcolumns in executePartitionByExpression before executing expression. Closes #103057. #105573 (Pavel Kruglov).
• Fix NOT_FOUND_COLUMN_IN_BLOCK exception when TTL expression references a subcolumn. #105578 (Pavel Kruglov).
• Fix a server crash (SIGSEGV) reachable by any user with CREATE TABLE rights when sending CREATE TABLE ... TO INNER UUID '...' without an ENGINE clause over HTTP or the native protocol. The same bug also crashed the client. The parser now reports a proper BAD_ARGUMENTS error instead of dereferencing a null pointer. #105579 (Groene AI).
• Fix estimateCompressionRatio window function losing accumulated data between rows. Closes #101738. #105581 (Pavel Kruglov).
• Fix crash when inserting tuples of different sizes in the same VALUES clause into a String column. Closes #101727. #105582 (Pavel Kruglov).
• Hive partition value extraction now honors the cast_string_to_date_time_mode setting and accepts ISO 8601 timestamps with timezone suffixes (e.g. +0000, +00:00, Z) in partition keys by default. #105584 (Alexey Milovidov).
• Fixes NOT_IMPLEMENTED error on toString from DateTime with Timezone containing NULL value. Closes #103712. #105587 (Yarik Briukhovetskyi).
• Fix OOB read in flattened Dynamic column Native deserialization. #105666 (Pavel Kruglov).
• Fix a LOGICAL_ERROR (Unexpected return type from if) raised during query planning for if expressions whose result type is Variant and whose second-or-third branch is a constant-condition if over a UInt64 literal that fits into Int64. Closes #105649. #105680 (Groene AI).
• Fix wrong results when reading an Iceberg table with iceberg_use_version_hint = 1 after another writer (such as the icebergLocal/icebergS3 table function) without the setting advances the table. version-hint.text is now kept in sync by every writer once the file exists, so subsequent readers using the hint see the latest snapshot. #105682 (Groene AI).
• Fixed a silent under-count in SELECT queries when use_query_condition_cache = 1 (default). A query of the shape PREWHERE pk_prefix = X WHERE non_pk IN (...) against a column with a bloom-filter skip index poisoned the QueryConditionCache for the pk_prefix = X predicate, so a subsequent benign SELECT count() ... WHERE pk_prefix = X returned an incorrect, under-counted result. Affected all 26.x releases. #104781. #105686 (Groene AI).
• Fixed singleValueOrNullMerge returning a concrete value instead of NULL when merging a state that had already observed multiple distinct values. #105734 (Minh Vu).
• Fix Template input format error recovery after malformed rows. #105735 (Yue Ni).
• Fix a crash in the mongodb table function, MongoDB storage and MongoDB dictionary source when the collection name is empty or contains NUL bytes. #105776 (Raúl Marín).
• Fix Keeper snapshot cleanup after failed writes so partial snapshots are cleaned up safely and failed writes can be retried without advancing latest_snapshot_meta. #105779 (Antonio Andelic).
• Fix ALTER TABLE ... CLEAR COLUMN being rejected for explicit SummingMergeTree and CoalescingMergeTree columns_to_sum columns. #105785 (Antonio Andelic).
• Fixed an assertion in DatabaseCatalog::tryGetDatabase (and resulting UNKNOWN_DATABASE error in release builds) when creating a parameterized view whose database name is supplied via a query parameter, for example CREATE VIEW {db:Identifier}.v AS SELECT {p:UInt64}. Parameters in the DDL parts of the CREATE statement are now substituted at create time, while parameters in the SELECT body remain available for substitution at view-call time. #105799 (Groene AI).
• Fix Bad get: has Decimal32, requested Decimal128 from sumMap and sumMapWithOverflow over a Nested(... Nullable(Decimal(P, S))) column when the aggregate state is serialised (e.g. parallel replicas, sumMapState via a binary-state formatter, external aggregation). #105816 (Groene AI).
• Fix Expected ColumnLowCardinality, got String / Bad cast from type DB::ColumnString to DB::ColumnLowCardinality errors when apply_mutations_on_fly = 1 is used on a table with pending on-fly UPDATE/DELETE mutations queued before an ALTER MODIFY COLUMN ... LowCardinality(...) mutation. #105847 (Raúl Marín).
• Iceberg writes now preserve NULL values in Nullable(T) partition columns. Previously, a NULL written by ClickHouse showed up as the default value of the inner type (0 for int) when read back by Spark or other Iceberg readers. Closes #105852. #105862 (Groene AI).
• Fixed Parquet and ORC filter pushdown for IN (subquery) predicates, allowing row-group/page/bloom-filter pruning to work for file, url, s3, and object-storage reads. #105863 (Arsen Muk).
• Fix read-in-order for Merge tables (with a new analyzer). #105867 (Nikolai Kochetov).
• Fixed Iceberg v2 merge-on-read position deletes returning wrong rows when a single delete file references multiple data files and several such delete files apply to the same data file. The streaming reader (use_roaring_bitmap_iceberg_positional_deletes = 0) now filters delete-file rows by file_path in C++ instead of relying on Parquet row-group pruning, restoring the ascending-positions invariant. #105888 (Groene AI).
• Fix Cannot find column error for distributed queries with IN Array(...) filter for the new analyzer. #105894 (Nikolai Kochetov).
• Fix a SIGSEGV in ALTER TABLE ... MODIFY COLUMN ... Nullable(...) on MergeTree tables when a column with STATISTICS is concurrently dropped by another ALTER. Closes #105912. #105917 (Groene AI).
• Fixed a server crash that could occur when a query reading from PostgreSQL — via the postgresql table function, the PostgreSQL table engine, or a dictionary with a PostgreSQL source — was cancelled (for example with KILL QUERY) and cancelling the remote PostgreSQL query failed. #105949 (Rory Shanks).
• Fix SYSTEM INSTRUMENT ADD formatting so handler arguments are separated by a single space, and reject invalid SLEEP instrumentation argument lists with more than two values or a range where the minimum is greater than the maximum. #105984 (Pablo Marcos).
• Fix possibly wrong results for queries that combine an outer join with a subsequent inner join referencing the outer join’s null-supplying side. Join reordering could pick a plan that pulled inner-join conditions into the outer join’s ON clause. #105992 (Vladimir Cherkasov).
• Fixed a possible crash due to a too-large string literal sent within the query. #105996 (Nikita Taranov).
• Fix INVALID_WITH_FILL_EXPRESSION exception when using INTERPOLATE () (empty) with a sorting prefix in ORDER BY and use_with_fill_by_sorting_prefix enabled. Sorting prefix columns are now correctly excluded from the implicit interpolation set, matching the behavior of explicitly named INTERPOLATE (col). #106001 (Yakov Olkhovskiy).
• Fix ALTER TABLE partition operations silently failing for Bool partition keys. Closes #101722. #106004 (Pavel Kruglov).
• Fix JSONExtractRaw and JSONHas for typed JSON paths with default values. Closes #101721. #106005 (Pavel Kruglov).
• Fix incorrect ’/‘-prefix addition on empty basepaths in data lake configurations. Closes #105989. #106013 (thewisenerd).
• Fix IcebergLocal table engine becoming read-only after a DETACH + ATTACH cycle or a server restart, which made every subsequent INSERT fail with Local object storage Local is readonly. (READONLY). #106016 (Groene AI).
• Fix Keeper failures during follower catch-up when the new request dispatcher response queue could fill before the response thread started. #106049 (Antonio Andelic).
• Better compatibility with the old analyzer. If the table has columns like x.a Array, x.b Array, x String, prefer arrays for ARRAY JOIN x. #106069 (Nikolai Kochetov).
• Fixed the filesystem cache being silently disabled for Azure Blob Storage (e.g. Delta Lake tables over Azure) because object metadata did not include the blob ETag. #106091 (thewisenerd).
• Fix system.dictionaries returning 0 rows with partial SHOW DICTIONARIES revoke. #106105 (Pavel Kruglov).
• Fixed a server crash when querying DeltaLake tables with allow_experimental_delta_kernel_rs enabled and a credential or option that contained invalid bytes (the Rust FFI panicked across the extern "C" boundary). #106109 (Raúl Marín).
• Fix incorrect results for queries against tables whose ORDER BY contains a monotonically decreasing function such as (c0 / -42) or intDiv(c0, -42). Predicates on the underlying column (for example, c0 < 0) could wrongly prune granules that contained matching rows, producing missing results. Closes #106084. Closes #106124. Closes #106080. #106136 (Nihal Z. Miaji).
• Fix support for using WASM SQL UDFs in MATERIALIZED VIEW definitions. #106161 (Yue Ni).
• Iceberg partition pruning now correctly handles WHERE partition_col = (SELECT ... FROM ...) filters where the analyzer wraps the scalar subquery result in an internal _CAST(Const, 'TargetType') with matching source and target types. Previously such filters disabled partition pruning and triggered a full table scan. #106204 (Groene AI).
• Fix clickhouse local’s --query_id parameter to allow it to specify custom query id. #106205 (Yue Ni).
• Fixed the S3 storage class (s3_storage_class / s3_storage_class_name) being ignored for objects written via multipart upload on S3 disks and object storage, which caused large objects to be created with the default STANDARD class. The option name is now accepted both as s3_storage_class and s3_storage_class_name for disks, object storage and backups. #106214 (Alexey Milovidov).
• Fixed Keeper sometimes getting stuck on startup when setting nuraft_max_log_gap_in_stream is set to non-default value (default is 0, i.e. disable pipelining of append_entries requests). #106220 (Michael Kolupaev).
• Validate corrupted DDSketch aggregate-function state (bin keys) on insertion instead of accepting malformed data. #106236 (Yarik Briukhovetskyi).
• Fixed coordinator mode mismatch when subqueries override in-order settings. #106243 (Nikita Taranov).
• Fix optimize_skip_unused_shards did not apply (and force_optimize_skip_unused_shards falsely failing) when a Distributed table is queried through a Merge table or the merge table function, with the predicate applied above it. #106250 (Nikolai Kochetov).
• Fix INTERPOLATE () throwing INVALID_WITH_FILL_EXPRESSION when ORDER BY columns are aliased in the SELECT list with the old analyzer. Closes #106248. #106252 (Yakov Olkhovskiy).
• Fixed an abort when deserialising a malformed AggregateFunction(uniqTheta, ...) state from RowBinary input or a query parameter. The bad input is now rejected with CORRUPTED_DATA instead of escaping as a std::exception and aborting the process via abortOnFailedAssertion. #106260 (Groene AI).
• Reject out-of-range IntervalKind bytes during RowBinaryWithNamesAndTypes type decoding (input_format_binary_decode_types_in_binary_format = 1) with a clear INCORRECT_DATA error instead of constructing a DataTypeInterval with an invalid kind that could subsequently trip undefined behavior in hash paths. #106261 (Groene AI).
• Fix SLRU downgrade failure rollback issue in 26.1+. Fix System/Data cache split priority with keep_free_space_ratio feature. #106286 (Kseniia Sumarokova).
• Read the null subcolumn as a JSON path in Nullable(JSON) instead of the null-map. Closes #106085. #106295 (Pavel Kruglov).
• Fix RestCatalog::empty returning the wrong result when an Iceberg REST catalog contains tables. #106301 (Lefteris).
• Fixed an exception for INNER JOIN queries with an empty left MergeTree table when enable_parallel_replicas, query_plan_use_new_logical_join_step, and query_plan_optimize_join_order_algorithm = 'greedy' are enabled. #106338 (Nikolai Kochetov).
• Fixed an incorrect conversion of subnormal Float16 values to Float32 (e.g. when reading them from Numpy .npy files), caused by an off-by-one mantissa shift. #106343 (Joanna Hulboj).
• Fix a crash and a possible NOT_FOUND_COLUMN_IN_BLOCK error when constraint-based optimization (optimize_using_constraints) is used with correlated subqueries. #106349 (Raúl Marín).
• Fixed an out-of-bounds read in sipHash64Keyed, sipHash128Keyed and sipHash128ReferenceKeyed when hashing a column whose arrays are all empty and the key is not constant. #106355 (Raúl Marín).
• Fixed SYSTEM RELOAD CONFIG discarding per-endpoint Azure Blob Storage settings (such as use_native_copy), which caused a disk’s configured settings to be ignored for BACKUP/RESTORE until the server was restarted. #106357 (Julia Kartseva).
• Fix regexpExtract(haystack, pattern) so that patterns without a capturing group return the whole match instead of throwing INDEX_OF_POSITIONAL_ARGUMENT_IS_OUT_OF_RANGE. #106374 (Groene AI).
• Fix LOGICAL_ERROR exception during cache predownload when a remote S3 object is overwritten with shorter content between listing and reading. #106375 (Nikita Fomichev).
• Fix memory usage growth in filesystem cache. #106387 (Kseniia Sumarokova).
• Fix multiple heap out-of-bounds reads in the Arrow IPC format reader (ArrowColumnToCHColumn). A malformed Arrow file could declare more rows than its buffers contain, declare list/struct/map child lengths inconsistent with their parent, supply non-monotonic list offsets, or truncate a child validity bitmap, causing reads past the end of heap allocations. This is reachable by any user with SELECT privilege via file(), format(), table functions, or ArrowFlight inputs. All data, offsets, view-struct, and validity-bitmap buffers are now validated before any raw pointer access, and list/struct/map shapes and offsets are checked for consistency. #106395 (Raúl Marín).
• Fixed backups failing with FILE_DOESNT_EXIST when a refreshable materialized view’s REPLACE target is collected on a Replicated or Shared database whose materialized view isn’t yet instantiated on the backup-initiating replica. #106411 (Julia Kartseva).
• Fix a logical error Column identifier ... is already registered when a mutation (DELETE/UPDATE) predicate contains an IN/EXISTS subquery that reads from a default table expression nested in another subquery. #106414 (Alexey Milovidov).
• Fix a logical error (Left and right columns have same names) in the join order optimizer that could occur for joins executed with parallel replicas when two relations in the join graph share column names. #106418 (Alexey Milovidov).
• Fix a logical error when building a polygon dictionary from source data containing NaN or infinite point coordinates. Such coordinates are now rejected with a clear error. #106423 (Alexey Milovidov).
• Fixed a bug where the used_privileges and missing_privileges columns of system.query_log could contain privilege strings leaked from unrelated earlier queries of a different user, database, or session. #106425 (Alexey Milovidov).
• Functions base58Encode, base58Decode and tryBase58Decode now respect max_execution_time and query cancellation on large inputs, and reject inputs larger than 10 KB instead of running for a very long time. The limit is configurable via the new setting function_base58_max_input_size (0 disables it). #106428 (Alexey Milovidov).
• Fix sporadic incorrect results for ORDER BY ... DESC queries when reading wide parts in reverse order with read_in_order_use_virtual_row_per_block = 1 and a small max_block_size. #106429 (Vladimir Cherkasov).
• Fixed a NOT_FOUND_COLUMN_IN_BLOCK exception when querying an Iceberg or S3 table with a compound WHERE containing IS NOT NULL on a column that is not in the SELECT list, using the Parquet V3 native reader. #106443 (Shaohua Wang).
• Allows user to specify headers for HTTPDictionary using named collections. #106459 (Jan Rada).
• Fix a case where a worker thread could stay attached to a stale thread group after CurrentThread::attachToGroup failed part way through, causing later tasks on the same thread to fail with Thread is already attached to a group. #106462 (Mikhail f. Shiryaev).
• Fixed an exception when a vector search query uses a vector index and uses another skip index like minmax and use_skip_indexes_on_data_read = 1. #106473 (Shankar Iyer).
• Malformed Avro enum values are now validated and rejected with an error instead of causing an out-of-bounds read and abort when deserializing corrupted Avro data. #106476 (Miсhael Stetsyuk).
• Fix inflated progress reporting when reading from Iceberg tables with _file or _path filters. Previously, total_bytes_to_read progress included all files from the manifest regardless of filtering. #106491 (Pedro Ferreira).
• Fixed a Bad cast exception for Redis dictionaries that use STORAGE_TYPE 'simple' with a cache/direct layout and a single string (complex) key; such dictionaries now work, and composite keys over simple storage report a clear error. #106501 (Vladimir Cherkasov).
• Fix wrong-results bug where WHERE c0 = const returned no rows for tables with ORDER BY f(c0) when f(const) evaluates to NaN, e.g. ORDER BY sqrt(c0) with a negative constant. The primary-key analysis incorrectly pruned every granule and poisoned the query condition cache for subsequent queries. #106507 (Groene AI).
• Fix LIMIT WITH TIES and fractional LIMIT WITH TIES not respecting the collation from ORDER BY ... COLLATE when determining ties. Rows that are equal according to the collation (for example '1' and '01' under numeric collation) were compared byte-wise, so some tied rows were wrongly dropped from the result. #106539 (Nihal Z. Miaji).
• Fix DISTINCT in order and LIMIT BY in order optimizations (including negative LIMIT BY) returning wrong results when the input is sorted with a collation (ORDER BY ... COLLATE). Rows that are equal according to the collation (for example 'a' and 'A' under a case-insensitive collation) are ordered by collation key and are not adjacent by value, so the in order optimization is now skipped when a collator is used. #106564 (Nihal Z. Miaji).
• Fixed wrong results for SELECT c, count() FROM t WHERE c GROUP BY c against tables with the implicit _minmax_count_projection, an explicit aggregate projection, or a normal projection: every group used to collapse into one row with a constant key and the total row count. #106590 (Groene AI).
• Fixes a bug where users could not use scalar subqueries in the first argument of IN where the second argument is a non-constant tuple. #106610 (Yarik Briukhovetskyi).
• Fix exception Mutation of Memory table produced incomplete output raised when running ALTER TABLE <memory_table> commands that have no per-row data effect on a Memory engine, namely APPLY PATCHES, APPLY DELETED MASK, MATERIALIZE STATISTICS, MATERIALIZE INDEX, MATERIALIZE PROJECTION, and REWRITE PARTS. Memory tables do not own those structures, so such commands are now treated as no-ops. #106621 (Groene AI).
• Fix session_timezone being ignored when serializing LowCardinality(DateTime) columns to text formats (CSV, TSV, JSONEachRow, etc.). Previously, after the first write of a LowCardinality(DateTime) column on a server, every subsequent query that wrote such a column rendered the wall-clock string in whichever timezone was first seen, regardless of session_timezone. #106634 (Groene AI).
• Fixed a LOGICAL_ERROR “Trying to get name of not a column: ExpressionList” raised by queries that pass an asterisk inside multiIf to a table function argument, e.g. numbers(multiIf(*, ...), 2). The query now rejects the unresolvable matcher with UNSUPPORTED_METHOD. #106647 (Groene AI).
• Fix server failing to start with Too many marks in file ...skp_idx_idx.cmrk4, marks expected 0 (bytes size 0) when a MergeTree table has a skip-index part with zero granules and a non-empty marks file on disk. #106675 (Groene AI).
• Reject pathological file glob patterns that would cause unbounded recursion in directory listing. A maximum recursion depth of 1000 is now enforced; queries that exceed it raise TOO_DEEP_RECURSION instead of crashing the server with a stack overflow. #106676 (Groene AI).
• Fix server abort Bad cast from type DB::ColumnNothing to DB::ColumnVector<char8_t> in FunctionIf::executeForConstAndNullableCondition when the if/multiIf condition constant-folds to Const(Nullable(Nothing)) (for example, an out-of-range subscript on an empty array, like arraySort(x -> x, [])[toNullable(1)]). #106678 (Groene AI).
• Fix exception 'Trying to read from input() twice.' raised when the table function input is wrapped in a non-MATERIALIZED CTE that is referenced from more than one place in the query. The query is now rejected with a clean INVALID_USAGE_OF_INPUT error at planning time. input is a one-shot client stream and can only be consumed by a single source in the query plan. #106682 (Groene AI).
• Make FORMAT apply to the EXPLAIN output of EXPLAIN ... INSERT ... SELECT ... FORMAT ... also when SETTINGS precede the FORMAT or the output format is Values. Follow-up to #101772. #106686 (Alexey Milovidov).
• Fixed a rare spurious RESOURCE_ACCESS_DENIED error (“Scheduler queue with resource request is about to be destructed”) for a query that was actually granted access to a workload resource, caused by a reused thread-local request object retaining a previous request’s failure. #106690 (Alexey Milovidov).
• Fix a race between destroying INATSConsumer objects and calling INATSConsumer::onMsg on them via callbacks in the NATS library. #106692 (Miсhael Stetsyuk).
• Fixed match, extract, extractAll and countMatches returning wrong results for regular expressions containing hex or octal escapes. #106709 (ofeliacode).
• Keeper’s internal Raft TLS now honors the openSSL.client.verificationMode setting. Previously peer certificate verification was always enabled for inter-Keeper Raft communication regardless of this setting, so none was silently ignored. Now none explicitly disables Raft peer-certificate verification, while an absent setting keeps the previous secure-by-default behavior. Configurations that explicitly set none will stop verifying Raft peer certificates after upgrade, matching the configured intent. #106726 (Antonio Andelic).
• Fix logical error in startup scripts in SYSTEM RELOAD CONFIG. Additionally, load startup scripts on SYSTEM RELOAD CONFIG (currently private-only functionality). #106727 (Miсhael Stetsyuk).
• Revert a change that made sumMap / the -Map combinator reject custom-named numeric value types (such as SimpleAggregateFunction(sum, T) and Bool) with ILLEGAL_TYPE_OF_ARGUMENT: Values for -Map cannot be summed, breaking previously-working aggregations. #106729 (Nikita Fomichev).
• Fixed multiple memory-safety and resource-exhaustion issues in format readers reachable from untrusted input: a heap out-of-bounds read in the native Parquet reader’s DataPageV2 definition/repetition level-length handling, a stack overflow on Parquet files with deeply nested schemas, and allocations that ignored max_memory_usage when parsing GeoParquet WKB/WKT geometry and Avro strings/bytes. #106739 (Raúl Marín).
• Fixed a heap buffer overflow (server crash) in decodeHTMLComponent when decoding strings containing the expanding HTML entities ≫⃒ or ≪⃒, reachable by any user with a single SELECT. #106741 (Raúl Marín).
• Reject CREATE TABLE, ALTER TABLE ADD INDEX, and CREATE INDEX with GRANULARITY 0 for skipping indexes with a clear BAD_ARGUMENTS error. Previously this triggered an Inconsistent AST formatting exception in debug builds. #106783 (Groene AI).
• Fixed Azure BACKUP/RESTORE ignoring endpoint settings for legacy-form azure_blob_storage disks. #106784 (Julia Kartseva).
• Fixed a Bad cast exception while pruning parts by MinMax statistics when a key column is LowCardinality and the predicate constant is LowCardinality(Nullable(...)). #106793 (Groene AI).
• Fix inconsistent columns (that leads to LOGICAL_ERROR later) on exception (i.e. MEMORY_LIMIT_EXCEEDED) during parsing. #106802 (Azat Khuzhin).
• Fix after configuring keeper_server.http_control.secure_port, the server returned an HTTP response when making requests to HTTPS clients. #106822 (linjiayu1025-collab).
• Fix logical error in parseDateTime with non-ASCII input bytes. #106856 (Pavel Kruglov).
• Fix SHOW CREATE ROW POLICY emitting restrictive/permissive in lowercase instead of uppercase, inconsistent with other keywords. #106865 (Valery Petrov).
• Fixes the case when parallel replicas wasn’t applied for view with UNION due to empty table in the UNION. #106900 (Igor Nikonov).
• Fix server crash (SIGSEGV) when reading truncated Protobuf data with input_format_allow_errors_num > 0. #106905 (Andrey Tsarevskiy | Андрей Царевский ).
• Fix THERE_IS_NO_COLUMN exception for distributed queries involving optimize_rewrite_aggregate_function_with_if optimization when the aggregate function argument requires a cast to Nullable type. #106908 (Yakov Olkhovskiy).
• Fix an exception (LOGICAL_ERROR) in the join runtime filter when the join key contains a Variant or Dynamic type nested inside a Tuple, Array, or Map and the right side of the join has a single distinct value. #106931 (Groene AI).
• Fixed a signed integer overflow (undefined behavior) in arrayLevenshteinDistanceWeighted and arraySimilarity when the weight arrays contain large integer values. The weighted distance is now accumulated in a wide integer for integral weights, so large integer weights no longer overflow and stay exact. #106934 (Groene AI).
• Fix a server crash (null pointer dereference) when running TRUNCATE or DROP on an EmbeddedRocksDB table whose RocksDB handle was released, for example a read_only table whose data directory was emptied by a prior TRUNCATE. #106940 (Groene AI).
• Fix an exception (std::length_error reported as a LOGICAL_ERROR) when reading from a *Cluster table function such as urlCluster with a very large max_streams_for_files_processing_in_cluster_functions setting. The number of streams is now bounded to a sane value. #106946 (Groene AI).
• Fix a server crash (null pointer dereference of DB::IConnections) in RemoteQueryExecutor when a distributed query is cancelled right before it is sent to a shard. #106950 (Groene AI).
• Fixed elapsed_us always being zero, and read_rows/read_bytes being undercounted, in system.processors_profile_log and system.query_log for asynchronous insert flush (AsyncInsertFlush) queries. #106982 (Christoph Wurm).
• Fix a crash (Source column is not Map / SIGSEGV) when merging sorted blocks that contain a Variant column with a Map variant whose local storage order differs from its global order. #107011 (Groene AI).
• Fix a logical error conflicted_part_name.has_value() raised during a synchronous insert into a ReplicatedMergeTree table when the inserted block was fully deduplicated and the conflicting part’s deduplication node had already been removed (for example by a concurrent DROP PARTITION). #107026 (Groene AI).
• Fixed an exception (logical error this->visited_views == right->visited_views) on INSERT when two materialized views on the same source table write to the same target table and a dependent view reads that target, with materialized_views_squash_parallel_inserts enabled. #107027 (Groene AI).
• Fixed a Block structure mismatch in UnionStep stream logical error (server abort on debug/sanitizer builds, Code: 49 on release builds) that occurred when one branch of a UNION/INTERSECT/EXCEPT read a Sparse-serialized column while the sibling branch read the same column as a full one (for example when pushing to a materialized view). #107041 (Groene AI).
• Fixed a LOGICAL_ERROR exception when inserting into a DeltaLake table with columns that do not match its write schema (for example a Nested column that flattens to subcolumns, or a table function with an explicit column subset). Such inserts now fail with a user-facing INCOMPATIBLE_COLUMNS error instead. Closes #87402. #107058 (Groene AI).
• Fix TYPE_MISMATCH error (“Cannot convert string … to type …”) for ORDER BY <numeric column> ... LIMIT n queries when lazy materialization placed another column before the sort column. The top-K threshold is now read from the correct sort column. #107060 (Groene AI).
• Fixed a syntax error when a FORMAT, SETTINGS, or INTO OUTFILE clause follows SHOW ROW POLICIES or SHOW MASKING POLICIES (e.g. SHOW ROW POLICIES FORMAT TabSeparated). #107061 (Groene AI).
• Fix a compound ALTER TABLE ... RENAME COLUMN a TO b, RENAME COLUMN c TO a that reuses a freed column name (a “swap”) between columns of different types. The materialized part recorded the wrong column type, so a later SELECT failed with Conversion between numeric types and IPv6 is not supported (or aborted on part load in debug builds). #107064 (Groene AI).
• Fixed non-deterministic results of the roundDown function when the boundaries array contains NaN. The same input value could return a finite boundary or NaN depending on the surrounding rows in a batch. NaN boundaries are now ignored, so the result depends only on the finite boundaries. #107065 (Groene AI).
• Fixed quantileTDigest and quantileTDigestWeighted throwing DECIMAL_OVERFLOW for Date and DateTime arguments when the interpolated quantile is fractional but in range (for example quantileTDigestWeighted(date, weight) on values that all fit in the type). The fractional result is now truncated to the result type, matching quantilesTDigestWeighted; genuine out-of-range values still raise an error. Closes: #106722. #107066 (Groene AI).
• Fix trimLeft, trimRight, and trimBoth (and aliases ltrim, rtrim, trim) throwing TOO_LARGE_STRING_SIZE when the custom trim character set is longer than 16 characters. Trim sets of any length are now supported again. #107071 (Nikita Fomichev).
• Fixed silent truncation of out-of-range integer values in Enum8/Enum16 type definitions. Enum8('a' = 200) now throws ARGUMENT_OUT_OF_BOUND instead of silently creating Enum8('a' = -56). #107081 (Groene AI).
• Fix wrong row order (and a LOGICAL_ERROR “Rows are not sorted with permutation” in debug builds) for multi-column ORDER BY ... LIMIT queries that sort by Nullable columns when several rows tie on the leading columns. #107094 (Groene AI).
• Fix a LOGICAL_ERROR (Expected the argument N to have X rows, but it has Y) when executing a function over a Dynamic column built by a JOIN over Dynamic (for example the non-joined rows of a RIGHT/FULL JOIN, or a correlated EXISTS subquery decorrelated into a join). #107095 (Groene AI).
• Fix spurious ZooKeeper session recreation on config reload. #107096 (Azat Khuzhin).
• Do not hold mutex across ZooKeeper reads in access entities refresh. #107097 (Azat Khuzhin).
• Fix a Logical error: Too large size passed to allocator exception on INSERT into a MergeTree table when adaptive_write_buffer_initial_size is set to an extremely large value. The adaptive write buffer initial size is now clamped to the buffer maximum. #107104 (Groene AI).
• Fix LOGICAL ERROR (Bad cast from type DB::ColumnString to DB::ColumnLowCardinality) when a Variant constant containing a LowCardinality member is compared to a key column whose key expression is a non-monotonic deterministic function (for example a minmax skip index over sipHash64(col)). #107111 (Groene AI).
• Fix a Bad cast from type DB::IColumn const* to DB::ColumnNullable const* logical error when a qualified asterisk (t.*) over a JOIN ... USING key is passed to an aggregate function and the other side of an outer JOIN has a Nullable key (with join_use_nulls = 0). #107129 (Groene AI).
• Fix ALTER TABLE ... ON CLUSTER batches that mix MODIFY SETTING/RESET SETTING with a comment change (for example MODIFY COMMENT 'x', MODIFY SETTING old_parts_lifetime = 123) being applied only on the leader replica, leaving the other replicas diverged. Also fix a positional or per-column-SETTINGS MODIFY COLUMN ... COMMENT being misclassified as a local comment-only metadata change, which left the column reorder out of the replicated metadata and could cause INCOMPATIBLE_COLUMNS on replica restart. #107142 (Groene AI).
• Fix a LOGICAL_ERROR (“Table expression … data must be initialized”) raised when a qualified asterisk matcher (for example x.*) referenced a recursive CTE by name inside its own recursive term. Such matchers now expand the recursive table’s columns, the same way a qualified column (x.a) or an unqualified matcher (*) already does in that position. #107144 (Groene AI).
• Fixed wrong query results caused by the query condition cache when on-fly mutations (apply_mutations_on_fly) or patch parts filtered rows before PREWHERE. A query reading with apply_mutations_on_fly = 1 could poison the cache so that a later query with apply_mutations_on_fly = 0 and the same predicate skipped marks it should have read and returned too few rows. The same fix also covers row-level security policies, which are prepended as a filter ahead of PREWHERE: a query run under a restrictive row policy could poison the cache for a later query that uses the same predicate without that policy. #107145 (Groene AI).
• Fixed BACKUP to AzureBlobStorage: copying a data file inside a backup wrote the destination object outside the backup directory. Backup object existence checks on S3 destinations now use exact HeadObject requests instead of prefix listing, preventing false matches of similarly-prefixed keys. #107153 (Pablo Marcos).
• Fixed a signed integer overflow in quantileExactExclusive, quantilesExactExclusive, quantileExactInclusive and quantilesExactInclusive that could produce a wrong result for Int64 inputs spanning a large range. #107154 (Groene AI).
• Fixed a LOGICAL_ERROR (“Unexpected exception in refresh scheduling”) that could put the server into a crash-loop on restart when a refreshable materialized view has a REFRESH ... DEPENDS ON <name> dependency whose unqualified name matches a temporary table or a CTE name. #107156 (Groene AI).
• Make sure we don’t run startup scripts on config reload, as that is semantically wrong and unexpected by users - and has been known to be error-prone. #107187 (Miсhael Stetsyuk).
• Cap the max value for queue_size for pre-reserve. #107205 (Elian Gidoni).
• Fix the Argument ... of GROUPING function is not a part of GROUP BY clause error for queries that use the grouping function with the group_by_use_nulls setting enabled. #107206 (Nikolai Kochetov).
• Fix incorrect result order for ORDER BY over UNION ALL with optimize_read_in_order enabled when the union pipeline was narrowed due to max_streams_for_union_step settings; narrowing is now skipped when the plan relies on sorted UNION output streams. Closes #106880. #107208 (Vladimir Cherkasov).
• Fixed a possible null pointer dereference while resolving proxy configuration during late server shutdown. #107231 (Pedro Ferreira).
• Fix lightweight UPDATE queries with legacy parallel replicas enabled for non-replicated MergeTree tables. #107246 (Groene AI).
• Fixed a server abort (std::out_of_range logical error) when inserting into an Iceberg table whose write block column names do not match the field ids of the latest schema (for example after a concurrent writer renames a column within the iceberg_metadata_staleness_ms window). The insert now fails with a clean query error instead of crashing the server. #107279 (Groene AI).
• Fix the stuck shutdown bug on server and local due to static thread pools keeping idle threads indefinitely if max_*_thread_pool_free_size > 0. #107291 (Miсhael Stetsyuk).
• Fix the s3 table function silently ignoring a lowercase positional partition_strategy (e.g. hive). #107297 (Julia Kartseva).
• Fix low_cardinality_allow_in_native_format=0 handling in parallel blocks marshalling. #107319 (Azat Khuzhin).
• Fix wrong (often empty) results from ORDER BY <col> LIMIT n when the use_skip_indexes_for_top_k optimization is active and a part with a minmax skip index on the sort column has had rows removed by a lightweight DELETE. The optimization no longer ranks the stale minmax of lightweight-deleted parts ahead of the parts that hold the live top rows. #107320 (Groene AI).
• Fixed a bug in ClickHouse Keeper where the snapshot metadata reported via last_snapshot (and zk_latest_snapshot_size in mntr) could move backwards after a stale or duplicated snapshot install, which could also let a same-index local snapshot overwrite a registered snapshot file in place while it was still being streamed to a peer or uploaded to S3. #107321 (Antonio Andelic).
• Fixed possible server stack overflow (crash) when reading a deeply nested schema or value in the MsgPack, BSON, ORC, Parquet, JSON, DeltaLake, Iceberg and Paimon formats. Such deeply nested input is now rejected with an exception. #107341 (Raúl Marín).
• Fix a possible logical error in SYSTEM SYNC DATABASE REPLICA ... STRICT and make the STRICT modifier actually take effect for database replicas. #107344 (Pedro Ferreira).
• Fixes a server abort in debug/sanitizer builds when reading a DeltaLake table whose ClickHouse schema names a column that is absent from the Delta column mapping, for example after a column was renamed or dropped in the Delta log. The query now fails with a catchable INCORRECT_DATA error instead. #107347 (Groene AI).
• Fix ORDER BY ... WITH FILL producing extra rows when an ORDER BY column before the fill column uses a COLLATE collation. The rows are now grouped by the sorting prefix using that collation, matching the sort order. #107365 (Groene AI).
• Fixed a crash (LOGICAL_ERROR in debug builds) and a silent wrong-results bug (in release builds) when reading an Iceberg table whose metadata re-binds an existing schema-id to a different schema across metadata versions. Such metadata is now rejected with ICEBERG_SPECIFICATION_VIOLATION. #107370 (Groene AI).
• Fix a LOGICAL_ERROR (Variant N (T) has size X, but expected Y) when a function such as toString or concat is applied to a Variant or Dynamic column that holds a single non-empty variant together with NULLs, and the function returns its input column unchanged. #107374 (Groene AI).
• Fix Logical error: 'Duplicate announcement received for replica number N' that could occur with parallel replicas when a scalar subquery contained nested subqueries reading the same table. #107381 (Groene AI).
• Fix getServerSetting to return the live effective value for runtime-changeable server settings (such as max_server_memory_usage, mark_cache_size, max_concurrent_queries, thread pool sizes, etc.), matching what system.server_settings reports. #107388 (Alexey Milovidov).
• Fix arrayResize with a Decimal size argument: the size is now interpreted by its real value (e.g. arrayResize([1, 2, 3], 1.5::Decimal(2, 1)) returns one element) instead of the raw unscaled representation. #107389 (Alexey Milovidov).
• Fix a LOGICAL_ERROR (block.rows() == getRows()) raised on an async INSERT into an Alias table when use_strict_insert_block_limits was enabled. #107400 (Groene AI).
• Fix a logical error (Unexpected return type from equals. Expected Nullable(UInt8). Got UInt8) when the disjunction (partial predicate) push-down optimization pushes a condition over a USING key whose type is widened by the JOIN. Also fix a server crash (segmentation fault) in the analyzer when resolving identifiers in JOIN ... USING queries that contain constant-foldable if/multiIf branches referencing unknown identifiers. #107407 (Groene AI).
• Fixed a heap buffer overflow (server crash) in windowFunnel when finalizing a crafted aggregate-function state with an out-of-range event type, reachable by any user with a single SELECT. #107412 (uwezkhan).
• Fixed undefined behavior when a non-finite floating-point value (such as nan or inf) is passed as the timestamp/duration argument of the prometheusQuery / prometheusQueryRange table functions. Such an argument now raises BAD_ARGUMENTS instead of producing a garbage timestamp. #107417 (Groene AI).
• Fix MaterializedPostgreSQL silently stopping replication of changes when the PostgreSQL database or table name contains upper-case letters (the pgoutput consumer requested an unquoted, lower-cased publication name that did not match the case-preserving publication). #107423 (Alexey Milovidov).
• Fix an exception (Logical error: Not-ready Set is passed as the second argument for function 'in') when a key expression (ORDER BY, PRIMARY KEY, PARTITION BY, or a skip INDEX) contained an IN operator with a table on the right-hand side, e.g. ORDER BY (x IN some_table). Such key expressions are now rejected at table creation time. #107424 (Groene AI).
• Fix incorrect results from the optimize_rewrite_aggregate_function_with_if optimization for aggregate functions that preserve NULL payload values (the *_respect_nulls family: anyRespectNulls, first_value_respect_nulls, anyLast_respect_nulls, last_value_respect_nulls). The optimization no longer rewrites f(if(cond, x, NULL)) into the -If form for such functions. #107430 (Groene AI).
• Fix a spurious filesystem error: in last_write_time: No such file or directory exception when listing a local-disk object storage directory (e.g. an Iceberg table on a local disk) while files are being concurrently replaced. A concurrently removed entry is now omitted from the listing instead of aborting it. #107432 (Groene AI).
• Fix THERE_IS_NO_COLUMN error when optimize_if_transform_strings_to_enum = 1 and the optimized if/transform-over-string-literals expression is a GROUP BY or ORDER BY key over a Distributed table or parallel replicas. #107455 (Groene AI).
• Fixed a rare server crash in DISTINCT processing that could occur when an allocation failed (for example, when hitting a memory limit) while the set of distinct keys was being initialized. #107467 (Groene AI).
• Fix LOGICAL_ERROR: Unexpected return type from materialize (and similar type-mismatch errors) when apply_mutations_on_fly = 1 is used on a table with a pending on-fly UPDATE whose target column is also read as a function input by an earlier on-fly UPDATE, before an ALTER MODIFY COLUMN ... LowCardinality(...) mutation. #107475 (Groene AI).
• Fixed stale AWS STS credentials when reading DeltaLake tables over S3: the engine now keeps the freshly-credentialed S3 client when refreshing its snapshot, so long-running reads no longer fail after the STS token’s TTL expires. The STS assume-role credentials provider now also honors credential refresh for any S3 access using STS. #107480 (Elmi Ahmadov).
• Fixed SELECT ... FINAL and OPTIMIZE TABLE ... FINAL returning duplicate rows after CREATE TABLE ... CLONE AS, ATTACH PARTITION ... FROM or MOVE PARTITION ... TO TABLE adopted parts from a plain MergeTree into a ReplacingMergeTree, SummingMergeTree or AggregatingMergeTree. The adopted part’s merge level is now reset to 0 when the source and destination engines differ, so the destination deduplicates it on the next merge. #107481 (Groene AI).
• Fix an integer underflow in the PostgreSQL wire protocol parser where a message with a length field smaller than 4 caused a size - 4 wraparound and an oversized resize/ignore. #107485 (uwezkhan).
• Query cancellation is now better tracked while waiting for the quorum in ReplicatedMergeTree. #107513 (Nikita Taranov).
• Fixed a Not-ready Set is passed as the second argument for function 'in' (LOGICAL_ERROR) when querying a table with a PARTITION BY key and an IN/NOT IN subquery wrapped inside a larger expression, for example WHERE (c0 IN (SELECT ...)) != 0. #107515 (Groene AI).
• Fixed currentUser(), user(), SESSION_USER and authenticatedUser() evaluating to an empty string on the asynchronous insert flush path (with async_insert = 1). This affected DEFAULT/MATERIALIZED column expressions and materialized views that reference these functions, which silently stored an empty string instead of the inserting user. #107541 (Groene AI).
• With enable_analyzer = 1 (the default), querying a table by its bare name when it only exists in another database now suggests the right table, e.g. SELECT * FROM functions reports Maybe you meant system.functions?. Previously the new analyzer gave a hint-less Unknown table expression identifier error, while the old analyzer already produced the helpful suggestion. #107550 (Groene AI).
• Account memory used by the rapidjson library (in prettyPrintJSON, JSONMergePatch and the rapidjson JSON parser) against the memory tracker, so pathological inputs are rejected with MEMORY_LIMIT_EXCEEDED instead of allocating without bound. #107555 (Raúl Marín).
• Fix a LOGICAL_ERROR (updateFormatPrewhereInfo called more than once) raised when querying a file(), url(), or object-storage source with both an explicit PREWHERE and a WHERE while optimize_prewhere_after_pushdown was enabled. #107568 (Groene AI).
• Fix a crash (null pointer dereference) that could happen when a distributed query plan was executed locally (make_distributed_plan + distributed_plan_execute_locally) with log_formatted_queries = 1. #107570 (Groene AI).
• Fixed a server abort (std::terminate, signal 6) during teardown of a distributed-plan query (make_distributed_plan = 1) when a worker status-check failed to re-schedule the next check (for example CANNOT_SCHEDULE_TASK on shutdown or MEMORY_LIMIT_EXCEEDED). The query now fails cleanly and the server keeps running. #107575 (Groene AI).
• Added missing bounds checks in Elf and DWARF parsing. #107579 (Michael Kolupaev).
• Added missing bounds checks in ORC reader. #107580 (Michael Kolupaev).
• Fixed an exception (Digest does not match logical error) that could happen on RENAME TABLE, RENAME DATABASE, or CREATE OR REPLACE TABLE involving a TimeSeries table inside a Replicated database. Renaming a TimeSeries table is now supported. #107583 (Groene AI).
• Fix an unauthenticated memory-exhaustion denial of service on the MySQL protocol port. #107599 (Shaohua Wang).
• Fix DROP TABLE of a TimeSeries table in a Replicated database, which previously leaked the inner tables and left the background drop task retrying forever (DROP TABLE ... SYNC would hang). #107604 (Groene AI).
• Fixed two path-traversal issues in the replicated part fetch protocol that could let a malicious replica write files outside the part directory. #107606 (Antonio Andelic).
• Fixed ‘Account must be specified error’ when reading a Delta Lake table over Azure. #107620 (Smita Kulkarni).
• Fixed ALTER TABLE ... REPLACE PARTITION on a plain MergeTree table resurrecting the replaced-out parts after a server restart, which made the table return both the replacement rows and the stale replaced rows. #107623 (Groene AI).
• Fix a server crash when reading a materialized view whose target is a Distributed table while the query runs with enable_analyzer = 0. #107653 (Groene AI).
• When several quotas are assigned to the same user or role, all of them are now enforced together (a query is rejected if any of them is exceeded), instead of only one quota being enforced and chosen non-deterministically. SHOW QUOTA and system.quota_usage now show all quotas enforced for the current user. #107664 (Alexey Milovidov).
• Fixed s3 and other object storage table functions throwing LOGICAL_ERROR instead of BAD_ARGUMENTS when a key-value argument is duplicated, e.g. s3('http://...', format = 'CSV', format = 'TSV'). #107670 (Groene AI).
• Fixed the MySQL interface being unusable with MySQL Connector/J 8.2.0 and newer (including 9.x). The info field of the OK packet is now length-encoded, matching the MySQL server, so the JDBC driver can connect. #107693 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR (“Input nodes size mismatch in dag”) when a query with make_distributed_plan = 1 joins on a function-wrapped key whose two sides have no common type (for example ON intDiv(-1, t1.key) = t2.key with a UInt64 right key). #107701 (Groene AI).
• Reading Arrow/ArrowStream data with empty String/Binary columns produced by Apache Arrow Java < 19.0.0 (including Apache Spark) no longer throws INCORRECT_DATA. #107764 (Raúl Marín).
• Fixed a BAD_GET exception (“Bad get: has String, requested UInt64”) from countmin column statistics estimation when a query compares a column against a literal of a different type that is not pre-coerced, such as numeric_col IN (SELECT '5') or string_col IN (SELECT 5). #107793 (Groene AI).
• Fixed the odbc and jdbc table functions hanging for minutes and ignoring query cancellation (KILL QUERY, max_execution_time) when the bridge becomes unresponsive while inferring the remote table structure. #107809 (Alexey Milovidov).
• Fixed an exception (Logical error: 'index >= result.start') when formatting a malformed query that mixes the positional and named secret-argument forms of the s3/gcs table functions, e.g. s3('url', 'a', 'b', secret_access_key = 'c'). #107818 (Groene AI).
• Fix set skip index not pruning granules over LowCardinality columns. #107868 (Konstantin Bogdanov).
• Fixed insert deduplication computing wrong hashes for String and Array columns with the server setting insert_deduplication_version = new_unified_hash: identical inserts could fail to deduplicate because the deduplication hash depended on the row’s position within the inserted block. #107915 (Sema Checherinda).
• Fix incorrect results of ORDER BY over Nullable columns with multiple sort keys on macOS (Apple Silicon), caused by a missing sign-extension in the JIT-compiled sort comparator. #107973 (Raúl Marín).
• Fix performance regression when reading Dynamic columns with multiple threads. Caused by #100730. Closes #107942. #107997 (Pavel Kruglov).
• The deprecated data lake setting storage_catalog_url is now correctly rejected by the catalog guard (previously only storage_catalog_type and storage_aws_access_key_id were checked), and the error message lists all deprecated settings. #108040 (Alexey Milovidov).
• Fix a Bad cast from type DB::ColumnSparse to DB::ColumnVector<char8_t> logical error when a LIKE query reads from a text index via the direct-read fallback path over a column stored sparse. #108068 (Groene AI).
• Always run access cache batch-finished handlers, even on empty batch. #108124 (Azat Khuzhin).
• Fixed a LOGICAL_ERROR (Column identifier is already registered) for some queries with UNION ALL. #100770 (Shaohua Wang).
• Fixed a use-after-free of the workload storage mutex during server shutdown. #101447 (Tuan Pham Anh).
• Fixed a deadlock in ALTER DATABASE MODIFY COMMENT with the Shared catalog. #103683 (Nikolay Degterinsky).
• An empty Unicode-quoted identifier now raises a SYNTAX_ERROR instead of CANNOT_PARSE_QUOTED_STRING. #104173 (leonard9893).
• Preserve the original parameter types of timeSeries* aggregate functions so that AggregateFunction types round-trip correctly across table reattach and parallel replicas. #104812 (Vitaly Baranov).
• Fixed a LOGICAL_ERROR for a materialized CTE under serialize_query_plan with parallel replicas. #104896 (Igor Nikonov).
• Fixed signed integer overflow in the timeSeries*ToGrid functions when the staleness (window) parameter is near INT64_MAX. #105319 (Groene AI).
• A compatibility setting value no longer reverts apply_row_policy_after_final to false, so row policies are always applied correctly with FINAL. #105637 (Yarik Briukhovetskyi).
• Fixed clickhouse chdig client by honoring SETPGROUP/RESETIDS/SETSIGDEF in the posix_spawn stub. #105858 (Azat Khuzhin).
• Fixed an exception during merge-tree sanity checks on projections when the merge/mutate executor was not initialized. #105919 (Mikhail Artemenko).
• During index analysis on projections, the minmax index is now loaded from the projection itself rather than from the parent part, producing correct results. #105940 (Mikhail Artemenko).
• Fixed an exception in addMonotonicChain for materialize(monotonic_chain(...)). #106050 (Nikolai Kochetov).
• Fixed RESTORE failing on backups that include MASKING POLICY dependents. #106086 (Julia Kartseva).
• Fixed possible memory corruption when expanding SQL user-defined functions with prefer_column_name_to_alias enabled. #106121 (Nikolai Kochetov).
• A dictionary using the bridge connection is now reloaded after a server restart, re-establishing the connection. Closes #106151. #106152 (Pedro Ferreira).
• Fixed the filesystem cache for LocalObjectStorage. #106239 (Kseniia Sumarokova).
• Fixed the lazy remote source for table functions with use_delayed_remote_source. #106470 (Nikolay Degterinsky).
• Fixed timeSeriesLastToGrid for timestamps before the grid start and out-of-window timestamps. #106504 (Vitaly Baranov). #106577 (Vitaly Baranov).
• Fixed a LOGICAL_ERROR (Inconsistent AST formatting) on a function name containing query parameters. #106635 (Vitaly Baranov).
• Fixed a logical error when a table is dropped before distributed plan task deserialization. #106944 (Alexander Gololobov).
• Use exact comparison for Nullable GROUP BY keys. #107050 (Nikolai Kochetov).
• Fixed a LOGICAL_ERROR (Trying to get name of not a column) raised when a table-function argument was not a column expression (for example an unresolved * matcher from multiIf/CASE). #107411 (Alexey Milovidov).
• Fixed unbounded growth of the time-zone cache (DateLUT) when many distinct invalid time-zone names are processed. #107464 (Alexey Milovidov).
• Write the part_log entry before reporting a plain MergeTree mutation as done. #107741 (Azat Khuzhin).
• arrayFold now sanity-checks its array argument against malformed input. #107932 (Michael Kolupaev).
• Fixed a rare data race and possible use-after-free on the cached skip-index archive reader of a MergeTree part, which could happen when a query read skip indices while the part was being moved or renamed. #107995 (Raúl Marín).
• Fix performance regression for Map subcolumns with PREWHERE. #107988 (Pavel Kruglov).
• Fix parseDateTimeBestEffort with a timezone throwing CANNOT_PARSE_DATETIME on NULL rows of toString(Nullable(DateTime64)). #108310 (Yarik Briukhovetskyi).
• Fixed the ArrowFlight table function and engine rejecting a named collection that omits the optional dataset key with a No such key 'dataset' error. #108041 (Alexey Milovidov).
• Fixed an Inconsistent AST formatting logical error for a no-argument window function inside a CODEC or engine declaration (e.g. CODEC(cume_dist() OVER (...))); such a function now keeps its parentheses so the query survives a format/parse round-trip. #107806 (Alexey Milovidov).
• Fix hasToken with a separator-containing needle silently returning results via a text index instead of raising BAD_ARGUMENTS. #108189 (Jimmy Aguilar Mena).
• The setting use_skip_indexes_on_data_read can now be reverted to its pre-26.1 default (false) via the compatibility setting, providing an escape hatch for a performance regression where the on-data-read path defeats minmax/set/bloom_filter skip-index mark-range pruning. #108330 (egor romanov).
• Fixed a possible crash (null pointer dereference) when the database/db override of a named collection passed to remote/remoteSecure is not a constant database name, e.g. remote(nc, database = (SELECT 1)). The query now fails with a clear error instead of crashing. #108271 (Groene AI).
• Fixed refreshable materialized view getting stuck if the ZooKeeper connection is lost at the wrong moment. #108234 (Michael Kolupaev).
• Fix Bad cast from type DB::ColumnVector<...> to DB::ColumnTuple when reading an Array(Tuple(...)) column whose value is filled with defaults, e.g. after ALTER TABLE ... ADD COLUMN or after an unfinished ALTER TABLE ... CLEAR COLUMN mutation applied on the fly. #107232 (Groene AI).
• Fix a Bad cast from type DB::ColumnDynamic to DB::ColumnNullable logical error when an explicit reference to a JOIN ... USING key whose common supertype is Dynamic is passed to an aggregate function, for example count(t.key) IGNORE NULLS. #107289 (Groene AI).
• Fix silent data loss on plain (non-replicated) MergeTree when REPLACE PARTITION, MOVE PARTITION, DETACH PARTITION, or DETACH PART is run on a partition that still has unapplied lightweight UPDATE patches. These operations now reject the command, as ReplicatedMergeTree already does. #107386 (Groene AI).
• Fix a crash (SIGSEGV in release builds, type-mismatch assertion in debug builds) in has over a Map with a Dynamic key when the lookup argument is LowCardinality, e.g. has(map('a'::Dynamic, ...), toLowCardinality('b')). #107956 (Groene AI).
• Fixed a LOGICAL_ERROR (“Block structure mismatch … between ConvertingTransform and RemovingReplicatedColumnsTransform”) when inserting into a materialized view whose TO target table declares a column with a wider Enum than the view’s SELECT produces. The valid Enum widening is now applied. #107648 (Groene AI).
• Fix NUMBER_OF_COLUMNS_DOESNT_MATCH error when querying a Distributed table (or using parallel replicas) that has several ALIAS columns expanding to the same expression and referencing them together with ORDER BY/GROUP BY/HAVING. #107913 (Yakov Olkhovskiy).
• Fixed undefined behavior (null pointer passed to memcpy) in the detectCharset and detectLanguageUnknown functions when the input string is larger than 32768 bytes and no character set can be detected. #108250 (Groene AI).
• Fix signed integer overflow (undefined behavior) in dateDiff with hour and minute units on extreme DateTime64 values close to the Int64 range limits. #108229 (Groene AI).
• Fix Too many marks for a text index on an empty merged part. #106867 (Azat Khuzhin).
• Fix a LOGICAL_ERROR (“Unexpected return type from if”) when reading a column under apply_mutations_on_fly = 1 after an ALTER UPDATE col = ... WHERE <cond> with a non-constant or false condition followed by ALTER MODIFY COLUMN col <new type>. #108128 (Groene AI).
• Fix a server abort (Logical error in IColumn::insertFrom) when casting an Array(Dynamic) or Array(Variant) to QBit with accurateCastOrNull, e.g. accurateCastOrNull(CAST(range(114), 'Array(Dynamic)'), 'QBit(Float32, 114)'). #108288 (Groene AI).
• Fixed a syntax error when an alias follows a subquery in DESCRIBE TABLE (...) AS .... #100205 (Yarik Briukhovetskyi).
• getClientHTTPHeader is now correctly treated as non-deterministic, so its result is no longer incorrectly reused by the query result cache. #108029 (Alexey Milovidov).

​

Build/Testing/Packaging Improvement

• Build delta-kernel-rs with the default-engine-native-tls feature so that its own HTTP client (reqwest) can reuse the OpenSSL that ClickHouse already links in. This is only a partial native-tls enablement: object_store still forces reqwest/rustls-tls-native-roots, so reqwest ends up with both the native-tls and rustls backends. delta-kernel-rs remains disabled under MSan for now because ring is still compiled in via object_store (both transitively through rustls and directly for AWS request HMAC signing), and its hand-written assembly does not generate the symbols MSan requires (tracked upstream at arrow-rs-object-store#585). #96856 (Austin Bonander).
• Add PGO (Profile-Guided Optimization) and BOLT (Binary Optimization and Layout Tool) post-link optimization for the clickhouse binary. Profiles are collected from CI workloads and applied as a best-effort step during release builds — both stages fall back to non-optimized output if a profile is stale or incompatible. As of today, there are no benefits from PGO. #100938 (Alexey Milovidov).
• Add --storage mode to clickhouse keeper-bench that benchmarks KeeperStorage in-process (no network, no raft, no state machine), using the same setup/generator config sections as the network mode. #103081 (Michael Kolupaev).
• Fix build with ENABLE_AWS_S3=OFF. #105390 (Yue Ni).
• Update mongo-cxx-driver to r4.3.0. #105522 (Konstantin Bogdanov).
• Fix build against RapidJSON. #105674 (Ilya Golshtein).
• Speed up the build by removing transitive includes from widely-used base/ headers and stripping unused code from the vendored Poco/Logger.h. #105702 (Raúl Marín).
• Updated libxml2 from 2.15.1 to 2.15.3. #105985 (Konstantin Bogdanov).
• Use expat 2.8.1 inside Poco. #105988 (Konstantin Bogdanov).
• Bump thrift to v0.23.0. #105993 (Konstantin Bogdanov).
• Use postgres tag REL_18_4. #105994 (Konstantin Bogdanov).
• Upgrade krb5 to 1.22.2. #106076 (Konstantin Bogdanov).
• Update bundled curl to 8.20.0. #106077 (Konstantin Bogdanov).
• Bumped mariadb-connector-c to 3.1.29. All the ClickHouse-relevant patches preserved. #106287 (Nikita Mikhaylov).
• Bump distroless base image digests for libssl3t64 CVE fixes. #106360 (Rahul Nair).
• Use wasmtime v45.0.1. #106836 (Konstantin Bogdanov).
• Bump openssl to 3.5.7. #106844 (Konstantin Bogdanov).
• Distroless Docker build now also updates the :X.Y-distroless and :X.Y.Z-distroless floating tags, not just the full-version tag. #106932 (Rahul Nair).
• Disable unused curl features, such as NTLM auth, cookies, alt-svc, HSTS, DNS-over-HTTPS, netrc, MIME, AWS SigV4. #107226 (Konstantin Bogdanov).
• Updated NuRaft with a fix for a snapshot-install livelock that occurred when snapshot IO ran on a background thread. The nuraft_use_bg_thread_for_snapshot_io Keeper setting remains disabled by default; CI now randomizes it to cover both modes. #107338 (Antonio Andelic).
• Fix building abseil when the ClickHouse checkout path contains a substring matching a header extension. #107349 (zhiqiang).
• Determine the L2 cache size from CPUID on x86_64 instead of the glibc-specific sysconf(_SC_LEVEL2_CACHE_SIZE), which is not available under musl libc. #107469 (Konstantin Bogdanov).
• Increase the stack size for musl builds to 8 MiB to support deeply-nested queries. #107470 (Konstantin Bogdanov).
• Derive the set of C library symbols localized in Rust static libraries from the actual reference libraries instead of a hand-maintained allowlist, so all compiler-rt builtins and platform libm functions are covered. #107571 (Raúl Marín).
• Switch the distroless server and keeper images to gcr.io/distroless/base-nossl-debian13 so they only ship the libraries ClickHouse actually links against and stop pulling in ones we statically link ourselves. #107837 (Rahul Nair).
• Re-enable 50 stateless tests on macOS (arm_darwin) that were stale-skipped, and use fsync instead of F_FULLFSYNC for directory syncing on macOS. #107887 (Raúl Marín).

​

ClickHouse release 26.5, 2026-05-21. Presentation, Video

​

Backward Incompatible Change

• Changed defaults of date_time_input_format and cast_string_to_date_time_mode from basic to best_effort. Queries that previously failed to parse non-basic datetime strings (e.g. 2024 April 4, Apr 15, 2020 10:30:00) may now succeed by default. To keep the old strict parsing behavior, set these settings to basic (or use compatibility). #89334 (Alexey Milovidov).
• Tuple element name null is now forbidden because it conflicts with the subcolumn name used for Nullable null maps, causing ambiguous subcolumn resolution. #98377 (Alexey Milovidov).
• Add a setting dynamic_disk_allow_from_env, dynamic_disk_allow_from_zk, dynamic_disk_allow_include to disallow usage of from_env, from_zk, include in dynamic disks. Backward Incompatible Change because prohibits by default the behaviour which used to be allowed by default. #99138 (Kseniia Sumarokova).
• You can no longer use the obsolete Arrow-based Parquet reader and writer. The native implementation will be used instead. #100949 (Alexey Milovidov).
• SHOW CREATE TABLE t now prefers the temporary table when both a permanent and a temporary table named t exist and no database is specified, matching the existing behavior of DESCRIBE TABLE. Additionally, DESCRIBE TEMPORARY TABLE syntax is now supported. #100966 (Alexey Milovidov).
• Reduced default http_max_fields from 1,000,000 to 1,000 and http_max_field_name_size from 128 KB to 4 KB to limit pre-authentication memory usage by HTTP connections. Added http_max_request_header_size and http_headers_read_timeout settings. Users who rely on the previous higher limits can restore them via settings. #103285 (Sema Checherinda).
• Added a histograms nested column to system.metric_log that snapshots every registered histogram metric per row, with a new system_metric_log_show_zero_values_in_histograms setting to control zero-value emission. Deprecates the system.histogram_metric_log table. #103770 (Miсhael Stetsyuk).
• CAST to DateTime or DateTime64 without an explicit time zone now preserves the time zone of its source argument (when the source is a DateTime/DateTime64 with an explicit time zone), matching the behavior of the toDateTime/toDateTime64 functions. Closes #55072. #104433 (Alexey Milovidov).
• Removed the kql table function. Use SET dialect = 'kusto' to run queries in the KQL dialect. #105101 (Alexey Milovidov).
• The window functions RANK and DENSE_RANK now reject arguments and throw NUMBER_OF_ARGUMENTS_DOESNT_MATCH, in line with the SQL standard. Previously, queries such as RANK(x) OVER (ORDER BY id) were silently accepted with the argument ignored. To restore the previous lenient behavior, set allow_rank_dense_rank_arguments = 1. Closes #49526. #104324 (Groene AI).

​

New Feature

• Added a new setting max_bytes_ratio_before_external_join, mirroring max_bytes_ratio_before_external_group_by and max_bytes_ratio_before_external_sort. It expresses the spill-to-disk threshold for hash joins as a fraction of available memory; combined with the absolute max_bytes_before_external_join, the smaller resulting threshold applies. #103862 (Alexey Milovidov). The new max_bytes_ratio_before_external_join setting is now enabled by default at 0.5, mirroring max_bytes_ratio_before_external_group_by and max_bytes_ratio_before_external_sort. Hash joins automatically spill to grace hash join once the right-side data exceeds half of the available system memory (when memory limits are configured). #104285 (Alexey Milovidov).
• Add table function filesystem. It allows to represent directory structure as a table, to query files’ metadata and contents with SQL. Originally #42039 by @perst20. See #42039. See #50208. #53610 (Alexey Milovidov).
• Allow passing bare function names to higher-order functions like arrayMap, arrayFilter, etc. For example, arrayMap(negate, [1, 2, 3]) is now equivalent to arrayMap(x -> negate(x), [1, 2, 3]). #101033 (Alexey Milovidov).
• Add setting send_table_structure_on_insert_with_inline_data and --inline-insert-data client option to allow the server to parse inline INSERT data itself over the native protocol, avoiding the round-trip to receive table structure and improving performance for many small inserts. #101034 (Alexey Milovidov).
• Add tokenizeQuery and highlightQuery functions for SQL query tokenization and syntax highlighting. tokenizeQuery returns lexer tokens with byte offsets and token types; highlightQuery returns parser-based syntax highlighting ranges with highlight categories (keyword, identifier, function, number, string, etc.). #101054 (Alexey Milovidov).
• Add url_base setting to resolve relative URLs in the url table function and URL table engine, following RFC 3986 semantics. #101113 (Alexey Milovidov).
• Support negative values in the LIMIT BY clause to select rows from the end of each group instead of the beginning. For example, LIMIT -2 BY id returns the last two rows per id. Negative offsets (LIMIT -1 OFFSET -1 BY id) and mixed signs (LIMIT -2 OFFSET 1 BY id) are supported as well. #103222 (Nihal Z. Miaji).
• Support json_value function to output with tuple and array to improve performance of multiple json query. #78362 (kevinyhzou). Support multi-path Tuple/Array JSONPath argument in JSON_VALUE, JSON_EXISTS, and JSON_QUERY, by @KevinyhZou. #101102 (Alexey Milovidov).
• A new kafka_autodetect_client_rack parameter is introduced. If set, discover the Availability Zone via cloud facilities and propagate it as the client.rack parameter of librdkafka to avoid cross-zone communications. #81323 (Ilya Golshtein).
• Add Read event type to system.blob_storage_log for tracking object storage read operations, controlled by new setting enable_blob_storage_log_for_read_operations. #96867 (Alexey Milovidov).
• Users can now see ZooKeeper watches issued by clickhouse-server using the new system.zookeeper_watches table. #99277 (Den Kalantaevskii).
• Added Shards profile event that counts the number of shards involved in distributed queries, summed across all tables. #99470 (Alexey Milovidov).
• WASM UDFs can be declared DETERMINISTIC and become a subject of constant folding. #100005 (Vasily Chekalkin).
• Add parallel_replicas_prefer_local_replica setting: when disabled, parallel replicas are selected purely by the load balancing algorithm, allowing even max_parallel_replicas = 1 queries to be directed to another host. #100139 (Alexey Milovidov).
• Add {disk,storage,http}_connections_rcvbuf and {disk,storage,http}_connections_sndbuf server settings to control TCP socket buffer sizes on outgoing HTTP connections, allowing operators to override kernel autotuning and cap per-connection memory usage. #100478 (Sema Checherinda).
• Support CREATE OR REPLACE MATERIALIZED VIEW with the same atomic swap semantics as CREATE OR REPLACE TABLE. Works with inner tables, TO tables, POPULATE, REFRESH, and ON CLUSTER. #100539 (DQ).
• Add s3_read_request_duration_microseconds and s3_read_request_bytes histogram metrics to observe S3 GET request connection lifetime and bytes consumed, visible in system.histogram_metrics and the Prometheus endpoint. #102058 (Sema Checherinda).
• Add Paimon, PaimonS3, PaimonAzure, PaimonHDFS, and PaimonLocal table engines with incremental read support backed by Keeper snapshot progress tracking. Incremental mode returns only new rows since the last committed snapshot. Targeted snapshot delta reads are available via paimon_target_snapshot_id, and per-query snapshot caps via max_consume_snapshots. Background metadata refresh is configurable with paimon_metadata_refresh_interval_sec. Gated by allow_experimental_paimon_storage_engine. #102343 (XiaoBinMu).
• Added a new Kafka table setting kafka_map_virtual_columns_on_write. When enabled, columns named _key, _timestamp, _headers.name and _headers.value in the Kafka table schema are produced as the corresponding Kafka message key, timestamp and headers on INSERT, and are excluded from the message payload. #103243 (Alexey Milovidov).
• Added SYSTEM PAUSE VIEW [db.]name and SYSTEM PAUSE VIEWS queries for refreshable materialized views. Unlike SYSTEM STOP VIEW, SYSTEM PAUSE VIEW does not interrupt the currently running refresh — the in-flight refresh is allowed to complete and only subsequent refreshes are prevented. Undone by SYSTEM START VIEW or SYSTEM START VIEWS, which now uniformly clear both the stopped and paused states. #103252 (Nikita Mikhaylov).
• Added function regexpPosition (with PostgreSQL-compatible aliases regexpInstr and regexp_instr) that returns the byte position of the N-th regex match in a string. Supports start offset, return-after-match mode, regex flags, and capture-group selection. #104172 (Abhinav Agarwal).
• New functions isPrime and isProbablePrime for primality checks. isPrime returns an exact result for unsigned integers up to UInt64. isProbablePrime also supports UInt128 and UInt256; for those wider types, 0 means definitely composite and 1 means probably prime. The optional second argument of isProbablePrime, rounds, controls the confidence (capped at 256); the default of 25 rounds bounds the false-positive rate for a random composite below 10^-15, and isProbablePrime is cancellable. #104234 (Nihal Z. Miaji). #104639 (Alexey Milovidov). #104806 (Nihal Z. Miaji).
• clear and /clear now clear the terminal in the clickhouse command-line tools instead of running as a mistaken query. #104318 (Tyler Hannan).
• Allow the file table function to accept an Array(String) of paths in SELECT queries. #104442 (Yue).
• Add deterministic and higher_order columns to system.functions table. #104479 (Pedro Ferreira).
• Add an optional encoding variant parameter to bech32Encode (bech32 / bech32m) and a raw decode mode to bech32Decode for non-SegWit address encoding (e.g. Cosmos SDK, Injective, Osmosis). #98986 (Yash ).
• You can now write data in AvroConfluent format, which was previously input-only. This enables producing Confluent Schema Registry-framed Avro messages directly from ClickHouse, for example when writing to Kafka. The schema is automatically registered with the registry. Use the new output_format_avro_confluent_subject setting to specify the subject name. #101935 (János Benjamin Antal).
• Added the prettyPrintJSON function to format a JSON string into a human-readable form. Useful for dashboards or reports where prettifying previously required an extra step on the client side. Closes #62523. #102594 (Dmitry Prokofyev).
• Added STRING_AGG as a case-insensitive alias of groupConcat for PostgreSQL / SQL-standard compatibility. #105125 (Alexey Milovidov).
• Individual subquery results can now be cached independently using SETTINGS use_query_cache = true on specific subqueries, without caching the entire outer query. A new setting query_cache_for_subqueries = true enables bulk propagation of use_query_cache into all subqueries. Note: use_query_cache on the outer query no longer auto-propagates to subqueries. #99804 (Vincent Voyer).
• Added lexer-based syntax highlighting to the query editor in the Web UI (play.html), modelled after clickhouse-client colors. #105105 (Alexey Milovidov).

​

Experimental Feature

• Added an experimental web terminal interface at /webterminal that provides an interactive clickhouse-client session in the browser over WebSocket. Disabled by default; enable with the allow_experimental_webterminal server setting. See it here. #100277 (Alexey Milovidov). #105191 (Alexey Milovidov). #105059 (Alexey Milovidov).
• The Kafka2 engine (experimental, with Keeper-based offset storage) now supports direct SELECT queries and the kafka_commit_on_select setting. #100276 (Alexey Milovidov).
• WASM UDFs can now coerce more numeric types: smaller integers to wider integers (e.g. Int8 to UInt64), and any integer to floating point (e.g. Int32 to Float32). #100435 (Vasily Chekalkin).
• Support LIKE function in DELETE FROM system.webassembly_modules query. #104397 (Vladimir Cherkasov).
• Support geo types for Iceberg. #103113 (Konstantin Vedernikov).
• Add prepared statements functionality to ArrowFlight SQL server. #103047 (Yakov Olkhovskiy).
• Improved KQL parser (for the Kusto language support) robustness by preserving parser depth and backtracks counters across KQL parser stages, so parser limits are tracked consistently for complex KQL queries. #103528 (Yakov Olkhovskiy).

​

Performance Improvement

• Reuse the Parquet footer metadata cache when reading local Parquet files via the file table function or File table engine. Previously the cache was only consulted for object-storage backends. #104260 (Alexey Milovidov).
• Push ORDER BY ... LIMIT n through LEFT/RIGHT joins when the sort key only references columns from the side preserved by the join, restricting how many rows the preserved-side input must produce before joining. Controlled by the new setting query_plan_top_k_through_join (default enabled). #104268 (Alexey Milovidov).
• Enable use_top_k_dynamic_filtering and use_skip_indexes_for_top_k settings by default to improve performance of ORDER BY ... LIMIT N queries. #99537 (Alexey Milovidov). Restrict use_top_k_dynamic_filtering to fixed-length sort columns by default, avoiding regressions on ORDER BY <var-length-column> LIMIT N queries where the per-row threshold comparison cost exceeds the I/O savings. The previous behavior is available via the new use_top_k_dynamic_filtering_for_variable_length_types setting. #104216 (Alexey Milovidov).
• Use the oversize-arena feature of jemalloc to reduce page faults. #103958 (Nikita Taranov).
• Limit simultaneously active streams in UNION ALL to reduce peak memory usage. #100176 (Alexey Milovidov).
• Slightly optimize the userspace page cache. #100300 (Alexey Milovidov). Now the userspace page cache is always better than the OS page cache.
• Cold reads of object storage through the userspace page cache (use_page_cache_for_object_storage = 1) are now significantly faster, because consecutive cache misses are coalesced into a single HTTP request instead of one request per page_cache_block_size block. #104230 (Alexey Milovidov).
• Speed up the index analysis involving type casting and function application. Closes #55653. #100366 (Alexey Milovidov).
• Speed up huge queries on Merge tables over a huge number of underlying tables. Closes #32465. #100369 (Alexey Milovidov).
• Remove vtable from setting field types, reducing Settings copy size by ~28x and improving cache locality via typed-array layout for all settings types. #102269 (Raúl Marín).
• Add max_threads_min_free_memory_per_thread and max_insert_threads_min_free_memory_per_thread settings to automatically reduce query parallelism when the server is low on free memory. #100383 (Alexey Milovidov).
• Reduce memory usage on low-memory systems (< 4 GiB). #100389 (Alexey Milovidov).
• Added OptimizeTrivialGroupByLimitPass. For trivial SELECT ... FROM t GROUP BY k LIMIT n queries (no HAVING, ORDER BY, or window functions), the analyzer now sets max_rows_to_group_by = n + offset with group_by_overflow_mode = 'any', so aggregation stops once n distinct keys have been produced instead of grouping the entire input. Controlled by the new setting optimize_trivial_group_by_limit_query (enabled by default). #104473 (Amos Bird) (Alexey Milovidov).
• Granule-level implicit min-max index for the _part_offset and _block_number virtual columns, including for projections. Enables fast pruning of granules based on virtual column predicates. #103952 (Mikhail Artemenko). #104746 (Mikhail Artemenko). #105137 (Mikhail Artemenko).
• For the prealloc_serialized family of aggregation methods, precompute per-row hashes during the batch-serialization pass and use them to (a) skip rehashing in emplaceKey/findKey and (b) software-prefetch the next row’s bucket. Speeds up multi-key string/serialized aggregation by hiding hash-table cache-miss latency. #104475 (Amos Bird) (Alexey Milovidov).
• Query condition cache for Iceberg tables. #102115 (Konstantin Vedernikov).
• Optimize cramersV, cramersVBiasCorrected, theilsU, and contingency when used with window functions. Closes #83521. #93384 (Nihal Z. Miaji).
• Added a query optimization that rewrites tupleElement(dictGet('dict', ('a', 'b', 'c'), key), N) into dictGet('dict', 'a', key), avoiding fetching unnecessary dictionary attributes. Controlled by the optimize_dictget_tuple_element setting (enabled by default). #100186 (Alexey Milovidov).
• Enable buffering for sorting steps on the initiator for distributed queries with sorting. #100661 (Nikita Taranov).
• Performance improvement for partial_merge JOIN. #100945 (Artem Zuikov).
• Slightly reduced memory over-allocation in partial_merge JOIN. #100963 (Artem Zuikov).
• Optimize allocations/deallocations by caching sampling settings instead of traversing the entire memory-tracker hierarchy. #101267 (Azat Khuzhin).
• Reduce memory allocation overhead during S3 multipart uploads by preallocating internal tracking containers. #101799 (Gagan Dhakrey).
• Add software prefetch in hash join probe phase to reduce memory access latency for large hash tables, controlled by setting enable_software_prefetch_in_join. #102444 (Xiaozhe Yu).
• Optimize MemoryTracker layout, improving its performance by ~25%. #103464 (Azat Khuzhin).
• Rewrite coalesce(a, b, ...) <op> const and ifNull(a, b) <op> const predicates before index analysis so per-column primary key and skip indexes on each argument can prune granules. Controlled by the new setting allow_key_condition_coalesce_rewrite (on by default). #103468 (Manuel).
• Significantly improved query performance when reading Iceberg catalogs with large JSON metadata files by optimizing escaped slash processing. #103998 (Mohaidoss).
• Avoid index uncompressed cache overhead when the cache is disabled (server setting index_uncompressed_cache_size = 0, which is the default). #104063 (Michael Kolupaev).
• Fix ORDER BY ... LIMIT queries with small limit reading excessive granules when combined with a non-selective WHERE filter. Previously, any filter above an in-order read disabled the per-mark-range task split, so the pipeline could not cancel between granules and read the whole part for each stream. #104112 (Vladimir Cherkasov).
• Fixed missing hash table sizes cache reuse after reordering JOIN sides. #104131 (Nikita Taranov).
• When fsync_after_insert is enabled, fsync of part files is now performed in parallel using the IO thread pool, speeding up insert finalization on wide tables. The full ClickBench hits load is about 22% faster. #104137 (Alexey Milovidov).
• Vectorise find_first_symbols, find_first_not_symbols, find_last_symbols_or_null, find_last_not_symbols_or_null, and splitInto on AArch64 using NEON. Previously these helpers had a SIMD path only on x86 (SSE2 / SSE4.2) and fell through to a scalar loop on ARM. Speeds up TSV parsing by ~2x, URL functions and splitByChar by ~1.3x on the ClickBench hits dataset; very dense JSON parsing (sub-16-byte field cadence) regresses slightly in line with the existing SSE2 trade-off. #104228 (Alexey Milovidov).
• Reading data through a trivial ALIAS column (e.g. some_alias['key'] where some_alias ALIAS m) now goes through the same per-subcolumn read path as referencing the underlying column directly, restoring large I/O savings for Map, Array, Tuple and Nullable aliased columns. #104245 (Raúl Marín).
• Reduced lock contention on the asynchronous remote-FS read path by making the per-query AsyncReadCounters lockless. #104374 (Nikita Mikhaylov).
• Slightly better query plan for projections in-order scan. #103723 (Mikhail Artemenko).
• Extend optimizeUseNormalProjections to also handle a specific case where nothing was filtered but the projection’s sorting key can remove the sorting step from the query plan. #104680 (Mikhail Artemenko).
• Allow json.path[] syntax sugar and explicit type hints to work correctly on typed JSON paths. #99179 (Pavel Kruglov).
• Optimize json.path[N].nested.path (which expands to tupleElement(tupleElement(json.path[N], 'nested'), 'path')) into json.path[].nested.path[N], which reads much less data. #99802 (Pavel Kruglov).
• Optimize SHOW TABLES access checks in system.parts by hoisting database-level grant checks out of the per-table loop. #100860 (Shaohua Wang).
• Remove a redundant per-table SHOW TABLES access check in the SELECT name / SELECT database, name fast path of system.tables. #100881 (Shaohua Wang).
• Enable JIT compilation on macOS. #100947 (Alexey Milovidov).
• Skip deferring the row policy after FINAL when it depends only on sorting-key columns and is deterministic, and skip the corresponding PREWHERE deferral that the row policy was forcing in that case. #102884 (Yarik Briukhovetskyi).
• Added support for hasAny and hasAll as filter predicates in text indexes. #103266 (Anton Popov).
• Route long-lived MergeTree heap state (per-part and per-table metadata) to a dedicated jemalloc arena, exposed via jemalloc.mergetree_arena.* async metrics. Reduces default-arena fragmentation in steady-state and keeps per-part survivors from pinning otherwise-decayable pages. #104136 (Raúl Marín).
• Improve ZooKeeper client timeout handling under heavy load when many requests are pipelined on a single session. The ZooKeeper client now uses a progress-based timeout: as long as any data is received from the server within session_timeout_ms, the wait is extended. A hard cap of 3 * session_timeout_ms per request still bounds caller latency. Closes #100466. #104351 (Antonio Andelic).
• Move per-event trace flags into a separate array and allocate them lazily, reducing the overhead of ProfileEvents tracing. #105030 (Azat Khuzhin).
• Add compareTrackAt for ColumnDecimal, improving comparison performance for Decimal columns. #105110 (Artem Zuikov).
• Added a SIMD implementation of MD5 (AVX2 / AVX512) that hashes multiple inputs in parallel, improving throughput. #105161 (Joanna Hulboj).

​

Improvement

• Improve the help output for all ClickHouse applications: --help now consistently returns exit code 0, prints to stdout, and a top-level clickhouse --help lists all subcommands. Adds a --no-sudo option for clickhouse start / restart (useful in Docker), and a clickhouse help subcommand. Continuation of #58244 from @qoega. #98148 (Alexey Milovidov).
• The default of input_format_column_name_matching_mode is changed from match_case to auto. Input formats that match input column names against the table schema (JSONEachRow, CSVWithNames, JSONColumns, BSONEachRow, RowBinaryWithNames, etc.) now first try a case-sensitive match and fall back to case-insensitive matching when the case-sensitive match misses. The previous strict behavior is preserved under compatibility. #104320 (Alexey Milovidov).
• Added STDDEV as a case-insensitive alias of stddevSamp for PostgreSQL/SQL-standard compatibility. #105120 (Alexey Milovidov).
• Added array_to_string as a case-insensitive alias of arrayStringConcat for PostgreSQL compatibility. #105121 (Alexey Milovidov).
• Added unnest as a case-insensitive alias of arrayJoin for PostgreSQL compatibility (function-call form). #105124 (Alexey Milovidov).
• The progress bar in clickhouse-client now shows temporary data on disk usage (e.g. for external sort, aggregation, or JOIN) next to RAM, including a per-host breakdown for distributed queries. #105190 (Alexey Milovidov).
• Added system.predicate_statistics_log, a new sampled log of predicate filter selectivity and MergeTree index-granule pruning per query. Disabled by default; enable via the predicate_statistics_sample_rate server setting. Needed for automatic index and projection recommendations. #98727 (Yarik Briukhovetskyi).
• Allow skipping local shard with missing table when skip_unavailable_shards is enabled. #100141 (Alexey Milovidov).
• Add startup warnings (visible in system.warnings) when a Linux mdraid array is being resynchronized or is in a degraded state, as both can affect disk I/O performance or indicate disk failures. #100941 (Alexey Milovidov).
• WASM UDFs now appear in system.functions with correct origin (WasmUserDefined), syntax, arguments, and returned_value columns populated from their ClickHouse type metadata. Previously they were either missing or listed with the wrong origin as duplicates. #101053 (Vasily Chekalkin).
• The generate_series table function now supports negative step values for generating descending sequences, e.g. SELECT * FROM generate_series(99, 0, -1). #101056 (Alexey Milovidov).
• Fix Context has expired exceptions raised by some functions (dotProduct, formatRow, structureToCapnProtoSchema/structureToProtobufSchema, user-defined functions) when used in deferred execution paths such as DEFAULT/MATERIALIZED expressions and table engine settings, by ensuring those functions hold a strong context where required. #101109 (Alexey Milovidov).
• Fix spurious Cache limits violated logical errors on priorities created with zero limits (such as the per-query filesystem cache priority used when enable_filesystem_query_cache_limit is on). #101428 (Alexey Milovidov).
• Auto-detect the region for s3express endpoints. #101520 (Pradeep Chhetri).
• Support field ids for data files in Iceberg writes. Closes #102322. #102362 (Konstantin Vedernikov).
• The left panel in the web UI (play.html) now scrolls independently, stays visible when the page is scrolled, and clicking a table scrolls to the query area. #102498 (Alexey Milovidov).
• Allow opening the Play UI in a new tab via Ctrl/Cmd/Shift+click or middle-click on the ClickHouse logo in the left panel. #102501 (Alexey Milovidov).
• New metrics to track memory used by projection primary keys and projection index granularity across all tables. #102587 (Narasimha Pakeer).
• Apply max_network_bandwidth_for_user and max_network_bandwidth_for_all_users to remote filesystem reads/writes. #103080 (Azat Khuzhin).
• Improve the resilience of backups when Refreshable Materialized Views are constantly refreshed by making decisions based on the snapshotted tables’ state rather than a global one. #103384 (Nikita Mikhaylov).
• SELECT * FROM system.databases now always lists data lake catalog databases regardless of the show_data_lake_catalogs_in_system_tables setting. Previously they were hidden by default, which was inconsistent with SHOW DATABASES that always showed them. #103444 (Alsu Giliazova).
• Align the default of the server setting concurrent_threads_soft_limit_ratio_to_cores in code with the shipped config.xml value (2), so the default max_min_fair concurrent threads scheduler caps query processing threads at 2x cores out of the box even when the shipped config.xml is not used. #103446 (Alexey Milovidov).
• Improved sanitizer robustness in parser and string-function edge cases by handling empty input in getURLScheme, avoiding null-pointer arithmetic in Lexer::nextToken max_query_size checks, and short-circuiting UTF-8 subsequence evaluation on empty haystacks before decoding. #103489 (Yakov Olkhovskiy).
• Improved correctness of monotonicity inference for divide(0, x) and intDiv(0, x): previously the functions were unconditionally reported as monotonic, but 0 / x is non-monotonic on any range that includes 0 because 0 / 0 is undefined (NaN/Inf for divide, division-by-zero exception for intDiv). The false claim caused KeyCondition::applyMonotonicFunctionsChainToRange to produce ranges with left > right, which tripped a LOGICAL_ERROR Invalid binary search result in MergeTreeSetIndex in debug builds when an IN / NOT IN expression on the primary key wrapped the key in divide(0, key) or intDiv(0, key). Release builds were unaffected — the assertion is gated by #ifndef NDEBUG. #103621 (Groene AI).
• Replace the HTTPConnectionPool*TCP{Rcv,Snd}BufBytes_{p50,p75,p90,p95} async metrics with a bucket-based histogram http_pool_tcp_buf_bytes (labels group, direction) in system.histogram_metrics. Per-group total async metrics are preserved. #103704 (Sema Checherinda).
• The server settings {disk,storage,http}_connections_{rcvbuf,sndbuf} are now reported as changeable without restart in system.server_settings. Their values can be updated via SYSTEM RELOAD CONFIG; the runtime application path was already in place since #100478. #103772 (Sema Checherinda).
• Add support for resolved symbols (Array(String)) in flameGraph. #103816 (Azat Khuzhin).
• CLI client can now specify <rainbow_parentheses>false</rainbow_parentheses> in its config in environments where terminal colors clash with the parentheses (similarly, clickhouse format now has --no_rainbow_parentheses). #103851 (Larry Snizek).
• Fix MultiVolnitsky UTF-8 case-insensitive search by rolling back partial putNGram inserts on failure and switching failed needles to fallback searchers, avoiding inconsistent state. #103864 (Yakov Olkhovskiy).
• Implement sched_getcpu via rseq TLS in glibc-compatibility musl. #104016 (Azat Khuzhin).
• Replicated refreshable materialized view in APPEND mode will no longer refresh twice if connection to zookeeper was briefly lost during refresh. #104051 (Michael Kolupaev).
• Allow Distributed tables without an explicit column list to validate sharding keys against the inferred remote table structure. #104111 (Yue).
• Fix REST catalog with Azure abfss path. #104120 (Konstantin Vedernikov).
• Honor role_arn and role_session_name auth settings in the Keeper S3 snapshot client, allowing snapshot uploads to use STS AssumeRole-based authentication. #104140 (Alexey Milovidov).
• Fix a memory leak when running queries with MATERIALIZED CTEs. #104153 (Alexey Milovidov).
• Fix lowerUTF8 / upperUTF8 exception when processing large non-ASCII data sets (e.g. during text-index builds) where the accumulated output buffer would exceed 2 GiB; reject excessively long single rows with a clear error. #104229 (Shaohua Wang).
• Harden MergeTree column-statistics loading against transient I/O failures to prevent permanently disabling statistics-based optimizations for affected parts. #104372 (zoomxi).
• INSERT INTO ... SELECT FROM input(...) no longer requires the CREATE TEMPORARY TABLE grant. #104470 (Alexey Milovidov).
• Added a new MergeTree setting concurrent_part_removal_threshold_for_remote_disk (default 16) that is used instead of concurrent_part_removal_threshold when at least one part being removed is stored on a remote disk. The previous threshold of 100 could make DROP TABLE and other part-removal operations on object storage backends stall for tens of seconds because removals were performed serially even though each one is a separate network round-trip. The new setting causes the parallel removal path to be entered much sooner on remote storage, while leaving local-disk behavior unchanged. #104676 (Groene AI).
• New setting defer_partition_pruning_after_final (default 1) makes the 26.3-introduced behavior of skipping partition pruning under FINAL opt-out. Set it to 0 to restore pre-26.3 partition pruning — substantially faster on event-log workloads where same-PK rows cannot span partitions. compatibility = '26.2' flips it to 0 automatically. #104705 (Nikita Fomichev).
• Fix intExp2 returning wrong results for out-of-range inputs (shifts >= 64 under JIT, and integers with magnitude exceeding INT_MAX on both code paths). #105054 (Raúl Marín).
• Quote shell arguments built from CLI options and parsed input in clickhouse install and clickhouse git-import so paths, user/group names, and commit hashes containing whitespace or shell metacharacters are handled correctly. #105232 (Raúl Marín).
• The Alias table engine is now non-experimental and available without the allow_experimental_alias_table_engine setting. #103488 (Alexey Milovidov).
• Added --queries-format option to clickhouse-benchmark to choose between the default tab-escaped one-query-per-line input (tsv) and parsing standard input as a script of multiple queries separated by ; (script). #99972 (Aleksandr Musorin).
• The element-wise tuple operators tuplePlus, tupleMinus, tupleMultiply, tupleDivide, tupleModulo, tupleIntDiv, and tupleIntDivOrZero are now variadic and accept two or more tuples of the same size, applied element-wise as a left-fold. They previously accepted exactly two arguments only. #104659 (Aruj Bansal).
• Added a new built-in web UI at /processors-profile that visualizes the pipeline of any past SELECT query as a heatmap, sourced from system.query_log and system.processors_profile_log. Each processor is colored by its elapsed_us and shows per-processor stats (rows, bytes, wait times) on hover. #104614 (Nikita Mikhaylov).
• Setting query_plan_use_logical_join_step (and its alias query_plan_use_new_logical_join_step) is now obsolete and has no effect; the logical join step is always used. #104017 (Vladimir Cherkasov).
• The user input on play.html and the other built-in web pages (dashboard.html, jemalloc.html, merges.html, binary.html, webterminal.html) is no longer pre-filled with the literal text default. It now shows a user placeholder hint and is empty by default. When left empty, the page does not send a user= URL parameter (and the WebTerminal auth JSON omits the user field), so the server applies its usual fallback to the default user — and HTTP credentials supplied via other channels (X-ClickHouse-User, HTTP Basic, per-handler <handler><user> configuration) are no longer overridden. #105254 (Alexey Milovidov).
• ORC reader: decouple offset-based read (readBigAt) from use_prefetch, so EC-encoded data on HDFS can be read correctly even when use_prefetch = false (e.g. by Gluten). #103348 (zhanglistar).
• Files downloaded from play.html with the CSVWithNames format now use the .csv file extension. #103737 (JackFielding).
• More input validation for the BigLake catalog. #105117 (Konstantin Vedernikov).

​

Bug Fix (user-visible misbehavior in an official stable release)

• Fix inconsistent formatting by remembering if an expression was parenthesized. #92340 (Alexey Milovidov).
• Fix accurate comparison between Decimal and Float types. #94293 (zoomxi).
• Allow positional arguments in distributed queries. #94359 (simonmichal).
• Fix assertion failure in DatabaseCatalog::updateDependencies when dropping and recreating a materialized view with the same name; also fix view dependencies being silently lost during RENAME TABLE or EXCHANGE TABLES when referential dependencies were empty. #98779 (Alexey Milovidov).
• Fixed a server abort (assertion failure in DatabaseCatalog::getTableImpl) that occurred when creating a table via ON CLUSTER with a UUID identical to an existing database’s UUID. #98861 (xiaohuanlin).
• Fixed usage of skip indexes on data read (with enabled setting use_skip_indexes_on_data_read) and existing patch parts created by lightweight updates. #99543 (Alexey Milovidov).
• Fix “Not-ready Set” exception when a filter with IN (subquery) is moved to PREWHERE by the query optimizer. #100375 (Alexey Milovidov).
• Fix “Cannot find column” error when combining ADD COLUMN with RENAME COLUMN in a single ALTER TABLE statement. #100387 (Alexey Milovidov).
• Fix views with mixed UNION and INTERSECT/EXCEPT operators returning wrong results after DETACH/ATTACH or server restart. #100390 (Alexey Milovidov).
• Fix “Trying to execute PLACEHOLDER action” exception that could occur during stress testing with the AST fuzzer when a correlated subquery appeared in an IN clause. #100398 (Alexey Milovidov).
• Fix logical error exception when inserting into an Iceberg table with a Date column partitioned by year, month, or day transforms. #100404 (Alexey Milovidov).
• Fix skip index being used with incompatible data after ALTER TABLE MODIFY COLUMN changes the column type, which could cause server crashes in sanitizer builds or incorrect query results. #100526 (Alexey Milovidov).
• Fix credential leak in query_log for paimonCluster, paimonS3Cluster, paimonAzureCluster, and deltaLakeS3 table functions. #100529 (JIaQi Tang).
• Fix DROP TABLE on Kafka engine tables potentially hanging indefinitely due to a deadlock in rd_kafka_consumer_close. #100604 (Alexey Milovidov).
• Replace the hardcoded source_table_engines list with runtime lookup via StorageFactory and DatabaseFactory. Add source_access_type to DatabaseFactory::EngineFeatures so that CREATE DATABASE with source engines (PostgreSQL, MySQL, S3, etc.) requires the same source grants as CREATE TABLE. Fixes GRANT TABLE ENGINE ON * failing with table_engines_require_grant=false. Closes #71544. #100746 (pufit).
• Fix server crash when reading from a table whose ALIAS column contained a correlated subquery nested inside a function call (e.g. ALIAS toString(intDivOrZero(x, (SELECT ...)))). CREATE TABLE and ALTER TABLE now reject any subquery at any depth in DEFAULT / ALIAS / MATERIALIZED column expressions with THERE_IS_NO_DEFAULT_VALUE. Previously, the shallow validation accepted nested subqueries, so the correlated case crashed at read time and the nested non-correlated case appeared to work but could lead to unexpected behavior; both are now rejected at DDL time. #100753 (Groene AI).
• Fixed a server crash (Logical error: Bad cast from type DB::FunctionNode to DB::ColumnNode) that could occur when using queries with correlated subqueries on tables that have CONSTRAINT ... ASSUME definitions, when optimize_substitute_columns and convert_query_to_cnf settings are enabled. #100756 (Groene AI).
• Fix an exception that could occur when inserting a row into a JSON column where the row introduced a new dynamic field alongside a typed field with an incompatible value, and the column was subsequently used as a GROUP BY key. #100758 (Jimmy Aguilar Mena).
• Fix workload IO scheduling being silently bypassed for S3/object-storage writes that go through the DiskObjectStorageTransaction path (used by s3_with_keeper disks and explicitly when use_fake_transaction=false). Previously, INSERT writes into MergeTree tables on such disks ignored CREATE RESOURCE / CREATE WORKLOAD throttling because the resource link was never injected into WriteSettings for the transactional write path. #100777 (JIaQi Tang).
• Fix shared schema-cache collision between Protobuf and ProtobufList: reading a message first as FORMAT Protobuf and then as FORMAT ProtobufList could fail because ProtobufList forced the cached schema into its envelope form. ProtobufList now falls back to the cached message type when no envelope schema is available. #100849 (Callum Cooper).
• Fix heap-buffer-overflow in Rust CXX bridge due to std::exception ABI mismatch. #100931 (Azat Khuzhin).
• FunctionVariantAdaptor now throws ILLEGAL_TYPE_OF_ARGUMENT when all Variant alternatives are incompatible with a function, instead of silently returning Nullable(Nothing). Previously, a WHERE function(variant_col) predicate where no alternative type was compatible would return 0 rows with no error, while the equivalent SELECT context already threw correctly. #100939 (Vasily Chekalkin).
• TRUNCATE ALL TABLES no longer fails when the database contains views. #100943 (Alexey Milovidov).
• Fix the optimize_rewrite_array_exists_to_has optimization to correctly handle type-incompatible cases (e.g. Date vs String) and re-enable it by default. #100944 (Alexey Milovidov).
• Fix S3 settings priority so that storage_configuration disk settings override global <s3> section, and user/profile/query-level settings override both. #100975 (Alexey Milovidov).
• Fix sparse serialization losing the sign of negative zero (-0.0) for BFloat16, Float32, and Float64 columns. #100983 (Takumi Hara).
• Fixed two bugs in the WKT geometry parser used when reading WKT-encoded GeoParquet and Arrow files: MULTILINESTRING geometries were incorrectly parsed as Polygon (causing an exception for typed columns and silent data corruption for mixed Geometry columns), and a malformed WKT string with no type keyword triggered undefined behaviour instead of a clean error. #100997 (Vasily Chekalkin).
• Fix grouping/GROUPING SETS queries on Distributed tables with a single shard that failed with “Method executeImpl is not supported for ‘grouping’ function”. #101030 (Alexey Milovidov).
• Fix LOGICAL_ERROR “Column identifier is already registered” in the planner that could occur when the same table expression was processed multiple times. #101048 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception “Column identifier is already registered” when additional_result_filter setting is used with UNION or EXCEPT queries. #101051 (Alexey Milovidov).
• Fixed server crash (Logical error: Bad cast from ColumnVector to ColumnNullable) when using * APPLY with aggregate functions and group_by_use_nulls=1 with GROUPING SETS, ROLLUP, or CUBE. #101062 (Groene AI).
• Fix infinite loop when input_format_csv_skip_first_lines or input_format_tsv_skip_first_lines exceeds the number of lines in the file. #101111 (Alexey Milovidov).
• Fix out-of-bounds access in lazy materialization optimization that could cause an exception in debug builds. #101144 (Alexey Milovidov).
• Fix type inference in recursive CTEs: column types are now iteratively widened via getLeastSupertype across the non-recursive and recursive sides of the UNION ALL until convergence, preventing integer overflow in expressions like x + 1. #101155 (Alexey Milovidov).
• Clamp MergeTree compress block size settings (max_compress_block_size, min_compress_block_size, marks_compress_block_size, primary_key_compress_block_size) to 256 MiB to prevent server crash when extreme values are set via CREATE TABLE SETTINGS. #101159 (Groene AI).
• Fix a LOGICAL_ERROR crash in QueryAnalyzer::resolve() when a ROW POLICY uses a scalar subquery in its USING clause (e.g. USING (SELECT 1)). Closes #100695. #101263 (Groene AI).
• Fix undefined behavior in MergeTreeDataPartWriterCompact::cancel when a stream allocation fails. #101292 (Alexey Milovidov).
• Fixes a case where timezone was not included during the assignment in the ALTER statement. Closes #101328. Related https://github.com/ClickHouse/ClickHouse/pull/100647. #101403 (Yarik Briukhovetskyi).
• Fix server crash (LOGICAL_ERROR) for INSERT SELECT queries with ORDER BY ALL when the SELECT pipeline produces multiple streams. #101443 (Groene AI).
• Fix incorrect query results caused by join reordering pushing an INNER JOIN filter (referencing only the preserved side of an outer join) into the outer join’s ON clause. #101504 (Vladimir Cherkasov).
• Skip applying lazy materialization for plans with arrayJoin, which could lead to limit not being respected. Close #101608. #101644 (Vladimir Cherkasov).
• Fix incorrect JOIN query results when mergeFilterIntoJoinCondition optimization silently dropped a WHERE equality condition with mismatched types. #101652 (Xiaozhe Yu).
• Fix the error test hint not working for syntax/parsing errors — previously -- { error SYNTAX_ERROR } on a malformed query would fail with “Expected server error” instead of matching the client-side parse error. #101675 (Groene AI).
• Fix INTO OUTFILE ... TRUNCATE not actually using atomic rename — the write went directly to the original file instead of a temp file, so on query failure the original content was destroyed. Now the data is written to a temp file and renamed only on success. #101884 (Pablo Marcos).
• Fixed an exception Unsupported DeltaLake type: varchar(n) when reading Delta Lake tables whose schema contains varchar(n) or char(n) column types. These types are now mapped to String, consistent with how the Delta Lake protocol stores them as plain byte arrays in Parquet. #101973 (Flavio Malavazi).
• Fix silent data loss when reading tar archives from S3 with schema_inference_mode=union and heterogeneous Parquet schemas — the Parquet metadata cache incorrectly reused the first file’s metadata for all subsequent files in the archive. #101990 (Ahaan Limaye).
• Fix server crash (LOGICAL_ERROR) when executing ALTER TABLE UPDATE/DELETE on Iceberg tables when no prior SELECT or INSERT was done on the table in the same server lifetime. #102113 (Alexey Milovidov).
• Fix possible crash when an ALTER query executed on tables from in-memory database (e.g. temporary tables). #102360 (Den Kalantaevskii).
• Fixed formatQuery producing duplicate grants like GRANT FILE ON *.*, FILE ON *.* TO x instead of GRANT FILE ON *.* TO x when formatting backward-compatible READ/WRITE grants on the same source. #102411 (Groene AI).
• Fix lost parallelism for aggregation after read-in-order queries when max_streams_to_max_threads_ratio is greater than 1. #102467 (Alexey Milovidov).
• Fixed a crash (LOGICAL_ERROR: Unknown virtual column) when selecting subcolumns (like .null for Nullable, .size0 for Array, tuple elements, or map keys/values) from a Buffer table engine. #102470 (Groene AI).
• Fix exception “Cannot fold actions for projection” in join reorder optimization when a LEFT/RIGHT JOIN with join_use_nulls is combined with other joins involving more than two tables. #102516 (Alexey Milovidov).
• Fixed silent overflow in scalar pointwise operations on numericIndexedVector (e.g. numericIndexedVectorPointwiseAdd) when the scalar is out of range. Such inputs now raise INCORRECT_DATA instead of returning corrupted values. #102546 (FriendLey).
• Fix inconsistent AST formatting for INSERT with SAMPLE and query-level OFFSET. Closes #102523. #102547 (zoomxi).
• Fix LOGICAL_ERROR exception in groupConcat when deserializing a malformed aggregate function state. #102558 (Christoph Wurm).
• Now we do not accept garbage at the end of Time64 CSV values with input_format_csv_use_default_on_bad_values=0. Closes #102490. #102596 (Yarik Briukhovetskyi).
• Fix UNKNOWN_ELEMENT_OF_ENUM exception when inserting default value to JSON column with Enum typed paths. Closes #102359. #102687 (Pavel Kruglov).
• Fix columns_substreams.txt corruption during column rename in some cases. Closes #102259. #102689 (Pavel Kruglov).
• Fix inserting into Time data type during JSON parsing. Closes #102016. #102690 (Pavel Kruglov).
• Fix hex encoding of content sample in cached schema filenames. Closes #101904. #102703 (Pavel Kruglov).
• Fixed sort order violation (crash in debug, silent data corruption in release) during SummingMergeTree merge when the ORDER BY key is a hash expression over a Float32 column containing signaling NaN values. #102791 (Groene AI).
• Make DETACH DATABASE ... SYNC and other operations that call waitDetachedTableNotInUse cancellable via KILL QUERY and responsive to server shutdown, preventing indefinite hangs when a concurrent query holds a table reference. #102804 (Antonio Andelic).
• Fixed possible incorrect result of ANY RIGHT JOIN. #102893 (Nikita Taranov).
• Fix S3 requests failing with ios_base::clear: unspecified iostream_category error instead of being retried, caused by Poco BufferedStreamBuf::flushBuffer not handling short writes from the socket layer. #102894 (Sema Checherinda).
• Fix input('auto') table function failing via HTTP interface in INSERT SELECT queries. #102902 (Miсhael Stetsyuk).
• Disable trivial LIMIT optimization with row policies/additional_table_filters (to allow parallel index analysis). #102921 (Azat Khuzhin).
• Hide secret key in HMAC SQL function. Fix #102927. #102997 (Mikhail f. Shiryaev).
• Fix MULTIPLE_EXPRESSIONS_FOR_ALIAS exception raised on distributed queries that reference the same quantile call multiple times (e.g. in SELECT, HAVING, and ORDER BY) with optimize_syntax_fuse_functions enabled. #103014 (tanner-bruce).
• Fixes a case found by CI where a not-ready set was passed when the filter depends on the left-side column. Closes #102966. #103029 (Yarik Briukhovetskyi).
• Fix castOrNull to JSON in some cases. Closes #101818. #103036 (Pavel Kruglov).
• Fix SELECT DISTINCT silently returning incomplete results when an aggregate projection matched the query and some parts of the table had no projection data (e.g. the projection was added on a table that already held data, and MATERIALIZE PROJECTION was not run). Closes #102951. #103052 (Nihal Z. Miaji).
• Fix TOO_FEW_ARGUMENTS_FOR_FUNCTION exception when a WHERE predicate such as AND(OR(A, A), A) collapses to a single top-level argument during common expression extraction in the analyzer. #103072 (Peng).
• Fix a logical error Function writeSlice expects same column types for GenericArraySlice and GenericArraySink raised when evaluating if/ifNull over tuples, maps, or arrays containing a QBit element. ColumnQBit::structureEquals incorrectly compared the inner tuple of one QBit column against the outer wrapper of the other, so structurally identical QBit columns were reported as different. #103084 (Groene AI).
• Make DETACH DATABASE with database_atomic_wait_for_drop_and_detach_synchronously respect KILL QUERY instead of hanging indefinitely when a table reference is held. #103095 (Alexey Milovidov).
• Fix segfault when deleting WASM module using non-identifier predicate (e.g. WHERE 1=1). #103101 (Joe Redfern).
• Fix backward compatibility break where old clients fail with UNEXPECTED_PACKET_FROM_SERVER when inserting into a newer server via remote() or distributed tables, caused by unconditional sendProgress at the end of processInsertQuery. #103148 (Sema Checherinda).
• IN with non-constant tuple second argument must not CAST tuple elements down to the LHS type, this can overflow. Closes #103055. #103169 (Yarik Briukhovetskyi).
• Fix use-of-uninitialized-value in protocol function. #103187 (Pavel Kruglov).
• Fix LOGICAL_ERROR (Bad cast … to ColumnLowCardinality) during MergeTree index analysis when a WHERE clause compares a LowCardinality key column with a constant cast to a type that contains nested LowCardinality (for example Variant(LowCardinality(Date), String)). #103211 (Groene AI).
• Fix Logical error: 'index < bucket_count' in the timeSeries*ToGrid aggregate function family (e.g. timeSeriesResampleToGridWithStaleness, timeSeriesChangesToGrid, timeSeriesResetsToGrid, timeSeriesRateToGrid) when called with extreme timestamp parameters that would overflow signed 64-bit arithmetic in the bucket count computation. Also cap the total number of grid buckets at 16 million to prevent accidental large-memory allocation from adversarial inputs. #103223 (Groene AI).
• Fix Logical error: 'Port is already connected' exception during pipeline expansion in the lazy FINAL path for ReplacingMergeTree. The bug affected queries with query_plan_optimize_lazy_final = 1 when ReadFromMergeTree::initializePipeline inserted internal transforms (e.g. a Resize) that wired the sub-pipeline’s processors together, causing LazyUnorderedReadFromMergeTreeSource::expandPipeline to try to connect output ports that were already connected. #103230 (Groene AI).
• Fix wrong results from LIMIT BY and DISTINCT when the input is a UNION ALL of sorted subqueries. The query plan optimizer incorrectly treated such a union as globally sorted, causing extra rows to be returned. #103231 (Nihal Z. Miaji).
• Fix possible crash during statistics calculation in Map type with lazy replication. Closes #102390. #103273 (Pavel Kruglov).
• Fix a data race for ClickHouse, MySQL, PostgreSQL and XDBC dictionary sources where clone() const reads from and isModified() const writes to the same invalidate_query_response string. #103277 (Miсhael Stetsyuk).
• Fix incorrect monotonicity detection for Date32. Closes #101265. #103283 (Yarik Briukhovetskyi).
• Cap pre-auth TCP Hello packet strings to 64 KB and add handshake_timeout_milliseconds server setting to limit total handshake time, preventing unauthenticated clients from consuming excessive memory or holding threads indefinitely. #103284 (Sema Checherinda).
• Fix Parquet ColumnIndex stats min_value > max_value for String columns. #103334 (Saurabh Kumar Ojha).
• Fix incorrect handling of NULL rows for Nullable(Tuple(...)) inputs in flattenTuple and tupleToNameValuePairs. flattenTuple now preserves the outer null map so NULL rows stay NULL. tupleToNameValuePairs now changes the result’s value type to Nullable(T) when possible when the input is Nullable(Tuple(...)), so NULL rows produce [('a', NULL), ('b', NULL)] instead of [('a', 0), ('b', 0)]. When the element type cannot be wrapped in Nullable (e.g. Array), default values are used instead of NULL. Closes #103312. #103383 (Nihal Z. Miaji).
• Check for malformed flattened Dynamic data in Native format. #103392 (Pavel Kruglov).
• Fix stale RPC causing replication stall after snapshot sync. #103406 (Seva Potapov).
• Fix a LOGICAL_ERROR “Cannot pop N rows from X” that aborted the server (in debug / sanitizer builds) when reading a malformed BSON document containing a value whose BSON type is incompatible with the target column type inside a Nullable(T) or Array(Nullable(T)) column. The malformed row now produces a clean ILLEGAL_COLUMN exception as intended. #103418 (Groene AI).
• Fix a server abort in arrayFill and arrayReverseFill when applied to an Array(String) column whose first row is empty together with a constant-false lambda. The aggregate / sub-array loop underflowed a size_t index to SIZE_MAX, which then caused an out-of-bounds read inside ColumnString::doInsertManyFrom. Same bug family as #12263. #103424 (Groene AI).
• Fixes ACCESS_DENIED / UNKNOWN_TABLE in downstream APPEND RMVs with SQL SECURITY DEFINER when an upstream REPLACE RMV’s EXCHANGE flips the target storage identity mid-resolution. #103427 (Alexander Gololobov).
• Populate _time column from url table function. #103437 (Nikita Taranov).
• Fix potential out-of-bounds read in the ALP codec decompression path when processing malformed input with an invalid bit-width. #103457 (Raufs Dunamalijevs).
• Fix a bug where SYSTEM SYNC FILESYSTEM CACHE '<name>' ON CLUSTER ... could lose the cache name while formatting the query, causing remote nodes to sync all filesystem caches instead of only the requested cache. #103469 (Asish Kumar).
• Fix processing of URL-encoded path in deltaLakeAzure. Closes #103509. #103525 (Smita Kulkarni).
• Fix heap-use-after-free in multiplying an AggregateFunction state by an integer (e.g. quantilesExactState(...) * N). The exponentiation-by-squaring loop merged the state with itself, which is undefined when the aggregate function’s merge reallocates its internal storage. Closes STID 0988-40af. #103536 (Groene AI).
• Fixed a rare LOGICAL_ERROR “Part X intersects previous part Y” raised during ReplicatedMergeTree table startup when two empty unexpected parts on disk had overlapping but non-containing block ranges. The exception aborted the table-attach thread and prevented the table from coming up. #103537 (Groene AI).
• Fix Logical error: Incorrect mark rows for part ... (debug-only assertion) that was triggered by mutations on MergeTree tables with non-adaptive index granularity (index_granularity_bytes = 0) whose last data mark was incomplete (most reliably reproduced via DETACH/ATTACH followed by a lightweight DELETE). #103538 (Groene AI).
• Fix clickhouse-local returning 0 rows silently when reading from /proc and /sys pseudo-files via the file() table function (e.g. SELECT * FROM file('/proc/cpuinfo', 'RawBLOB')). #103548 (Ashrith Bandla).
• Fixed handling of the max_string_length argument for the GenerateRandom table engine. #103550 (Alex Kuleshov).
• Fix masking nested credentials in logs. #103552 (Vitaly Baranov).
• Fix SVE detection using SVE instructions when unavailable. #103568 (Raúl Marín).
• Fix GenerateRandom storage args parsing. #103574 (Konstantin Bogdanov).
• Fix infinite loop on WITH FILL for data starting with ±inf. #103580 (Konstantin Bogdanov).
• Fix JSONExtract into a Variant type silently truncating fractional JSON numbers. Previously, JSONExtract('{"x": 3.14}', 'x', 'Variant(Int64, Float64)') returned Int64=3 and JSONExtract('{"x": 3.14}', 'x', 'Variant(String, Int64)') returned Int64=3, dropping the fractional part. The fractional value is now preserved losslessly: a Float64/Decimal member claims it when present, otherwise a String member captures the original JSON. Variant types containing only integer members (e.g. Variant(Int64, Int32)) and direct integer extraction (JSONExtract(json, 'Int64'), JSONExtractInt, etc.) are unchanged. #103620 (Groene AI).
• Fix SYSTEM INSTRUMENT REMOVE without arguments producing std::bad_optional_access (error code 1001) instead of SYNTAX_ERROR (error code 62). #103622 (Pablo Marcos).
• Fix logical error / undefined behavior in windowID and tumble window view when called with a timezone string as the 3rd argument. #103641 (Alexey Milovidov).
• S3 client logging now treats HTTP 400 responses that include a non-empty x-amz-bucket-region header (wrong SigV4 signing region) as an informational wrong-region case with a clearer log line instead of the generic error-style path. The STS web identity credentials provider is added to the S3 credentials chain only when web identity is configured, which reduces spurious warnings for deployments that do not use it. Closes #99140. #103673 (MeltonSmith).
• Fix SYSTEM SYNC REPLICA <db>.<tbl> IF EXISTS to silently succeed when the database does not exist, matching the existing behaviour for a missing table and the precedent set by DROP TABLE IF EXISTS. Previously the query threw UNKNOWN_DATABASE despite IF EXISTS. Closes #103629. #103689 (Groene AI).
• Fix logical error Arguments of 'plus' have incorrect data types thrown by FunctionBinaryArithmetic::executeImpl2 when a MergeTree table had Array(LowCardinality(...)) in its sort key and a WHERE clause used plus / minus between that column and an Array constant of a different element type. KeyCondition::getMonotonicityForRange now strips LowCardinality recursively before invoking the inner numeric dispatch. #103701 (Groene AI).
• Fix dynamic cache resize race in 26.1+. #103702 (Kseniia Sumarokova).
• Fixed an issue where text and bloom_filter skip indices on ALIAS columns whose expression contained a lambda with captured constants (e.g. arrayMap((k, v) -> concat(k, '=', v), mapKeys(m), mapValues(m))) were silently ignored. #103708 (Anton Popov).
• Fix position and positionCaseInsensitive with a start_pos close to UINT64_MAX getting stuck or causing a segmentation fault due to overflowing pointer arithmetic. #103766 (Raúl Marín).
• Fixes undefined behavior while parsing dateTime with a double fractional part. #103773 (Yarik Briukhovetskyi).
• Fixed a data-correctness regression in 26.x where JIT-compiled if and multiIf with a Decimal result type and a non-Decimal (integer or float) literal in one branch silently returned a value 10^scale too small. The slow (non-JIT) path was unaffected. Workaround on affected versions: SET compile_expressions = 0. #103809 (Groene AI).
• Fixes forwarding the tokenizer from the text index to supported functions. #103826 (Elmi Ahmadov).
• Fix wrong results from numbers, generate_series, and similar range-honoring sources when the WHERE clause uses IN or NOT IN over a tuple whose elements deduplicate to a single key column (for example, WHERE tuple(number, number) NOT IN (tuple(1, 2))). Closes #103660. #103835 (Groene AI).
• Keep the auto-spilling hash join’s actual memory peak under max_bytes_before_external_join. Previously, statistics-driven preallocation, in-place hash table doubling, and unbounded GraceHashJoin in-memory buckets could each push the query past the configured cap and trip MEMORY_LIMIT_EXCEEDED. #103838 (Alexey Milovidov).
• Fix a hang in clickhouse-local shutdown when system logs (such as text_log or filesystem_cache_log) are configured. SystemLogs::flushImpl was calling BaseDaemon::instance().flushTextLogs(), which throws std::bad_cast outside clickhouse-server and left the saving threads running until pthread_cond_destroy blocked the destruction of the system log queues. #103874 (Alexey Milovidov).
• Reject MATERIALIZED CTE queries whose body resolves to different inferred column types in different references with a clear TYPE_MISMATCH error instead of crashing the server with a Bad cast LOGICAL_ERROR. This previously happened when the CTE body referenced identifiers from outer scope (for example an alias from the calling subquery’s projection) that were inlined as different constants per reference. #103879 (Groene AI).
• Fix a use-of-uninitialized-value in the JOIN-conversion query plan optimizers (tryConvertAnyOuterJoinToInnerJoin, tryConvertAnyJoinToSemiOrAntiJoin) when the filter on top of an ANY OUTER JOIN contains a non-deterministic function such as rand, now, or rowNumberInAllBlocks. The optimizer no longer attempts to constant-fold such filters and leaves the JOIN unchanged, which also prevents incorrect conversions to INNER/SEMI/ANTI JOIN that could silently drop rows. #103880 (Groene AI).
• Fix a use-after-free in StorageKafka2 that could crash the server when a Keeper session is replaced while the consumer is holding ephemeral topic-partition locks. #103890 (Groene AI).
• Fixed a LOGICAL_ERROR “Primary key type mismatch” thrown when joining an EmbeddedRocksDB table via JOIN ... USING (key) against a subquery whose key column has a different type (e.g. Nullable(UInt64), Int64, Decimal) than the storage’s primary key. The planner now declines DirectKeyValueJoin on type mismatch and falls back to HashJoin, which handles the type conversion. #103928 (Groene AI).
• SET max_threads = DEFAULT (and the same for max_final_threads and max_parsing_threads) no longer loses the auto state. Previously, after resetting one of these settings, system.settings would report the resolved core count (e.g. 32) instead of 'auto(32)', and the server would behave as if the value had been explicitly pinned. #103991 (Groene AI).
• Fix SIZES_OF_ARRAYS_DONT_MATCH when reading AggregateFunction(topK(N), String) columns whose persisted alpha_size was grown by a pre-25.12 deserialize+serialize cycle. #104002 (Raúl Marín).
• Fixed SQL injection vulnerability in PostgreSQL, Cassandra and XDBC dictionary sources: string keys containing single quotes were escaped with \' (backslash), which these backends treat as a literal backslash rather than an escape sequence, allowing arbitrary SQL to be injected into dictionary lookup queries. ExternalQueryBuilder now emits SQL-standard '' escaping for those backends; ClickHouse and MySQL dictionary sources continue to use backslash escaping. #104009 (Shaohua Wang).
• Fixed possible type mismatch in aggregation when a column type was altered to/from LowCardinality, and min-max projection wasn’t rebuilt. #104013 (Nikita Taranov).
• Fix infinite loop of query optimizations caused by not propagating prevention of unused column removal when merging expressions. #104083 (János Benjamin Antal).
• Fix LOGICAL_ERROR in StreamingStorageRegistry::renameTable when batch-renaming streaming tables (S3Queue, Kafka, RabbitMQ) by using UUID-based identity tracking instead of name-based. #104101 (Nikita Taranov).
• Fixed wrong results when a CTE constructed with UNION ALL contained a SELECT DISTINCT branch and the outer query projected only a subset of the CTE’s columns. The new analyzer’s RemoveUnusedProjectionColumnsPass was incorrectly removing the un-referenced column from the inner DISTINCT projection, causing rows that should have remained distinct (same value in the projected column but different value in the dropped column) to collapse into one. #104114 (Groene AI).
• Fix several edge case correctness issues in setting optimize_inverse_dictionary_lookup where the optimization could silently drop rows or suppress exceptions. Closes #103270. Close #103085. #104133 (Nihal Z. Miaji).
• Fix write out of bounds while deserializing quantileTiming state. #104141 (Alexey Milovidov).
• Fix undefined behavior when an out-of-range Float64 value is converted to a wide integer type (UInt64, Int64, Int128, UInt128, Int256, UInt256). Previously, values equal to Float64(numeric_limits<T>::max()) (which rounds up to a value above the actual maximum) bypassed the bounds check and produced UB in the subsequent cast. This affected aggregate-function parameter parsing (topK, histogram, uniqUpTo, groupArrayInsertAt, etc.) and integer-typed settings via SET <setting> = <Float64>. Closes #103817. #104154 (Groene AI).
• Fix LOGICAL_ERROR exception “Unexpected return type from comparison. Expected UInt8. Got Const(Nullable(UInt8))” when comparing a nested Tuple(Tuple(Nullable(...))) (or deeper nesting) with a String literal. The comparison’s return type is now correctly inferred as Nullable(UInt8), matching the runtime behaviour. #104171 (Groene AI).
• Fixed a server abort (UndefinedBehaviorSanitizer: reference binding to null pointer, segfault in release builds) on SELECT ... FROM <ReplacingMergeTree(version, is_deleted)> FINAL PREWHERE is_deleted = <expr> AND <other column expr> when query_plan_optimize_lazy_final = 1. The lazy-FINAL non-intersecting reading step lost the is_deleted column from its output header because prewhere consumed it as an input without re-exposing it as an output. The downstream addIsDeletedFilter step then dereferenced a null ActionsDAG::Node pointer. #104177 (Groene AI).
• When calling system.failpoints table, it would use a failpoint, thus possibly disabling it. #104237 (Pedro Ferreira).
• Fix MemorySanitizer: use-of-uninitialized-value in detectLanguage* functions when the input contains a UTF-8 character ending exactly at the buffer boundary. #104257 (Raúl Marín).
• Fix a logical error (Bad cast from type DB::CachedObjectStorage to DB::S3ObjectStorage) that aborted argument parsing of data-lake table functions and engines (icebergS3, deltaLakeS3, etc.) when called with SETTINGS disk = '...' against a disk whose underlying object storage is wrapped by a decorator such as a filesystem cache. Closes #89300. #104258 (Groene AI).
• Fix parsing of parenthesized subscript expressions on columns named values, for example (values['a']), so they are no longer interpreted as SQL-standard VALUES table expressions. #104312 (Desel72).
• Fix a server abort/LOGICAL_ERROR in the filesystem cache background eviction thread (SLRUFileCachePriority::collectEvictionInfo) that could fire when keep_free_space_size_ratio or keep_free_space_elements_ratio was high enough to trigger eviction while all cache entries had already been promoted to the SLRU protected queue (probationary empty). #104313 (Groene AI).
• Fix exceptions (NOT_FOUND_COLUMN_IN_BLOCK, LOGICAL_ERROR, AMBIGUOUS_COLUMN_NAME) when using projections with UNION ALL views, window functions, or alias-column name collisions. Regression from #88798. #104317 (Amos Bird).
• Fixes a possible underflow in parsing postgres arrays. #104322 (Grant Holly).
• Functions that return a non-Nullable type (such as Array, Tuple, or Map) now accept Nullable arguments. Affected functions include extractAll, extractAllGroups, extractAllGroupsHorizontal, extractAllGroupsVertical, extractGroups, splitByChar, splitByString, splitByRegexp, splitByWhitespace, splitByNonAlpha, and alphaTokens. NULL input rows produce the default value of the result type (e.g. an empty array) instead of raising “Nested type is not allowed inside Nullable type”. #104326 (Alexey Milovidov).
• Fix iceberg stats with partitioned table. This closes #104321. #104329 (Konstantin Vedernikov).
• Fix a fatal logical error at server startup on macOS (and similar jemalloc builds) where Jemalloc::verifySetup incorrectly reported a jemalloc_enable_background_threads mismatch because optional background_thread / max_background_threads mallctls are absent; verification is skipped when those mallctls are unavailable, and getValue no longer leaves the output uninitialized on failure. Closes #102183. #104330 (SAYON DEEP).
• Fix a correctness regression where SELECT queries with max_rows_to_read_leaf and read_overflow_mode_leaf = 'throw' could incorrectly throw TOO_MANY_ROWS even when a skip index would have reduced the read below the leaf cap. The symmetric non-leaf settings (max_rows_to_read / read_overflow_mode) were already handled. #104331 (Groene AI).
• toUUID, toUUIDOrNull, toUUIDOrZero, toUUIDOrDefault, CAST to UUID and Nullable(UUID), accurateCastOrNull to UUID, and the input formats that parse a UUID from text (Avro, MsgPack, JSONExtract, …) now reject strings that have the right length but contain non-hexadecimal characters. Previously such inputs were silently turned into a fabricated UUID by walking off the end of the hex digit lookup table; now toUUID throws CANNOT_PARSE_UUID, and the Or* variants return NULL / the zero UUID / the supplied default. #104370 (Groene AI).
• Fix Bad cast LOGICAL_ERROR in caseWithExpression (and SQL CASE expr WHEN ... THEN ... ELSE ...) when the THEN-only supertype and the (THENs + ELSE) supertype land in different ColumnDecimal storages — for example THEN values (UInt16, Int8) with an ELSE of Decimal(9, 2). Closes #104335. #104378 (Groene AI).
• Internal failures reported via logExceptionBeforeStart (asynchronous insert flushes, materialized-view refreshes, parse errors raised during internal queries) now correctly increment the FailedInternalQuery, FailedInternalSelectQuery, and FailedInternalInsertQuery ProfileEvents alongside the user-visible FailedQuery, FailedSelectQuery, and FailedInsertQuery counters. Previously these internal counters stayed at zero for failures occurring before query execution started, undercounting a major class of internal failures. #104399 (Groene AI).
• Fix Code: 36. BAD_ARGUMENTS Expected literal, got {name:Type} thrown when a query parameter is used inside the base_backup setting of a BACKUP or RESTORE statement (e.g. BACKUP ... SETTINGS base_backup = S3({backup_name:String}, ...)). The regression was introduced in 26.1.5 by PR #99205. Closes #103324. #104413 (Groene AI).
• Fix a server abort in IcebergLocal/IcebergS3 writes when ALTER TABLE ... DROP COLUMN is followed by an INSERT with iceberg_metadata_staleness_ms greater than zero. ALTER now invalidates the local Iceberg metadata files cache so subsequent reads and writes see the new schema. #104419 (Groene AI).
• Fix DROP ROLE, DROP USER, DROP SETTINGS PROFILE, DROP ROW POLICY, DROP QUOTA and DROP MASKING POLICY to remove references to the dropped entity from any other access entity that referenced it (e.g. a user’s DEFAULT ROLE list, a settings profile’s TO list, a row policy’s grantees) and persist the cleanup to disk. Previously the in-memory state appeared correct because SHOW CREATE filters unknown UUIDs, but the on-disk .sql files retained dangling ID('<dropped-uuid>') entries and the references were resurrected on the next server restart, surfacing as ACCESS_ENTITY_NOT_FOUND errors during distributed query execution. #104427 (Groene AI).
• Fixed a server abort triggered by azureBlobStorage, AzureBlobStorage-engine, and DeltaLake-on-Azure with a connection string whose BlobEndpoint URL has an empty, non-numeric, or out-of-range port (e.g. BlobEndpoint=http://host:abc/). The server now returns a clean BAD_ARGUMENTS error instead of aborting in debug/sanitizer builds. #104460 (Groene AI).
• Allow an unquoted identifier as the user name in the remote and remoteSecure table functions, mirroring how the database and table arguments accept unquoted identifiers. Previously such a query failed with a misleading authentication error referring to the default user. #104465 (Alexey Milovidov).
• Fixed Logical error: Incorrect ASTSelectWithUnionQuery (modes: M, selects: N) triggered when a SQL user-defined function body contains a parenthesized inner UNION ALL (e.g. CREATE FUNCTION f AS x -> (SELECT 1 UNION ALL (SELECT 1 UNION ALL SELECT 1))). #104477 (Groene AI).
• Fix uniqThetaIntersect returning the cardinality of the first argument instead of 0 when the second argument is an empty uniqTheta state — for example the result of uniqThetaMergeStateIf(s, predicate) when the predicate excludes every row. #104529 (Groene AI).
• Fix SHOW TABLES and system.tables silently truncating the listing for DataLakeCatalog databases backed by Iceberg REST catalogs (iceberg-rest, onelake, biglake). When the catalog server paginates the list-tables or list-namespaces response (e.g. Microsoft Fabric / OneLake beyond ~50 tables per namespace), tables on later pages were silently invisible to SHOW TABLES and system.tables, even though they were queryable via direct SELECT. RestCatalog now follows the next-page-token continuation token defined by the Iceberg REST OpenAPI spec, matching the existing behavior of PaimonRestCatalog and UnityCatalog. #104531 (Groene AI).
• The input table function now infers its structure from the surrounding INSERT query’s FORMAT clause when that format has a fixed schema (LineAsString, RawBLOB, JSONAsString, etc.), so users no longer have to repeat the structure as input('line String') for these formats. #104532. #104533 (Groene AI).
• Updated datatype of fields in Iceberg history from Int32 to Int64. Closes #94176. #104579 (Smita Kulkarni).
• Fixed Inconsistent AST formatting LOGICAL_ERROR when parsing a function call where a lambda follows a comma, e.g. SELECT substring(x, `x` -> `x`). The parser used to silently merge the preceding arguments into the lambda’s left-hand side, producing a single-argument call that could not be re-parsed back to the same AST. It now preserves the function’s original arity. #104626 (Groene AI).
• CHECK TABLE t now prefers a TEMPORARY table over a permanent one with the same name when no database qualifier is given, matching the precedence already used by SHOW CREATE TABLE, DESCRIBE TABLE, OPTIMIZE TABLE, and ALTER TABLE. Previously CHECK TABLE t skipped temporary tables entirely, so it failed with UNKNOWN_TABLE on a temporary Log or File table even though those engines support CHECK. Follow-up to #100966. #104637 (Groene AI).
• Fix Keeper termination and restart loop when get /keeper/availability_zone is sent with quorum_reads=true to a Keeper without <placement> configured. #104663 (myeongjun).
• Fix a TOCTOU data race in FutureSetFromTuple::buildOrderedSetInplace that results in a logical error. #104673 (Miсhael Stetsyuk).
• Fix several functions returning different results for the same input depending on whether arguments arrived as columns or constants: bitRotateLeft / bitRotateRight (boundary shift counts), length(FixedString) / concatWithSeparator with LowCardinality(Nullable) inputs, roundDown on NaN, and rightUTF8 on invalid UTF-8. #104710 (Raúl Marín).
• Fixed anyHeavy returning a non-heavy value when the most frequent value was the column default and the data lived across multiple MergeTree parts (sparse-column read path). #104712 (Raúl Marín).
• Fix several defects in setting-constraint handling: MergeTreeSettings constraints declared on a setting’s canonical name could be bypassed by writing to an alias of that setting; the disallowed_values constraint check threw an exception on clamp paths (secondary queries, ON CLUSTER workers, SQL SECURITY DEFINER views) instead of silently dropping the change. #104737 (Raúl Marín).
• Fix a LOGICAL_ERROR exception (Metadata is not initialized) raised by OPTIMIZE TABLE, ALTER TABLE ... DELETE, ALTER TABLE ... ADD COLUMN and other ALTER variants on a lazily-attached Iceberg / IcebergLocal / DeltaLake / Hudi table whose metadata had not been loaded yet (typical after a server restart, or after a previous metadata write failed and left a corrupted metadata file on disk). The operation now either proceeds normally if the metadata loads successfully, or surfaces the underlying load failure as a regular user-facing exception instead of aborting the server in debug / sanitizer builds. #104738 (Groene AI).
• Fix segfault due to a use-after-free bug in AvroConfluentRowInputFormat. #104751 (Miсhael Stetsyuk).
• Scalar variants of numericIndexedVectorPointwiseMultiply, numericIndexedVectorPointwiseDivide, numericIndexedVectorPointwiseEqual, and numericIndexedVectorPointwiseNotEqual now raise INCORRECT_DATA when called with a UInt64 scalar above Int64::max. #104784 (FriendLey).
• Fix a bug where manually overriding a setting via its alias name (e.g. SET enable_analyzer = 1 instead of SET allow_experimental_analyzer = 1) after applying a compatibility setting could cause that override to be reverted by a subsequent change of the compatibility setting. #104829 (Raúl Marín).
• Fix AWS logger being disabled after https://github.com/ClickHouse/ClickHouse/commit/0e8ad4355c9d. #104837 (Konstantin Bogdanov).
• Fix three more functions returning different results for the same input depending on whether arguments arrived as columns or constants: transform (and caseWithExpression through it) with a constant Date/Date32/Enum/FixedString default, comparison operators between a String and a constant FixedString, and if/ifNull/nullIf with a FixedString branch under a constant condition. #104858 (Raúl Marín).
• Fix Bad cast from type DB::ColumnConst to DB::ColumnNullable server abort during partition pruning for MergeTree tables when the partition expression contains a function chain that collapses to a single constant value (such as floor(NULL, toRelativeYearNum(...))). #104861 (Groene AI).
• Fix a ThreadSanitizer data race in the ContextData copy constructor: table_function_results was copied from the source object without acquiring table_function_results_mutex, so a concurrent Context::executeTableFunction writer could race against the unsynchronized read in the copy constructor. #104879 (Groene AI).
• Fix data part consistency checks for types with dynamic structure and detect corrupted columns_substreams.txt. Resubmit of https://github.com/ClickHouse/ClickHouse/pull/103858 with additional changes. #104888 (Pavel Kruglov).
• Fix race between DROP and UNDROP in DatabaseCatalog. #104915 (Azat Khuzhin).
• Fix filesystem cache dynamic resize with partially downloaded segments, including restore accounting after failed eviction. #104921 (Antonio Andelic).
• DETACH TABLE on a temporary table (without the TEMPORARY keyword) now correctly raises SYNTAX_ERROR, matching the behavior of DETACH TEMPORARY TABLE. Previously it silently set an internal is_detached flag and returned without error. To remove a temporary table, use DROP TEMPORARY TABLE or DROP TABLE (the latter resolves the temporary table via Context::ResolveExternal). Closes #103475. #104943 (Groene AI).
• Make the filesystem table function honor max_memory_usage / max_server_memory_usage when loading file content. Previously, large or numerous parallel content reads could push past the limit without raising MEMORY_LIMIT_EXCEEDED and end up OOM-killed instead. #104956 (Alexey Milovidov).
• Fixed a LOGICAL_ERROR thrown when filtering system.detached_tables by uuid (e.g. SELECT count() FROM system.detached_tables WHERE uuid = '...'). The query now returns the expected result instead of aborting the server. #104979 (Groene AI).
• flattenTuple no longer raises a LOGICAL_ERROR when called on a tuple whose nested structure contains only empty Tuple() leaves (e.g. Tuple(c0 Array(Tuple())) or Tuple(c0 Tuple())). Such inputs now produce a user-facing ILLEGAL_TYPE_OF_ARGUMENT exception explaining that the flatten result would be an empty tuple. #104989 (Groene AI).
• Fix a server abort and a silent-wrong-result bug when querying loop(remote(...)) (or any loop() wrapping a storage that can defer aggregation) with GROUP BY. The outer planner used to add MergingAggregatedStep based on the inner storage’s reported processing stage, but LoopSource always materialises its inner select with QueryProcessingStage::Complete and emits plain column chunks, so MergingAggregatedTransform tripped a LOGICAL_ERROR (Chunk info was not set for chunk in MergingAggregatedTransform) under enable_parallel_replicas = 1 and silently dropped the outer aggregation otherwise. #105001 (Groene AI).
• clickhouse-benchmark --reconnect (bare, no value) was inadvertently broken in 25.4 by a change that turned --reconnect into an integer option, requiring a value. The bare form now works again and is equivalent to --reconnect=1 (reconnect on every query). #105006 (Groene AI).
• Fixed malformed JSON output for column names ending with incomplete UTF-8 sequences. #105012 (Pablo Marcos).
• Fix silent data loss after EXCHANGE TABLES or CREATE OR REPLACE TABLE of a materialized view’s source table. The MV’s source-view dependency edge is now kept on the original name so it continues to fire on inserts. Regression introduced by #98779; restores the pre-regression behavior. #105029 (Sema Checherinda).
• Fix CANNOT_COMPILE_CODE Could not find symbol __fixunsdfti when JIT-compiling expressions that convert a Float to UInt128, such as toUInt128(<Float64 expression>). The unsigned 128-bit float-to-int compiler-rt builtins were missing from the JIT symbol resolver. Closes #105031. #105048 (Raúl Marín).
• Fix clickhouse-local not printing log messages from a failed query when send_logs_level is set. #105067 (Alexey Milovidov).
• Fix incorrect results when the same parameterized view is referenced more than once in the same query with different argument values. Previously, the analyzer collapsed the calls into one, silently dropping all but the first filter. #105170 (Alexey Milovidov).
• Fix TOTALS row being rendered twice at the bottom of the result table in the play.html web UI. #103803 (Alexey Milovidov).
• Reject non-finite vectors (NaN, ±Inf) during vector search (either as searched vectors or as reference vectors); they previously caused undefined behavior in usearch. #104079 (Groene AI).
• PromQL: fix the aggregation operator for empty vectors. #104425 (Vitaly Baranov).
• PromQL: fix error handling in the Prometheus query API. #104741 (Vitaly Baranov).
• Fix a race in MergeTreeTransaction::afterCommit where, after a connection loss between writing the commit CSN to ZooKeeper and finalizing the transaction, the COMMIT response could reach the client before the new creation_csn / removal_csn became visible in system.parts. #104708 (Tuan Pham Anh).
• Fix exponential memory growth in the KQL parser when converting nested array indexing (arr[arr[arr[...]]]). #105142 (Alexey Milovidov).
• Fix incorrect results for RIGHT ANY JOIN when the right table has multiple rows per key and the output block is split due to size limits. Closes #99431. #102064 (Vladimir Cherkasov).
• Check for stack overflow in Avro reader during nested-type deserialization. #102417 (Pavel Kruglov).
• Fix duplicate rows in system.completions for MergeTree settings — each setting name appeared twice because both getMergeTreeSettings and getReplicatedMergeTreeSettings were dumped despite having identical setting names. Closes #102013. #102015 (Groene AI).
• Fix StorageObjectStorageQueue (S3Queue, AzureQueue) blocking shutdown until partially-processed files were fully read from object storage. The source now aborts the read immediately on shutdown; deduplication ensures rows already streamed into the destination table before shutdown are not duplicated when the file is retried on next start. #103126 (Tuan Pham Anh).
• Fix multi-block inserts to the Alias table engine with insert_deduplication_token so that all blocks are kept. #103246 (Enric Calabuig).
• Fix JSONHas and JSONExtractBool on native JSON columns returning the extracted value (cast to UInt8) instead of 0/1. #103313 (zxuhan7).
• Protect the CustomSeparated input format against malformed or adversarial data that omits format_custom_row_after_delimiter. Header detection, schema inference, and variable-column rows previously accumulated fields unboundedly and could allocate many gigabytes of memory before failing. Reads now fail with INCORRECT_DATA once a single row contains more than 1,000,000 fields. #103404 (Groene AI).
• Fix RESTORE of replicated MergeTree tables so that restored part attachment uses backup_restore_keeper_max_retries instead of the regular insert Keeper retry budget. #104610 (Pablo Marcos).
• Fix wrong results from WHERE p AND <LowCardinality(Nullable(int)) constant> on MergeTree tables. Previously such queries returned zero rows because part pruning derived a NULL default from the LowCardinality(Nullable(...)) type and synthesized a notEquals(x, NULL) guard, pruning every part. #104767 (Groene AI).
• Fix incorrect aggregation result when grouping by a non-injective function of a column that is also the partition key (e.g. PARTITION BY a with GROUP BY intDiv(a, 2) or a % 2). Values from different partitions that mapped to the same group were not merged, producing duplicate group rows under allow_aggregate_partitions_independently = 1. #104869 (Nihal Z. Miaji).
• Fix a race in Keeper snapshot transfer where a snapshot could be removed or moved while it was being sent to a recovering follower. #104941 (Antonio Andelic).
• Fix a silent wrong-result bug in hilbertEncode and mortonEncode where a non-constant Tuple first argument (range mask) used row 0’s values to drive the bit-shift for every row. The functions now compute the per-row mask values, so the result no longer depends on the input being constant or on the block size. #104992 (Groene AI).
• Fix clickhouse-local silently ignoring user-level configuration (profiles, users, quotas, access-control settings) when the configuration file is auto-discovered from ./clickhouse-local.xml, ~/.clickhouse-local/config.xml, or /etc/clickhouse-local/config.xml. Previously these settings were applied only when --config-file was passed or ./config.xml existed in the current directory; now all discovery paths behave consistently. #105008 (Groene AI).
• Fix BFloat16 columns silently returning zero matches when compared against a string literal (e.g. WHERE bf16_col = '49.9'). #105042 (Raúl Marín).
• Fix CLEAR COLUMN and TTL handling during merges for SummingMergeTree, AggregatingMergeTree, and CoalescingMergeTree when merge-required columns are absent or expired. Closes #101953. #105203 (Antonio Andelic).
• Fix a race between DNSCacheUpdater::run and Context::reloadClusterConfig by adding an explicit null-guard for shared->clusters_config before dereferencing it. #105220 (Mikhail f. Shiryaev).
• Fix best_effort date-time parsing for toDateTime64 with ms and ns precision. #105233 (Kaviraj Kanagaraj).
• Fix OSIOWaitMicroseconds reporting thread-lifetime I/O wait instead of per-query I/O wait. #105246 (Mikhail f. Shiryaev).
• Hudi table engine now raises INCORRECT_DATA (a regular query-level exception) instead of LOGICAL_ERROR when a parquet file in the table directory does not match the Hudi [FileId]_[FileWriteToken]_[Timestamp].[extension] naming convention. Previously such file names caused an exception in debug builds. #105266 (Groene AI).
• Fix deadlock in ParallelFormattingOutputFormat on schedule failure: when scheduleFormatterThreadForUnitWithNumber throws (e.g. CANNOT_SCHEDULE_TASK), the unit was left in READY_TO_FORMAT with no formatter thread to process it, leading to a hang. The schedule call is now wrapped in try/catch and any failure is routed through onBackgroundException so the collector exits cleanly. #105275 (Azat Khuzhin).
• Fix silent data loss in Distributed async inserts when recovering from an abnormal shutdown: if the last .bin file in a saved batch was intact but a middle one was corrupted, DistributedAsyncInsertBatch::recoverBatch would only validate the last file’s header and then sendBatch would mark the entire batch — including the intact files — as broken, losing their rows. Each file’s header is now validated individually so only the actually broken file is moved to broken/ and the surviving rows reach the remote shard. #105281 (Groene AI).
• Fix a regression in clickhouse extract-from-config --try: when the config used from_env attributes and had no include_from element (or had one pointing to a missing file), all from_env substitutions were silently dropped and returned empty strings. This broke the Docker entrypoint port discovery in 26.2.5. Closes #101704. #105283 (Groene AI).
• Fix INSERT INTO a SQLite table failing with a SQLite syntax error when the inserted value is an Enum, JSON, or AggregateFunction whose text representation contains a single quote. The output_format_values_escape_quote_with_quote setting is now honored by the corresponding serializations (previously only String and FixedString honored it). #105285 (Groene AI).
• Fix Code: 44. ILLEGAL_COLUMN: Cannot add column ...: column with this name already exists thrown by distributed queries with allow_push_predicate_ast_for_distributed_subqueries = 1 (the default) when a GLOBAL IN tuple is matched against a subquery whose projection contains duplicate column names, e.g. (x, y) GLOBAL IN (SELECT number, number FROM numbers(5)). #105290 (Groene AI).
• Fix reinterpret to Array(LowCardinality(...)) returning a confusing NOT_IMPLEMENTED error at runtime; it now returns ILLEGAL_TYPE_OF_ARGUMENT during type checking. #105301 (Raúl Marín).
• minMap/maxMap (array form) and minMappedArrays/maxMappedArrays now treat NaN consistently with ORDER BY: NaN is treated as last (returned only when all values are NaN). Previously, results depended on the position of NaN in the data due to IEEE 754 unordered comparison semantics, and disagreed with the Map-argument form of minMap/maxMap that was fixed in #100448. #105331 (Raúl Marín).
• Fix a memory leak in the per-table ColumnsDescription cache for MergeTree-family tables with Nested columns. Entries were never evicted across ALTER ADD COLUMN/DROP COLUMN cycles when share_nested_offsets = 1 (the default). #105376 (Groene AI).
• Fix BAD_ARGUMENTS exception “It’s a bug! Only integer types are supported by __bitWrapperFunc” thrown when a SELECT with a TYPE set(N) skip index has a WHERE atom whose result type is Float, BFloat16, or any other non-integer type (for example WHERE c0 + 0.1 or WHERE log(c0)). The skip index now falls back to the regular filter path in that case. #105384 (Groene AI).

​

Build/Testing/Packaging Improvement

• Stop using the system compiler-rt libraries and headers. Closes #91475. #102857 (Konstantin Bogdanov).
• Refresh the distroless Docker base image to fix OpenSSL CVEs in libssl3t64. #103583 (Rahul Nair).
• Reliably receive the GPG key during releases by chaining multiple Ubuntu keyservers, instead of timing out on keyserver.ubuntu.com. #103834 (Mikhail f. Shiryaev).
• Remove and forbid build-time CMake checks (check_*, try_compile, try_run). The compiler and toolchain are fixed, so feature detection at configure time is unnecessary and is now blocked project-wide; any version-specific behaviour must be gated on CMAKE_CXX_COMPILER_VERSION explicitly. #103980 (Alexey Milovidov).
• Bump libarchive from 3.8.6 to 3.8.7. #104047 (Robert Schulze).
• Update mongo-c-driver to 2.3.0. #104300 (Raúl Marín).
• Update LLVM dependencies for 22.x. #104381 (Joshua Carp).
• Make the embedded-client and PTY descriptor classes build on macOS and FreeBSD by using portable POSIX posix_openpt / grantpt / unlockpt and removing Linux-only #if guards. #104436 (Alexey Milovidov).
• Upgrade librdkafka to version 2.14.1. #105222 (János Benjamin Antal).

​

ClickHouse release 26.4, 2026-04-30. Presentation, Video

​

Backward Incompatible Change

• The IN operator now uses exact value semantics for Bool type: only 0 and 1 values in the set match Bool values. Previously, numeric values greater than 255 in the IN set were incorrectly clamped to true when compared against Bool, so SELECT CAST(1, 'Bool') IN (256) returned 1. Now it correctly returns 0. Closes #92980. #93115 (Ashrith Bandla).
• The H3 library has been updated to v4, which improves the precision of length, area, and other metric calculations. This change is backward incompatible because the new results differ from previous ones. #100348 (Alexey Milovidov).
• Disallows using SELECT as a bareword identifier in a WITH expression list element. #101059 (Aruj Bansal).
• This patch changes how the merge table will handle virtuals. If the underlying table contains _table or _database, these columns will be read from storage; otherwise, they will be filled after the read step using the expression step. #101742 (Mikhail Artemenko).
• The IN operator now rejects lossy Decimal conversions inside composite types (Tuple, Array, Map) as well, making its behavior consistent with top-level scalar comparisons. Previously, precision checks were enforced only for top-level scalar values: for example, CAST('33.3', 'Decimal64(1)') IN (33.33) correctly returned 0, but CAST(['33.3'], 'Array(Decimal64(1))') IN ([33.33]) incorrectly returned 1 because the lossy conversion happened inside an Array. Now both cases correctly return 0. #101812 (Nihal Z. Miaji).
• Reduced default http_max_fields from 1,000,000 to 1,000 and http_max_field_name_size from 128 KB to 4 KB to limit pre-authentication memory usage by HTTP connections. Added http_max_request_header_size and http_headers_read_timeout settings. Users who rely on the previous higher limits can restore them via settings. #103285 (Sema Checherinda).

​

New Feature

• Add automatic spilling to hash and parallel hash joins by converting them to grace hash join when memory limit is reached. This behavior is controlled by max_bytes_before_external_join. #97813 (János Benjamin Antal).
• Add Arrow Flight SQL support. #91170 (Yakov Olkhovskiy).
• Add incremental read support for Paimon table engines with Keeper-backed snapshot progress tracking, including targeted snapshot delta reads via paimon_target_snapshot_id, and extend test coverage for type mapping, partition pruning, and incremental read scenarios. #93655 (XiaoBinMu).
• The stem function is now non-experimental (previously, setting allow_experimental_nlp_functions had to be enabled). #102399 (Jimmy Aguilar Mena). You can now stem all the words/tokens in columns String, FixedString, Array([Fixed]String), Nullable, LowCardinality and Const easily with the stem function. #99137 (Jimmy Aguilar Mena).
• Implement new behavior of max_insert_block_size_rows, max_insert_block_size_bytes, min_insert_block_size_rows, min_insert_block_size_bytes in squashing under compatibility setting use_strict_insert_block_limits. #94207 (Kirill Kopnev).
• Add function arrayAutocorrelation(arr [, max_lag]) that computes the normalized autocorrelation of a numeric array for each lag. Supports integer, float, and decimal array types. #94776 (Wenyu Chen).
• A SQL function obfuscateQuery. Closes #98010. #98305 (Xuewei Wang).
• Add support for Map and JSON/Object types as dictionary attributes. Now dictionaries can store and retrieve complex types including Map(String, String), Map(String, Array(String)), JSON, and Nullable(JSON) types in both FLAT and HASHED layouts. #98627 (yanglongwei).
• Added two new MergeTree settings — replicated_fetches_min_part_level and replicated_fetches_min_part_level_timeout_seconds — that allow replicas to skip fetching freshly-inserted (unmerged) parts from peers, reducing replication overhead during heavy ingestion. #98625 (tanner-bruce).
• Add MergeTree skip index support for JSON columns using JSONAllPaths with bloom_filter, tokenbf_v1, ngrambf_v1, and text (inverted) index types, enabling granule skipping based on the set of JSON paths present in each granule. #98886 (Pavel Kruglov).
• The printf function now supports non-constant format strings, allowing different format patterns per row based on column values. #98991 (Yash ).
• Add a new projection index, commit_order, that reorganizes data in insertion order. #99004 (Mikhail Artemenko).
• Add highlight function that wraps occurrences of search terms in a text string with HTML tags (default <em>/</em>). Supports ASCII case-insensitive matching, automatic merging of overlapping matches, and custom open/close tags. #99131 (Peng).
• Implement quotas by normalized query hash to protect public ClickHouse services from abuse. 1. Support NORMALIZED_QUERY_HASH as a quota key type - separate quota buckets per unique normalized query, so CREATE QUOTA q KEYED BY normalized_query_hash tracks each distinct query independently. 2. Support QUERIES_PER_NORMALIZED_HASH as a quota resource type — limits max executions of any single normalized query within an interval, so MAX queries_per_normalized_hash = 100 prevents any one query pattern from running more than 100 times. #99586 (Alexey Milovidov).
• Users can now write join queries using the NATURAL JOIN syntax, which automatically matches on all columns sharing the same name and deduplicates those columns in the result. #99840 (Peter Nguyen).
• Support SET TIME ZONE 'tz' as alias for SET session_timezone. #99883 (phulv94).
• Added support for parameterized queries in the Web UI (play.html): query parameters like {name:Type} are detected and input fields are shown for filling in their values. #100041 (Alexey Milovidov).
• Support SQL standard VALUES clause as a table expression in FROM, e.g. SELECT * FROM (VALUES (1, 'a'), (2, 'b')) AS t(id, val). #100143 (Desel72).
• Add PostgreSQL-compatible units to the EXTRACT operator: EPOCH, DOW, DOY, ISODOW, ISOYEAR, WEEK, CENTURY, DECADE, MILLENNIUM. Also fix EXTRACT(WEEK FROM date) which previously threw an error. #100274 (Alexey Milovidov).
• Added support for SQL-standard compound interval literals with TO range qualifiers, e.g. INTERVAL '1:30' HOUR TO MINUTE. Internally decomposed into sums of intervals. #100453 (Desel72).
• Add asynchronous metrics for kernel TCP receive and transmit buffer memory (sk_rmem_alloc, sk_wmem_alloc) of HTTP connection pool sockets, reported as p50/p75/p90/p95 percentiles and totals per connection group. #100575 (Sema Checherinda).
• Added a jemalloc profiling web UI for ClickHouse Keeper, available at /jemalloc on the HTTP control port. #100606 (murphy-4o).
• Implement command SYSTEM FLUSH OBJECT STORAGE QUEUE db.table PATH 'x' for ordered and unordered modes. #100709 (Bharat Nallan).
• Added function JSONAllValues that returns all values from a JSON column as Array(String), with values serialized in text representation and ordered by their path names. Added support of text index for JSONAllValues expression on JSON columns. When a text index is created on JSONAllValues(json_column), it is automatically used to filter queries on JSON subcolumns (e.g., json_column.key1 = 'value'). #100730 (Anton Popov).
• Adds a new setting input_format_column_name_matching_mode which allows different case sensitivities for input formats. #99346 (manerone).
• Add watch command to clickhouse-keeper-client with watch support in get, exists, and ls commands. #100834 (Den Kalantaevskii).
• Added getChildrenRecursive (ListRecursive) request to ClickHouse Keeper and lsr command to clickhouse-keeper-client. This closes #99916. #100998 (Konstantin Vedernikov).
• Add new function arrayTranspose taking a two-dimensional array (matrix) and transposing it: SELECT arrayTranspose([[1, 2, 3], [4, 5, 6]]). #101214 (Vitaly Baranov).
• auto_statistics_types mergetree setting defaults to 'minmax, uniq' — minmax and uniq statistics are created automatically for all suitable columns in new tables - materialize_statistics_on_insert defaults to false — statistics are now built during merges rather than at insert time, reducing insert overhead. use SET materialize_statistics_on_insert = 1 to restore the old behavior. #101275 (Han Fei).
• Add prefer_dependency_replica refresh setting for materialized view dependency chains to reduce missing data from cross-replica replication lag. #101591 (Seva Potapov).
• Adds a hasPhrase (alias matchPhrase) function for phrase search (continuous sequences of tokens). Search is brute-force, i.e. not supported by the text index yet. #101997 (Elmi Ahmadov).
• Add s3_read_request_duration_microseconds and s3_read_request_bytes histogram metrics to observe S3 GET request connection lifetime and bytes consumed, visible in system.histogram_metrics and the Prometheus endpoint. #102058 (Sema Checherinda).
• Date and Date32 values can now be added to Time and Time64 values using the + operator, producing a DateTime or DateTime64 result. For example, SELECT toDate('2024-01-15') + toTime('14:30:25') returns 2024-01-15 14:30:25. The result is computed in the session timezone, and out-of-range results are handled according to the date_time_overflow_behavior setting. Closes #95914. #102421 (Nihal Z. Miaji).
• The text index is now GA and stays enabled regardless of the compatibility setting, preventing unexpected disabling during backup restores or when running in compatibility mode. #101518 (Nikita Fomichev).

​

Experimental Feature

• Add ALTER TABLE ... EXECUTE remove_orphan_files for Iceberg tables to identify and remove unreferenced files from object storage. #99127 (murphy-4o).
• Add query_plan_optimize_join_order_randomize setting that randomizes statistics used for join reordering, useful for testing. #100643 (Vladimir Cherkasov).
• Add AI function support to ClickHouse, allowing users to call OpenAI and Anthropic endpoints using SQL. aiGenerate is included as the first such function. #100831 (George Larionov).
• Add AI functions: aiClassify, aiExtract, and aiTranslate for utilizing LLM APIs in ClickHouse. #100832 (George Larionov).
• Added system.histogram_metric_log, a new system table that periodically snapshots all histogram metrics (e.g. S3/Azure latencies, keeper request processing stages durations). Also, the value column of system.histogram_metrics changes to Float64 as it’s more flexible and compatible with the Prometheus data model. #103046 (Miсhael Stetsyuk). The table structure is likely to be changed in future releases.

​

Performance Improvement

• ClickHouse is now able to prune entire data parts in SELECT queries based on min/max statistics. #94140 (zoomxi).
• Reduce lock contention during readonly operations on ReplicatedMergeTree tables with finished mutations. #95771 (Eduard Karacharov).
• Respect optimize_read_in_order when reading projections. Closes #89453. #95885 (Andrey Zvonov).
• Small set of improvements in Hash Join and Concurrent Hash Join. #96663 (Yarik Briukhovetskyi).
• Optimize DISTINCT transform by disabling LowCardinality columns optimization when input data is almost distinct. #97113 (Nihal Z. Miaji).
• Performance optimization for LIKE queries from #97723. Now these queries can use text indices. #98149 (Elmi Ahmadov).
• Vectorized math functions (exp, log, sigmoid, tanh) are now accelerated on AArch64 (using NEON/SVE) and on FreeBSD/Darwin, where they previously used a slower scalar fallback. #98230 (Raúl Marín).
• Queries filtering on MergeTree primary key columns with regexp alternations over literal strings, such as ^(abc-1|abc-2), can now use primary key pruning when the alternatives share a common prefix. #98988 (Yash ).
• Generalize ORDER BY ... LIMIT top-k dynamic filtering to support Nullable, String, and COLLATE types. #99033 (murphy-4o).
• Speedup hash join on Int32 and Int64 keys with small range by using a direct-index hash table. #99275 (Hechem Selmi).
• Faster discontinuous queries for LowCardinality columns with a single dictionary. #99285 (Ivan Babrou).
• Speed up var*Stable and stddev*Stable functions for Float64 columns by devirtualizing the inner loop. Note: this enables compiler optimizations (FMA/registers) that alter floating-point results at the ULP level. #99460 (Riyane El Qoqui).
• Use optimised Firedancer base58 encode for inputs of 32/64 bytes (automatic for base58Encode). Allow using optimised base58 decode if decoded result is 32/64 bytes (explicit with base58Decode('...', 32) or alike). #99461 (Joanna Hulboj).
• Enable linker section-based optimizations (-ffunction-sections, -fdata-sections, --icf=all) to reduce binary size and improve instruction cache utilization. #99474 (Alexey Milovidov).
• Fix negative scaling for short queries with aggregation on machines with many cores. When a query reads few marks, the pipeline no longer expands to max_threads after aggregation, avoiding overhead from mostly-empty streams. #99493 (Alexey Milovidov).
• Improve the performance of queries with parallel replicas by correctly selecting the reading task size. #99801 (Nikita Taranov).
• Allow prefetching when reading a remote file through the userspace page cache. #99919 (Alexey Milovidov).
• Avoid unnecessary computation of String .size subcolumn during subcolumns enumeration. #99941 (Pavel Kruglov).
• Make clickhouse-client’s progress bar less jittery when working from a hotel with clusters with a very large number of replicas. #100145 (Alexey Milovidov).
• Start MemoryWorker in clickhouse-local when page cache is enabled, so that the userspace page cache can be actually used. #100306 (Alexey Milovidov).
• Optimize queries by pushing the LIMIT clause down into the UNION ALL. #100364 (Alexey Milovidov).
• Add JIT compilation support for String and FixedString column comparisons in ORDER BY, improving merge-phase sort performance by 6–17% for string-heavy sort keys. Co-authored with @lgbo-ustc. #100577 (Raúl Marín).
• When read_in_order_use_virtual_row is enabled together with the new read_in_order_use_virtual_row_per_block setting, virtual row boundary information is now emitted after each block read from MergeTree, allowing the merge to reprioritize sources mid-stream for parts whose data is fully filtered out by WHERE/PREWHERE/JOIN. Close #99945. #100603 (Vladimir Cherkasov).
• Faster Float-to-String conversion for large integer values by extending the itoa fast path with dragonbox-compatible rounding. #100649 (Raúl Marín).
• Replace dragonbox with zmij for 1.5x-3x faster Float-to-String conversion. #100650 (Raúl Marín).
• Faster Int128/UInt128 to string conversion by replacing software division with Barrett reduction and unrolling the conversion loop. #100671 (Raúl Marín).
• Avoid spawning redundant threads in uniqExact parallel merge. #100686 (Jiebin Sun).
• Add batch parallel merge for uniqExact. #100687 (Jiebin Sun).
• Better parallelization of queries with simple views (with underlying MergeTree table) executed with parallel replicas. #100815 (Igor Nikonov).
• Implements support of parallel replicas over simple views (including eligible UNION ALL views over MergeTree tables) when parallel_replicas_allow_view_over_mergetree=1. This allows to parallelize view’s outer query instead of inner one which increases query parallelization across nodes. #100958 (Igor Nikonov).
• Optimise reading in order of the primary key for full_sorting_merge when filters with IN are present in the query plan. #101261 (Nikita Taranov).
• Optimization allocations/deallocations by caching sampling settings instead of traverse all memory tracker hierarchy. #101267 (Azat Khuzhin).
• Fix significant INSERT performance regression when deduplicate_insert = 'enable' (default since 26.2) by deferring data hash computation from squashing to the sink and using batch column hashing via updateHashWithValueRange, reducing overhead from ~2.5s to ~0.5s for 5M rows with 22 columns. #101494 (Sema Checherinda).
• Reduce profiled lock overhead by using try_lock to avoid timing uncontended acquisitions and removing hold-time measurement. #101502 (Antonio Andelic).
• Replace hand-written AVX-512 intrinsics in arrayDotProduct with platform-independent auto-vectorizable loops, adding AVX2 and ARM NEON support. #101571 (Peng).
• Improve performance in INSERT VALUES for Map, Array, and Tuple columns when values are passed as escaped strings (e.g. '{\'key\':1}'), avoiding unnecessary fallback to the SQL expression parser. #102119 (Joanna Hulboj).
• Fixed excessive RabbitMQ table engine CPU usage. #102711 (Jaap Elst).
• The JOIN order optimizer now infers transitive equi-join predicates from existing join conditions. For example, given A.x = B.x AND B.x = C.x, the equivalence A.x = C.x is recognized, allowing the optimizer to consider direct joins between transitively-connected tables. This can improve plan quality for star and snowflake schemas where dimension tables connect through a shared fact table. The feature is controlled by the new enable_join_transitive_predicates setting (off by default). #98479 (Alexander Gololobov).
• Optimize TRUNCATE DATABASE TABLES LIKE by pre-cancelling merges in parallel. #98597 (Shaohua Wang).
• Add monotonicity support for multiply, enabling primary key pruning for key * constant expressions. #98983 (Amos Bird).
• Cache dictionaries no longer take an exclusive lock in hasKeys; this reduces lock contention by using a shared lock for cache reads. #100796 (liuguangliang).
• Inline VIEW subquery in the query tree to allow more optimisations to be applied to the VIEW. #100830 (Dmitry Novik).
• Optimize cache loading on server startup. #101500 (Kseniia Sumarokova).
• Implement lazy column materialization for ReplacingMergeTree with FINAL in case the predicate is selective enough. #101647 (Nikolai Kochetov).
• Re-enable the optimize_rewrite_array_exists_to_has optimization (off by default since 23.10). It rewrites arrayExists(x -> x = elem, arr) into the much faster has(arr, elem) and now correctly skips the rewrite when the array element type and elem are not compatible for has (e.g. Date vs String), so previously breaking queries continue to work. Closes #71431. #100944 (Alexey Milovidov).

​

Improvement

• Improved EXPLAIN PLAN pretty=1 output: print top-level query output columns, show join relation labels/symbols with estimated result rows and locality, and include per-step output columns for join/source steps. The changes cover Information Deficit part from #98117. #99462 (Kirill Kopnev).
• Add MergeTree table setting share_nested_offsets (default true). When set to false, Array columns with dotted names (e.g. n.a, n.b) are treated as independent columns instead of sharing offset files and validating equal array sizes as part of legacy Nested semantics. #98416 (Amos Bird).
• Users can now specify multiple authentication methods in users.xml/yaml configuration (in SQL it was always possible). #91998 (Flip-Liquid).
• Auto reload Raft inter-node connections which are using TLS. #93455 (Evgeny).
• Extend cast_keep_nullable to work with Dynamic/JSON types. When set, casting NULL from types that can be Nullable will return NULL, otherwise NULL will throw CANNOT_INSERT_NULL_IN_ORDINARY_COLUMN error. #96504 (Seva Potapov).
• Reduced the memory footprint from the internal data structures (ISerialization objects) by introducing an object pool. #96563 (Nikita Mikhaylov).
• Add support of password and identity fields to keeper-client XML config. #96800 (Grigorii Sokolik).
• Improve Iceberg writes for the unity catalog. #98162 (Konstantin Vedernikov).
• Add setting finalize_projection_parts_synchronously to allow synchronous finalization of projection parts during INSERT, reducing peak memory usage for tables with many projections while preserving existing async behavior by default. #98228 (Amos Bird).
• Add projections_duration_ms column to system.part_log that records per-projection merge/rebuild duration in milliseconds. #98292 (Amos Bird).
• Improve canceling queries using ExpressionTransform and NumbersRangedSource by KILL QUERY and cancel query (Ctrl+C) in clickhouse-client. #98908 (Roman Vasin).
• Replace the hardcoded source_table_engines list with a runtime lookup via StorageFactory::getAllStorages(). This adds access checks for some missing table engines and closes #71544. #98984 (pufit).
• Add a setting to control type mismatch behavior for Variant and Dynamic (throw or return null). #99085 (Bharat Nallan).
• Improve Iceberg and Spark compatibility: fix inconsistent path handling caused by mixed usage of storage paths and metadata paths; enforce that Iceberg tables write down a table location that is either a URL or an absolute path; add a fallback for counting file sizes in Azure because some ClickHouse readers don’t support byte counting after traversal; handle version-hint.txt in a manner compatible with Spark; introduce type-level abstractions that make it harder to mix up path types in the future; add tests for Azure and Local that verify cross-engine interoperability without intermediate uploading/downloading; fix usage of position deletes, which previously relied on path inference heuristics where that approach is inappropriate. #99163 (Daniil Ivanik). #100420 (Daniil Ivanik).
• Fix possible race condition in IPartitionStrategy::cached_result introduced in https://github.com/ClickHouse/ClickHouse/pull/92844. #99400 (Arthur Passos).
• Users can now write ClickHouse Interval datatypes in the Arrow Format. #99519 (Peter Nguyen).
• Adds native support for importing and exporting UUID data types in Arrow and Parquet formats. Users can now directly query and transfer UUID data between ClickHouse and other data tools without requiring manual string conversions or workarounds. Automated logical inference for top-level UUIDs, and support for explicit schema hint for nested UUIDs. #99521 (Ivan).
• Support 7z archives on object storage. Closes #70968. #99600 (Alexey Milovidov).
• Add ObjectStorageListedObjects, ObjectStorageGlobFilteredObjects, ObjectStoragePredicateFilteredObjects, and ObjectStorageReadObjects ProfileEvents for introspection of object storage (S3, Azure, etc.) file listing and reading pipeline. #99778 (Sema Checherinda).
• Fix merge table function failing with UNKNOWN_IDENTIFIER error when querying columns not present in all underlying distributed/remote tables. #99833 (Alexey Milovidov).
• Now we include commit time in total mutation execution time metric for ReplicatedMergeTree. It was lost after #96376. #99936 (alesapin).
• Add a write-ahead log for blob objects pending removal in MetadataStorageFromDisk, improving durability and consistency between metadata and remote object storage when objects are deleted. #100019 (Maksim Kita).
• Disable AI SQL generation (?? command) in the embedded client (SSH and WebSocket protocols) to prevent access to the server’s environment variables. #100290 (Alexey Milovidov).
• Change the interface for Iceberg inserts with the catalog. Deprecate settings: storage_catalog_type, storage_aws_access_key_id, etc. #100334 (Konstantin Vedernikov).
• Render tabs as 4 spaces when pasting into clickhouse-client. Closes #100405. #100416 (Raúl Marín).
• Avoid scanning the whole remote data lake catalog for “Maybe you meant …” table hints when show_data_lake_catalogs_in_system_tables is disabled. #100452 (Alsu Giliazova).
• Apply distributed_index_analysis_min_indexes_bytes_to_activate after partition pruning. #100477 (Azat Khuzhin).
• Fix assertion failure in Parquet bloom filter push down when using empty IN/NOT IN clauses. #100543 (zoomxi).
• MinMax column statistics now store the minimum and maximum values as Field (typed) instead of Float64. The serialized format includes the column type name alongside the values. The statistics file version is bumped to V2; files written by older versions require re-materialization (ALTER TABLE … MATERIALIZE STATISTICS ALL). fix #53140. #100605 (Han Fei).
• Update cppkafka to include fix for Consumer close deadlock. #100612 (Azat Khuzhin).
• Object information used for parsing data files in Iceberg now contains the number of file rows and file size in bytes parsed from manifest file. #100645 (Daniil Ivanik).
• Add use_separate_cache_arena configuration parameter to be able to control separation of the cache memory arena. #100664 (Seva Potapov).
• Adds native support for importing Apache Arrow’s StringView and BinaryView data types into ClickHouse String columns, improving compatibility for Arrow-based ingestion. #100762 (Ivan).
• A few Keeper server settings are now hot-reloaded if config file is changed at runtime: max_requests_batch_size, max_requests_batch_bytes_size, max_request_size, quorum_reads. #100773 (Michael Kolupaev).
• Increment profile events MemoryAllocatedWithoutCheck/MemoryAllocatedWithoutCheckBytes in release build. #100899 (Pavel Kruglov).
• Cgroupv2 memory tracking now excludes slab_reclaimable from kernel memory, giving a more accurate measure of non-reclaimable memory usage. #100901 (Antonio Andelic).
• use_partition_pruning = 0 now also disables MinMax index pruning and count optimization on partition key columns, in addition to disabling pruning based on partition keys. #100904 (Nihal Z. Miaji).
• pretty=1 in EXPLAIN [PLAN] now prints expressions in a human readable format. #100927 (Kirill Kopnev).
• accurateCastOrNull and accurateCastOrDefault now support Tuple target types, including nested Tuples with Nullable elements. Previously these functions rejected Tuple targets because Tuple could not be inside Nullable. Closes #100820. #100942 (Nihal Z. Miaji).
• Fix chart duplication in Play UI when switching between light and dark themes. #101058 (Alexey Milovidov).
• Update chdig to v26.3.1 (perfetto UI, sparklines to summary for CPU/Memory/Merges/Queries), system.warnings, regexp search in logs). #101092 (Azat Khuzhin). Update chdig to v26.4.3 (perfetto improvements, fixes for sharing via pastila.nl, flamegraph diffs, change settings in realtime). #103145 (Azat Khuzhin).
• You can now have a trailing comma in the WITH clause before a SELECT query. #101093 (Aruj Bansal).
• Add compress_per_column_in_compact_parts MergeTree setting to control how compressed blocks are organized within Compact parts. When true (default, preserving current behavior), each column starts a new compressed block, allowing selective decompression. When false, all columns within a granule are packed into the same compressed block, improving compression ratio and read performance for workloads that always read all columns. #101114 (Amos Bird).
• Show table info balloon in Play UI only when hovering over the table name, not the entire row. #101118 (Alexey Milovidov).
• Add engine-specific icons and improve table list UX in the Play UI sidebar. #101134 (Alexey Milovidov).
• Support Nullable(Tuple) for Arrow, ArrowStream, ORC, legacy Parquet formats. #101272 (Nihal Z. Miaji).
• Display TOTALS row as a table footer in the web UI (play.html). #101286 (Alexey Milovidov).
• Support multi-query mode in the web UI (play.html): run multiple queries at once with parallel execution of SELECT-like queries and per-query result display. #101290 (Alexey Milovidov).
• Fix column resize in play.html web UI after the result table was refactored into a web component. #101295 (Alexey Milovidov).
• Add ability limit amount of jemalloc profile flushes on MEMORY_LIMIT_EXCEEDED per time interval. #101396 (Azat Khuzhin).
• Added keeper settings nuraft_streaming_mode (by default false), nuraft_max_log_gap_in_stream, nuraft_max_bytes_in_flight_in_stream. Closes #90743. #101427 (Kseniia Sumarokova).
• Added CGroupMemoryUsedWithoutPageCache async metric that reports cgroup memory usage excluding both the kernel OS page cache and the ClickHouse userspace page cache, mirroring MemoryResidentWithoutPageCache. Also clarified the CGroupMemoryUsed metric description. #101513 (Francesco Ciocchetti).
• Add parser-level syntactic sugar for the SQL standard OVERLAY function syntax. The overlay function already exists; this adds support for the keyword-based form using PLACING, FROM, and FOR as separators. #101681 (Desel72).
• Added column alias INDEX_LENGTH to system table information_schema.tables, analogous to existing uppercase aliases in this table. #101705 (Robert Schulze).
• System table information_schema.tables now ignores inactive table parts. This makes the shown table size values more realistic. #101706 (Robert Schulze).
• The ngrams function now rejects invalid ngram lengths. Example: SELECT ngrams('abc', 0) now returns an error. #101922 (Robert Schulze).
• A follow-up for #91820 and #90837: filter unsupported algorithms from the error message; run FIPS-specific tests in FIPS builds. #102067 (Mikhail f. Shiryaev).
• Limit cell height to three lines in the Web UI (play.html), with expandable cells on click. #102154 (Alexey Milovidov).
• Added new option that allows to force (virtual/path) style for S3 endpoints. Resolves #82019; #76007 Continue of https://github.com/ClickHouse/ClickHouse/pull/83168. #102378 (Konstantin Vedernikov).
• The restore_replace_external_engines_to_null setting now also skips restoring databases with external engines (e.g. DataLakeCatalog, MySQL, PostgreSQL, S3) instead of failing or initiating external connections. #102400 (Nikita Fomichev).
• Add text index analysis support for the hasPhrase function via the HINT mode. #102438 (Elmi Ahmadov).
• Treat STATISTICS as read-only in ColumnDependency to fix LOGICAL_ERROR during MATERIALIZE STATISTICS ALL. #102627 (Konstantin Bogdanov).
• Create and populate system.asynchronous_metric_log in keeper-as-server mode. #102664 (Miсhael Stetsyuk).
• Add default_system_log_flush_policy.skip_alias_columns config option to allow omitting ALIAS columns from system log tables, fixing S3-backed system logs that reject ALIAS columns. #102669 (Miсhael Stetsyuk).
• Don’t enable auto statistics for system tables. They rarely have chance to use them. #102862 (Han Fei).
• Support array tokenizer for the LIKE optimization. #102880 (Elmi Ahmadov).
• Send MemoryAllocatedWithoutCheck even in release builds. #103064 (Azat Khuzhin).
• Expose per-thread untracked_memory in system.stack_trace. #103065 (Azat Khuzhin).

​

Bug Fix (user-visible misbehavior in an official stable release)

• Fix Block structure mismatch in stream error caused by unnecessary columns returned from Lazy materialization. Fixes #95191. #96682 (Nikolai Kochetov).
• Fix a logical error with data masking policy query with ON CLUSTER. #97594 (Bharat Nallan).
• Fix a bug when using Unity catalog on top of GCS. #98456 (Melvyn Peignon).
• DataLakeCatalog now respects the server’s http_forbid_headers configuration when validating the auth_header setting. #98827 (Michael Anastasakis).
• Fix N+1 HeadObject calls for S3 brace-expansion globs. #99219 (Konstantin Bogdanov).
• Validate setting changes in create queries when the engine itself also supports settings. #99279 (János Benjamin Antal).
• Fixed ALTER TABLE UPDATE/DELETE failing with Missing columns error when a table has a MATERIALIZED column whose expression depends on an EPHEMERAL column. #99281 (Yash ).
• Credentials in JDBC, ODBC, and NATS connection strings are now masked in query logs and SHOW CREATE output, preventing accidental exposure of sensitive information. For URI-style connection strings (e.g. {scheme}://{user}:{password}@{host}), only the password portion is masked while the rest remains visible for easier debugging. The nats_token setting is now also masked. #99344 (János Benjamin Antal).
• Fix parseDateTimeBestEffort incorrectly parsing words starting with month prefixes in DD-month-YYYY format. Closes #99345. #99350 (Pavel Kruglov).
• Fix ignoring of TABLE_UUID_MISMATCH for non analyzer. #99380 (Azat Khuzhin).
• Fix a bug where explicit settings sent alongside compatibility in the same request could be silently ignored when their value matched the server default. #99402 (Raufs Dunamalijevs).
• Fixes cases where numbers with leading zeros in hive partitioning path were causing errors. Fixes #98801. #99458 (Yarik Briukhovetskyi).
• Fix heap-use-after-free when a table is dropped concurrently with a running read query (19 occurrences in CI over the last 90 days). #99483 (Alexey Milovidov).
• Fixed a bug in Keeper where a read request could get stuck (causing session to time out) if a different unrelated session on the same server was closed at just the wrong moment. #99484 (Michael Kolupaev).
• Validate column structure before applying patches. #99531 (Seva Potapov).
• Fix vertical merge rows_sources assertion failure when SYSTEM STOP/START MERGES toggles rapidly during merge of table with Dynamic columns. #99532 (Alexey Milovidov).
• Fix incorrect partition pruning for toWeek() that caused queries with WHERE toWeek(date, mode) = N to return empty results for weeks 49-52 on tables partitioned by toYYYYMM(date). #99542 (Takumi Hara).
• Fix exception in functions operating on ColumnReplicated with unreferenced rows produced by JOIN. #99564 (Hechem Selmi).
• Fix CLEAR COLUMN not rebuilding projections and not reevaluation materialized columns that depend on the cleared column, which could cause exceptions or data corruption during subsequent merges. #99565 (Desel72).
• Fixed an exception (Bad get: has Tuple, actual type String) in ConditionSelectivityEstimator when a query uses IN with a single scalar query parameter (e.g. WHERE col IN ({p:String})) on a table that has column statistics and use_statistics is enabled. #99614 (Ilya Yatsishin).
• Part with unknown projections should not be marked as lost forever. #99623 (Sema Checherinda).
• Fix a rare logical error exception during vertical merge when SYSTEM STOP MERGES and SYSTEM START MERGES are executed concurrently. #99628 (Desel72).
• Fix dangling reference in injectRequiredColumns causing crash during merge. #99679 (Tuan Pham Anh).
• Fix undefined behavior in Avro format reader when reading numeric values that overflow the target column type. Now queries fail on overflows instead of silently producing incorrect values. #99697 (asyablue22).
• Fix parsing of shell-style quotes in arguments for the executable table function. #99794 (Nikita Semenov).
• Fix false-positive abort in NativeReader when deserializing a Native format stream with a row-count mismatch: changed from LOGICAL_ERROR to INCORRECT_DATA so the error is handled as a data error rather than triggering abort() in sanitizer/debug builds. #99822 (Rahul Nair).
• Fix process abort in Tuple column deserialization when the serialization kind in the binary stream is DETACHED. #99823 (Rahul Nair).
• Fix false LOGICAL_ERROR exception during filesystem cache dynamic resize due to a race condition in SLRU sub-queue promotion. #99850 (Alexey Milovidov).
• Fix async insert queries reporting zero written_rows, read_rows, and result_rows in query_log and client output. #99879 (Sema Checherinda).
• Fix exception “Bad cast from type X to Y” in KILL QUERY when the internal query against system tables returns columns wrapped in ColumnConst. #99881 (Alexey Milovidov).
• Fix logical error with correlated subquery within untuple argument. #99917 (Vladimir Cherkasov).
• Fixes an exception when calling right, rightUTF8, or other substring functions with a length of INT64_MIN (-9223372036854775808), which previously caused undefined behavior due to integer overflow. The functions now correctly report an ARGUMENT_OUT_OF_BOUND error. #99934 (Jimmy Aguilar Mena).
• Now ClickHouse should properly handle spark-style tables (where we have full absolute path for each file or relative path to common table path). Fixes #92348. #99935 (alesapin).
• Fix “Inconsistent AST formatting” exception for ALTER TABLE ... MODIFY QUERY with nested subqueries containing SETTINGS when the ALTER itself also has SETTINGS. #99938 (Nikita Mikhaylov).
• Revert #97114 “Move join step row estimation before check for 1 child” due to suspected performance regression. #99957 (Alexander Gololobov).
• Fixes a bug where ClickHouse could skip files if the Content-Length header was missing in their HEAD request response (for example, because of decompressive transcoding in GCS). #99971 (Yarik Briukhovetskyi).
• Fix assertion failure (exception in debug builds, incorrect results in release builds) when multiplying NumericIndexedVector aggregate states by an even integer constant, caused by self-XOR on aliased Roaring bitmaps in pointwiseAddInplace. #99976 (Desel72).
• Prevent Unexpected return type exception in legacy filter pushdown through chained JOIN USING when key types change after join conversions. #99999 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception “Unexpected node type for table expression … Actual IDENTIFIER” when a scalar subquery is used inside an unresolved table function argument, e.g. SELECT * FROM remote('localhost', view(SELECT 2 AS x), concat(x, (SELECT 1))). #100014 (Alexey Milovidov).
• Fixed INSERT with VALUES failing when the data was followed by a trailing SQL comment (-- or /* */) on the next line. The comment is now skipped instead of being parsed as another row. #100016 (Pratima Patel).
• Fix exception in arrayRemove when comparing tuples with NULL components. #100017 (Alexey Milovidov).
• Fix cross-user data leak in system.asynchronous_inserts: any user with SELECT on the table could see pending async insert entries belonging to other users. Entries are now filtered by the current user, unless the user has the SHOW_USERS privilege. #100024 (Shaohua Wang).
• Fixes case where Time64 to UInt64 cast could clamp values to 24 hours. #100025 (Yarik Briukhovetskyi).
• Fix local server crash when CREATE DICTIONARY has a definition with list value containing non-existing function. #100036 (Yakov Olkhovskiy).
• Fix CSV, MsgPack format not being able to parse Nullable(Tuple) properly. Closes #99753. #100038 (Nihal Z. Miaji).
• Fix CREATE VIEW failing with UNKNOWN_IDENTIFIER when using a WITH function-expression alias (e.g. tuple(...)) as the right-hand side of IN. #100042 (Peng).
• Fix timeseries aggregate functions (e.g. timeSeriesResampleToGridWithStaleness) failing with ILLEGAL_TYPE_OF_ARGUMENT when used with initializeAggregation or AggregatingMergeTree under parallel replicas. #100053 (Alexey Milovidov).
• Make correctly processing negative values inside NumericIndexedVectorDataBSI. #100086 (Daniil Ivanik).
• Fix accurateCastOrDefault and to*OrDefault functions not preserving Const column type for constant inputs. #100132 (Alexey Milovidov).
• Omitted query parameters with LowCardinality(Nullable(T)) type now correctly default to NULL, same as Nullable(T). #100144 (Denys Melnyk).
• Fix use-of-uninitialized-value in StringSearcher.h. #100225 (Konstantin Bogdanov).
• Allow cancellation of scalar subqueries and other analysis-time pipelines via Ctrl+C. Previously, pressing Ctrl+C during a long-running scalar subquery had no effect until the subquery completed. Also fix the progress bar and JSON statistics to correctly report rows read during scalar subquery execution, both in clickhouse-client and clickhouse-local. Co-authored with @YjyJeff. #100230 (Raúl Marín).
• Fixed a LOGICAL_ERROR exception in queries involving Dynamic columns with cross joins and runtime filters, caused by ColumnVariant::filter sharing variant column pointers instead of cloning them in the hasOnlyNulls optimization path. Closes https://github.com/ClickHouse/ClickHouse/pull/100147. #100234 (Pavel Kruglov).
• Fixed array of variant bug which may reinterpret data type upon calling arrayFirst/arrayLast function. For example, previously Array(Variant(Date, Bool)) is converted to Bool when the actual underlying variant type is Date. #100255 (timothygk).
• A few minor changes to functions: h3 functions now validate boundaries better; readWKB checks the size limits (a new setting, max_wkb_geometry_elements); random generator functions limit the maximum iterations in their computations. A follow-up for #93543. #100270 (Alexey Milovidov).
• Fixed an issue where cutURLParameter could incorrectly skip parameters when they appeared as substrings of other parameters. #100280 (Nikita Semenov).
• Fix exception when Iceberg metadata file path setting contains a null byte. #100283 (Alexey Milovidov).
• Fixed the quadratic number of run queries when distributed_index_analysis is used with predicates containing IN subqueries. #100287 (Anton Popov).
• Fix “Block structure mismatch” exception when using GROUP BY ... WITH TOTALS HAVING combined with UNION DISTINCT and nullable expressions. #100293 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception in estimateCompressionRatio when block_size_bytes parameter is extremely large. #100298 (Alexey Milovidov).
• Fix “Inconsistent AST formatting” exception in debug builds when using GROUP BY CUBE(...) WITH ROLLUP or similar combinations. #100376 (Alexey Milovidov).
• Fix exception when creating a view with column aliases and SELECT * or EXCEPT/INTERSECT queries. #100386 (Alexey Milovidov).
• Fix DROP TABLE hanging indefinitely on Kafka engine tables when consumers are stuck in a rebalance after a heartbeat error. #100388 (Alexey Milovidov).
• Fix ReadBuffer is canceled. Can't read from it. exception in backup/restore operations using zip archives. #100400 (Alexey Milovidov).
• Fix TOO_MANY_ROWS exception for SELECT count() queries with max_rows_to_read / force_primary_key when data is split across multiple parts with non-aligned granule boundaries. #100408 (Alexey Milovidov).
• Fix system.completions to correctly filter databases, tables, and columns by access rights in all grant combinations: per-table, per-db, and per-column revoke. #100432 (Shaohua Wang).
• Fix SEGFAULT in NuRaft due to race condition. #100444 (Pablo Marcos).
• min/max/argMin/argMax now treat NaN consistently with ORDER BY: NaN is always skipped (returned only when all values are NaN). Previously, results depended on NaN position in the data due to IEEE 754 unordered comparison semantics. #100448 (Raúl Marín).
• Fixed a copy-paste bug where delta_lake_snapshot_end_version set without delta_lake_snapshot_start_version was silently ignored instead of producing a BAD_ARGUMENTS error. #100454 (Mohammad Lareb Zafar).
• StorageRabbitMQ::shutdown is not idempotent (it unconditionally accesses weak pointers and then destroys the corresponding shared pointers), but is now called twice: once in StreamingStorageRegistry and then in DatabaseCatalog. This fix makes the method idempotent and adds defensive null checks. #100455 (Miсhael Stetsyuk).
• Fix LOGICAL_ERROR exception when using accurateCastOrNull with QBit target type. #100470 (Raufs Dunamalijevs).
• Fix LOGICAL_ERROR exception “Stream … not found” when inserting into a table with nested Array(JSON) columns in wide parts with optimize_on_insert=0. #100475 (Pavel Kruglov).
• Validate file entry paths in backup metadata to reject path traversal, absolute paths, and empty names during RESTORE. #100483 (Pablo Marcos).
• Fix LIMIT m OFFSET n WITH TIES syntax not working. This syntax is equivalent to LIMIT n, m WITH TIES which already worked. #100491 (Nihal Z. Miaji).
• Fix exception “No set is registered for key” when using IN with Nullable(Tuple) columns that have named fields and LowCardinality elements. #100523 (Alexey Milovidov).
• Fix heap-buffer-overflow in usearch sorted_buffer_gt::insert() that could crash or silently corrupt memory during vector similarity search. #100537 (Dustin Healy).
• Fix EXECUTE AS ignoring FORMAT and INTO OUTFILE clauses specified in the query. #100538 (pufit).
• Fix inconsistent AST formatting for SAMPLE with query-level OFFSET. Closes #100576. #100579 (Pavel Kruglov).
• Fix polaris catalog with azure. Since 25.12 this catalog with azure started to add bucket at the beginning of the path. For example abfss://polaris-polaris@<some_url>.windows.net/polaris-polaris/<other-path> instead of abfss://polaris-polaris@<some_url>.windows.net/<other-path>. This PR cuts bucket in the path. #100583 (Konstantin Vedernikov).
• Fix type mismatch exception in transform when default column is const on some blocks. Closes #100574. #100616 (Pavel Kruglov).
• Fixes NOT_FOUND_COLUMN_IN_BLOCK cases where the projection SELECT part has columns that do not exist in the original SELECT part of the query. Closes #100194. #100623 (Yarik Briukhovetskyi).
• Validate Npy format shape dimensions against file size and overflow limits to prevent denial of service from crafted .npy files with unreasonably large dimensions. Also reject empty shapes and cap per-row memory to 2 GiB. #100625 (Raúl Marín).
• Fix session_timezone being ignored when parsing DateTime values during async inserts (TCP) and all inserts over HTTP. #100647 (Sema Checherinda).
• Allow passing sharding key to cluster() and clusterAllReplicas() table functions when using a table function as the source (e.g. cluster('name', view(...), sharding_key)). #100665 (Sergey Veletskiy).
• Fix server crash (assertion failure) when using parametric aggregate functions with Array combinator and NULL arguments, such as quantileIfArrayArray(0.5)([[NULL]], [[1]]). #100679 (nerve-bot).
• Fix exception when computing common supertype for empty and non-empty tuples with use_variant_as_common_type enabled. #100699 (Antonio Andelic).
• Server no longer fails to start when an Azure blob storage disk is configured but the endpoint is temporarily unreachable (e.g. DNS failure). #100701 (Raúl Marín).
• Fix undefined behavior in positiveModulo when the unsigned divisor does not fit in the signed result type. #100705 (Raúl Marín).
• Fix server crash (logical error “Unexpected return type from __topKFilter”) when use_top_k_dynamic_filtering is enabled and the ORDER BY column has Dynamic or Variant type. #100742 (Groene AI).
• Fix server crash when using has() function with PREWHERE/WHERE on a Tuple key containing LowCardinality elements. #100760 (Groene AI).
• Fix file_offset_of_buffer_end <= getFileSize() assertion failure (exception in debug builds) when reading from Log or StripeLog tables on S3 object storage with concurrent writes. #100763 (Alexey Milovidov).
• Fix an exception in the statistics selectivity estimator when a WHERE clause contains a function expression (e.g. toDecimal64(col, 3)) on a table with statistics enabled. The estimator now skips such predicates instead of attempting an invalid type cast. #100764 (Han Fei).
• Fixes a rare case where join with reordering can produce a wrong result. #100790 (Yarik Briukhovetskyi).
• Fix incorrect AggregateFunction argument types in optimized trivial count, which caused NUMBER_OF_ARGUMENTS_DOESNT_MATCH exception when querying expressions like count(v0 + v1) on distributed tables. #100794 (YjyJeff).
• Some catalogs can show some secrets in the SETTINGS section of select * from system.databases query result. This PR prevents such behavior. #100800 (Konstantin Vedernikov).
• Fix undefined behavior (signed integer overflow) in toStartOfInterval when using Week, Quarter, or Year intervals with an origin argument and extreme interval values. #100817 (Raúl Marín).
• Fix If, Distinct, DistinctIf, IfState aggregate function combinators with Tuple return type and one or more Nullable argument not being able to read older serialized states after introduction of Nullable(Tuple). Closes #98917. #100826 (Nihal Z. Miaji).
• Fix segfault in s3Cluster and distributed queries due to connection pool use-after-free. #100837 (Konstantin Bogdanov).
• Fix null pointer dereference segfault when loading dictionaries during server shutdown. Context::getUserDefinedSQLObjectsStorage (dereferences user_defined_sql_objects_storage) is called by dictionary threads concurrently with the main thread calling Context::shutdown (sets user_defined_sql_objects_storage to null). We need to make sure we disable future updates in the dictionaries loader, kill the currently running dictionary queries and join the dictionary loading threads - all before running Context::shutdown. Similar to what we do with normal queries. #100839 (Miсhael Stetsyuk).
• Fix buffer overflow in ULIDStringToDateTime when input contains non-ASCII bytes. #100843 (Konstantin Bogdanov).
• Fix crash (LOGICAL_ERROR) when querying a Merge table (or merge() table function) that wraps multiple tables including a Distributed table, with distributed_group_by_no_merge=1 enabled. #100859 (Groene AI).
• Cast_keep_nullable when enabled will not throw when casting dynamic null to variant. #100864 (Seva Potapov).
• Fix clickhouse-keeper-client get, exists, and ls commands printing duplicate watch_id error messages to stdout instead of stderr. #100893 (Mohammad Lareb Zafar).
• Fix exception in intDiv/intDivOrZero on arrays of nullable tuples, e.g. SELECT intDiv([divide((1, 2), ... AND NULL)], 2). #100895 (Raúl Marín).
• Evaluate engine arguments for StorageAlias before storing the definition, so that expressions like currentDatabase() are resolved to literals before being saved to the database. #100902 (Nikolay Degterinsky).
• Fix processAndOptimizeTextIndexFunctions when query_plan_merge_expressions = 0, where ExpressionStep is directly above ReadFromMergeTree. Fixes #100879. #100909 (Jimmy Aguilar Mena).
• Update replxx to include fix for out-of-bounds access in do_complete_line. #100925 (Azat Khuzhin).
• Fix wrong results when JOIN with shard-by-PK optimization uses query condition cache and some parts are filtered out by cached conditions. #100926 (Groene AI).
• Fix divide and intDiv returning ILLEGAL_DIVISION when used in filter expressions during index analysis in some cases. #100928 (Nihal Z. Miaji).
• Fixed “Target table doesn’t exist” errors for materialized views with inner tables during async startup, caused by incorrect startup dependency ordering. #100946 (Nikolay Degterinsky).
• Fix undefined behavior (signed integer overflow) in parseDateTimeBestEffort when parsing datetime strings with more than 18 fractional-second digits. #100948 (Vasily Chekalkin).
• Fixed a crash when using a text search index with an IN clause containing a tuple subquery, e.g. WHERE (id, str) IN (SELECT (id, str) FROM ...), or when the number of columns in the subquery does not match the tuple on the left side of IN. #100959 (Anton Popov).
• Fixed crash when building a polygon dictionary from a MergeTree table that uses sparse columns serialization. #100964 (Anton Popov).
• Fix logical error “Invalid action query tree node” when using INTERSECT ALL / UNION ALL with constant-folded expressions. #100977 (Alexey Milovidov).
• Fix sumCountOrDefault aggregate function with one or more Nullable argument not being able to read older serialized states after introduction of Nullable(Tuple). Closes #100882. #101021 (Nihal Z. Miaji).
• Fix crash (Logical error: isConst/isSparse/isReplicated assertTypeEquality) in merge algorithms when lazy column replication (enable_lazy_columns_replication) produces ColumnReplicated columns that flow into merge-sort pipelines with late-arriving inputs. #101036 (Groene AI).
• Fix incorrect UNKNOWN_IDENTIFIER error when the same alias is used for multiple expressions in SELECT; the correct MULTIPLE_EXPRESSIONS_FOR_ALIAS error is now reported. #101040 (Alexey Milovidov).
• Fix ALIAS columns with DateTime/DateTime64 types not applying timezone conversion when the declared timezone differs from the expression timezone. #101043 (Alexey Milovidov).
• Fix row policies not being recorded in query_log for views, subqueries, and INSERT ... SELECT. Even though row policies were applied during query planning, they were not propagated from sub-planners to the parent planner for logging. Row policies (for logging only) are now kept in QueryAccessInfo so that both planners and sub-planners can populate them. #101044 (Narasimha Pakeer).
• Fix exception in DirectJoinMergeTreeEntity when pipeline blocks contain ColumnConst columns that are merged with regular columns. #101046 (Alexey Milovidov).
• Fix spurious space in CTE column alias formatting (WITH t (a, b) → WITH t(a, b)). #101049 (Alexey Milovidov).
• Fix remote/cluster table functions failing with nested table functions like merge when the analyzer is enabled. #101055 (Alexey Milovidov).
• Fix OFFSET being applied twice in distributed queries when prefer_localhost_replica=1, producing fewer rows than expected. #101071 (Nihal Z. Miaji).
• Fix crash when using the Regexp format with an invalid regular expression in format_regexp setting. #101074 (Nihal Z. Miaji).
• Fix “Illegal type Decimal64 of start parameter” error for timeseries aggregate functions when using serialize_query_plan=1 with parallel replicas. #101083 (Groene AI).
• Fix exception in optimizeLazyMaterialization when a projection with PREWHERE is used with ORDER BY ... LIMIT. #101115 (Anton Popov).
• Fix server crash (SIGABRT) when using aggregate functions with the internal-only Null combinator (e.g. sumNull, avgNull) and aggregate_functions_null_for_empty = 1 setting enabled. #101147 (Groene AI).
• Fix a use-after-free in the filesystem cache write path that could cause reads from freed memory when logging completed file segments (detected by MemorySanitizer in BuzzHouse). #101161 (Groene AI).
• Fix server crash with “Trying to attach external table to a ready set without explicit elements” when distributed index analysis encounters a GLOBAL IN predicate whose set was built without explicit elements. #101178 (Groene AI).
• Fix MAX/MIN aggregate functions on Decimal columns returning incorrect results when JIT compilation is enabled (after the compilation threshold is reached). #101203 (Raúl Marín).
• Fix minmax_count_projection and trivial COUNT(*) optimizations being permanently disabled after a lightweight delete, even after all parts with a mask of lightweight delete were merged away. #101212 (Anton Popov).
• Fix a case which can lead to Having zero bytes, ... logical error from cache arising from a remote object being overwritten in between list and read which previously resulted in a stale object metadata. #101219 (Kseniia Sumarokova).
• Fix server crash (LOGICAL_ERROR: Bad cast from ColumnVector to ColumnLowCardinality) when querying a MergeTree table with ORDER BY CAST(lc_column, 'Type') where lc_column has a LowCardinality type. #101220 (Groene AI).
• Fix cleanup of stale processing nodes in S3Queue. #101230 (Kseniia Sumarokova).
• Fix UB in mergeTreeAnalyzeIndexes() in case of invalid optimizations argument. #101253 (Azat Khuzhin).
• Fix Logical error: 'partitions_count > 0' exception when performing consecutive ALTER TABLE UPDATE on a partitioned Iceberg table. #101278 (Desel72).
• Fix wrong query results when a large integer constant (e.g. 256, 2147483648) is used as a boolean predicate in a WHERE clause with AND on MergeTree tables. For example, SELECT count() FROM t WHERE (2147483648 > b) AND 2147483648 would incorrectly return 0 instead of matching all rows. #101287 (Groene AI).
• Fix insert-select from delta lake cluster with replicated merge tree. #101299 (Konstantin Vedernikov).
• Fixed crash with “Logical error: Reading from materialized CTE before materialization” when a scalar subquery references a chain of dependent materialized CTEs. #101305 (Groene AI).
• Fix data race on storage_id in IStorage::getDependentViewsByColumn. #101385 (Nikolay Degterinsky).
• Fix BACKUP FROM SNAPSHOT AST formatting and cloning. #101405 (Pablo Marcos).
• Fix LOGICAL_ERROR crash “Current component is empty” when querying system.part_moves_between_shards with enforce_keeper_component_tracking enabled. #101462 (Groene AI).
• Fix segmentation fault in DataTypeDynamic::create() when the fuzzer generates a malformed Dynamic type AST. #101464 (Groene AI).
• Throw an error when delta_lake_snapshot_version or CDF version settings are used without DeltaKernel enabled, instead of silently returning wrong data. #101489 (Desel72).
• Fix NOT_FOUND_COLUMN_IN_BLOCK exception when using ARRAY JOIN with JOIN USING and analyzer_compatibility_join_using_top_level_identifier setting enabled. Close #101240. #101507 (Vladimir Cherkasov).
• Fix Iceberg INSERT retry loop failing when the table was created with iceberg_metadata_file_path and the target metadata version already exists. #101548 (Groene AI).
• Strip Nullable from result column in arrayIntersect and related functions to avoid serialization/deserialization mismatch. #101569 (George Larionov).
• Fix server crash (LOGICAL_ERROR) when SELECT-ing from a materialized view backed by an IcebergLocal table engine. #101577 (Groene AI).
• Fix incorrect error message when calling intExp10 with NaN argument — it said intExp2 instead of intExp10. #101582 (Krishna Chaitanya).
• Fix allow_statistics=0 not blocking ALTER TABLE ADD STATISTICS and ALTER TABLE DROP STATISTICS after refactoring in #100288. #101585 (Krishna Chaitanya).
• Fix KeeperMap CREATE TABLE failing with “Cannot create metadata for table” when leftover ZooKeeper nodes from a pre-25.1 partial drop are missing the drop_lock_version node. #101623 (Antonio Andelic).
• Fix possible logical error during reading Map subcolumns. Closes #100769. Closes #101336. #101641 (Pavel Kruglov).
• Fix exact subcolumn match priority over prefix match in getSubcolumnData to avoid possible crash. Closes #101271. #101645 (Pavel Kruglov).
• Fix crash (LOGICAL_ERROR: “ColumnUnique can’t contain null values”) when comparing a LowCardinality column with a Variant NULL constant while use_variant_default_implementation_for_comparisons is disabled. #101690 (Groene AI).
• What: Added empty-stream guard to Bzip2ReadBuffer so it returns EOF instead of throwing UNEXPECTED_END_OF_FILE when the inner stream is empty. #101691 (ClickGap AI Bot).
• What: Fixed the inverted description text for the alterable column in system.s3_queue_settings and system.azure_queue_settings — swapped 0 and 1 meanings to match the actual code behavior. #101703 (ClickGap AI Bot).
• Fix positiveModulo(tuple, number) incorrectly dispatching to division instead of modulo. #101709 (ClickGap AI Bot).
• Fixes a crash when thread_pool_size is configured on a cache-wrapped disk. Previously, FileCacheSettings::loadFromConfig() rejected thread_pool_size as an unknown setting, preventing the server from starting. The setting is a valid IDisk parameter that controls the number of threads used for disk-to-disk copy operations during background part moves. #101712 (Francisco).
• Fix RANGE_HASHED dictionary creation silently accepting a non-existent MAX range attribute and using the wrong type configuration when min and max range attributes had different types. The bug was a copy-paste error in buildRangeConfiguration that looked up min_attr_name instead of max_attr_name for the max attribute. #101732 (Yakov Olkhovskiy).
• Fix use-after-free crash in CPU lease scheduler when the wait timer outlives the worker thread whose ProfileEvents::Counters it references. #101761 (Antonio Andelic).
• Fixes bug in arrayLevenshteinDistanceWeighted and arraySimilarity functions. Closes #101725. #101767 (Mikhail f. Shiryaev).
• Fix Prometheus Query API ignoring POST form bodies. #101794 (James Cunningham).
• Fix exception escaping from S3 Client::~Client destructor causing server termination. #101798 (Gagan Dhakrey).
• Fix use-after-scope in parallel deserialization of Object type dynamic paths, which could cause crashes when reading tables with many dynamic paths. #101823 (Antonio Andelic).
• Fixed incorrect output in function formatDateTime with formatter%W with certain (non-default) formatting settings. #101847 (Robert Schulze).
• Fix shouldPatchFunction false negative in SYSTEM INSTRUMENT ADD when the search string first appears inside a template argument of the demangled symbol name. #101885 (Pablo Marcos).
• Fix UDF registry loss when ZooKeeper session expires during periodic refresh — all user-defined functions could become unavailable until a full refresh succeeds. #101891 (Nikita Fomichev).
• Fixed system.codecs description for AES_256_GCM_SIV to report AES-256 instead of AES-128. #101917 (Jimmy Aguilar Mena).
• Fix using wrong extremes in min-max index created on JSON column leading to wrong query result. Closes #101700. #101918 (Pavel Kruglov).
• The splitByString tokenizer now rejects empty separator strings. #101928 (Robert Schulze).
• Fix materialize_skip_indexes_on_merge=false not suppressing text (full-text) indexes during merge. Previously, only non-text skip indexes (minmax, set, bloom_filter) were suppressed; text indexes continued to be built, wasting CPU and I/O. #101932 (Groene AI).
• The sparseGrams tokenizer generated longer tokens than the provided max length (this was due to a hard-coded +2 in the implementation). #101934 (Elmi Ahmadov).
• Fix SIGSEGV in MergeTreeDataPartWriterWide::cancel when a stream constructor throws during addStreams, leaving a null entry in column_streams. #101936 (Antonio Andelic).
• Fixes an exception when querying Merge or Distributed tables with a full-text index and combined filter conditions that mix has*Tokens with LIKE, while query_plan_direct_read_from_text_index is enabled. #101939 (Jimmy Aguilar Mena).
• Fix undefined behaviour when parsing native protocol query packets with invalid QueryProcessingStage values. #101972 (Raúl Marín).
• Close TCP connection when an exception occurs during initial query parsing to prevent reading garbage from a desynchronized stream. #101989 (Raúl Marín).
• Fix SLRU race bug in filesystem cache 26.1+, which can lead to space reservation logical error. In debug build it can also lead to failed assert: 'Previous state is Evicting, but expected state to be Active while setting Evicting flag for 2c1e3484ecdc6b78a8978fa5b17c5097:0:339 (state: Evicting)'.. #101991 (Kseniia Sumarokova).
• Fix exception when casting a string with trailing data to empty Tuple() type. #102011 (Alexey Milovidov).
• Fixes incorrect row ordering in queries that use ORDER BY with the grace_hash join algorithm. Affected queries could return results in the wrong order, producing silently incorrect output. #102036 (János Benjamin Antal).
• Fixes a LOGICAL ERROR (Unexpected size of index type) that could occur in RIGHT JOIN and FULL JOIN queries when the max_bytes_in_join setting was configured. #102042 (Jimmy Aguilar Mena).
• Fixes cases where Time with negative values was returning wrong result on comparison with DateTime. Closes #101670. #102056 (Yarik Briukhovetskyi).
• Fix crash in UDF refresh caused by ZooKeeperRetriesControl retrying on a stale (expired) ZooKeeper session without renewing it. #102059 (Nikita Fomichev).
• Fix missing spaces when formatting unlock snapshot. close https://github.com/clickhouse/clickhouse/issues/101723. #102063 (Han Fei).
• Fix crash (SIGSEGV) when querying a view with a WHERE clause and the inner query produces columns with different types than the view metadata (e.g. Nullable from LEFT JOIN with join_use_nulls). #102085 (Miсhael Stetsyuk).
• Fix VectorSimilarityIndexCache entries never being evicted after part removal due to mismatched cache keys. #102152 (Seva Potapov).
• NACK broken messages when using rabbitmq storage. #102157 (Seva Potapov).
• Fix logical error when parsing incorrect empty tuple string. #102289 (Nihal Z. Miaji).
• Fix incorrect aggregation results (duplicate rows) when using optimize_aggregation_in_order=1 with GROUP BY columns ordered differently from the table’s sorting key. #102299 (Groene AI).
• Fix crash in IcebergLocal ALTER TABLE ... UPDATE when using Avro format, caused by LowCardinality/Nullable wrapper types not being unwrapped before serialization. #102337 (Desel72).
• Fixed a segmentation fault in mutations on materialized columns without an expression. Closes #102185. #102342 (zoomxi).
• Coalescing merge tree fix for array type. This closes #89509. #102384 (Konstantin Vedernikov).
• Fix segfault (or LOGICAL_ERROR in debug builds) when reading Parquet files with bloom filter push down enabled and WHERE clause equality/inequality conditions. The crash occurred due to an out-of-bounds memory access in the Parquet prefetcher’s bloom filter data retrieval, and could also cause non-deterministic wrong query results. #102385 (Groene AI).
• Fix LOGICAL_ERROR abort during SLRU filesystem cache dynamic resize caused by shared eviction statistics across sub-queues and incorrect recovery path for failed candidates. #102396 (Antonio Andelic).
• Fix failure to initialize on a fresh replica Alias tables without a target table in the Database Replicated. Closes #101320. #102397 (Nikolay Degterinsky).
• Fix out-of-bounds read in string search functions (countSubstrings, position, etc.) when searching for a needle consisting entirely of null bytes. #102401 (Raúl Marín).
• Full text index settings (enable_full_text_index, allow_experimental_full_text_index, use_skip_indexes_on_data_read) are no longer disabled when the compatibility setting points to a version older than 26.1. Previously this could prevent SharedDatabaseCatalog from creating tables with text indexes. #102422 (Nikita Fomichev).
• Fix out of bounds read in printf with trailing %. #102472 (Raúl Marín).
• Fixes a chassert exception ReadBuffer is canceled in debug builds in AsynchronousMetrics, caused by rewind not resetting the buffer cancellation flag. #102524 (Yuri Fedoseev).
• Fixed hasToken / hasTokenOrNull with separator-only needles (e.g. '()', '!!!') on columns with a text index: previously the index silently skipped all granules instead of throwing BAD_ARGUMENTS (for hasToken) or returning NULL (for hasTokenOrNull). #102544 (Jimmy Aguilar Mena).
• Fix OOMs on huge multi requests in the keeper. For OpenTelemetry tracing, we unconditionally allocate >1 KiB for OpenTelemetry spans in ZooKeeperRequest objects - meaning, for really huge multi requests, we try to allocate >10 GiB extra memory. To fix this, we now keep shared data in static memory and use std::unique_ptr over std::optional in ZooKeeperOpentelemetrySpans. #102586 (Miсhael Stetsyuk).
• Fix NamedCollection CurrentMetric being inflated by CREATE NAMED COLLECTION IF NOT EXISTS on existing collections, and not being initialized for collections loaded from config or SQL storage at startup. Closes #102507. #102598 (Pablo Marcos).
• Fix exception in getStructureOfRemoteTable when local shard returns empty columns due to concurrent DDL. #102604 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception when multiple concurrent CREATE TABLE IF NOT EXISTS queries target the same S3Queue table on a Shared database. #102610 (Nikita Taranov).
• Fix LOGICAL_ERROR crash “Unexpected number of rows in column subchunk” in native Parquet V3 reader when reading nullable columns with a WHERE filter. #102628 (Groene AI).
• Fix AzureWriteMicroseconds profile event description saying “read” instead of “write”. #102639 (Miсhael Stetsyuk).
• Bug with row policy that gives ‘Not found column in block’ exception in some special cases. #102648 (Yarik Briukhovetskyi).
• Fixed a server exception in ClusterDiscovery when a static cluster (defined in config) temporarily had no live nodes. #102661 (Kseniia Sumarokova).
• Fix wrong date data type inference in case of overflow after timezone adjustment. Closes #102601. #102674 (Pavel Kruglov).
• Fix SYSTEM WAIT VIEW hanging forever when the refreshable materialized view is dropped while the wait is in progress. #102681 (Nikolay Degterinsky).
• Fix CASE with Dynamic expression returning ELSE for all rows. Closes #102511. #102684 (Pavel Kruglov).
• Fix flattened Dynamic type serialization with binary encoded data types. Closes #101911. #102692 (Pavel Kruglov).
• Fix format_schema_source=‘query’ silently ignoring multi-row results. Closes #101905. #102698 (Pavel Kruglov).
• Report actual exit code via SSH client instead of mapping all errors to 1. Closes #101741. #102700 (Konstantin Bogdanov).
• Fix losing HTTP headers for dynamic/predefined query handlers. Closes #101846. #102706 (Konstantin Bogdanov).
• Apply Poisson sampling correction to collapsed jemalloc heap profiles to match jeprof output. Previously the collapsed format underestimated actual allocation sizes by not accounting for the sampling probability. #102759 (Antonio Andelic).
• Fix the crash in the function hasPhrase with a NULL argument. #102802 (Nikita Taranov).
• Fix server crash (SIGSEGV) when reading Avro files with recursive schemas containing cyclic symbolic type references. Now such schemas are detected and rejected with a clear error message instead of crashing. #102853 (Groene AI).
• Fix server crash (LOGICAL_ERROR assertion) when a function on a Variant column hits a memory limit or other non-type-conversion exception during result casting in FunctionVariantAdaptor. The exception is now propagated correctly instead of being misclassified as an internal error. #102855 (Groene AI).
• Fix server crash in debug/sanitizer builds when std::length_error is thrown during schema inference (e.g., from extreme input_format_msgpack_number_of_columns values or malformed input data). #102859 (Groene AI).
• Make null representation in serialization of replicated and sparse columns respect settings (e.g. format_tsv_null_representation). #102888 (Hechem Selmi).
• Backported in #103499: Fix S3 requests failing with ios_base::clear: unspecified iostream_category error instead of being retried, caused by Poco BufferedStreamBuf::flushBuffer not handling short writes from the socket layer. #102894 (Sema Checherinda).
• Fix minmax_count_projection and trivial COUNT(*) optimizations being permanently disabled after a lightweight delete, even after all parts with a mask of lightweight delete were merged away. #102900 (Anton Popov).
• Fix random crashes in jemalloc due to LTO. #102913 (Azat Khuzhin).
• Optimize row policy OR-chains to IN in the new analyzer. #102915 (Azat Khuzhin).
• Fix jemalloc metadata corruption caused by page cache freeing with wrong alignment that may lead to crashes. #102918 (Azat Khuzhin).
• Plain INSERTs without materialized views no longer request excessive ConcurrencyControl slots and threads (max_threads instead of max_insert_threads), preventing CC slot starvation and thread count blowup on clusters with high INSERT throughput. #102961 (Sema Checherinda).
• Reintroduce ArrowMemoryPool to allow throwing MEMORY_LIMIT_EXCEEDED to avoid kernel OOM. #102999 (Azat Khuzhin).
• Fix cast_string_to_date_time_mode being ignored for CAST to Nullable(DateTime). Closes #101840. #103035 (Pavel Kruglov).
• Added support for ALIAS columns in text index direct read optimization. #103037 (Anton Popov).
• Backported in #103454: Fix SELECT DISTINCT silently returning incomplete results when an aggregate projection matched the query and some parts of the table had no projection data (e.g. the projection was added on a table that already held data, and MATERIALIZE PROJECTION was not run). Closes #102951. #103052 (Nihal Z. Miaji).
• Fix wrong results returned by WHERE x AND toNullable(N) on MergeTree tables when N is an integer wider than UInt8 (e.g. 256, 65535, 2147483648, or any negative integer). The filter incorrectly discarded all rows because splitFilterNodeForAllowedInputs used a NULL zero when converting the Nullable remainder of a reduced AND to boolean, turning the comparison into NULL under three-valued logic. #103077 (Groene AI).
• Fixed incorrect argument type reported in error messages of string search functions (e.g. locate, position) when arguments are passed in swapped order (locate(needle, haystack) with function_locate_has_mysql_compatible_argument_order = 1). #103102 (Alex Kuleshov).
• Fix waitForPause hanging indefinitely when disableFailPoint is called with no thread paused at the failpoint. #103119 (Shaohua Wang).
• Cap pre-auth TCP Hello packet strings to 64 KB and add handshake_timeout_milliseconds server setting to limit total handshake time, preventing unauthenticated clients from consuming excessive memory or holding threads indefinitely. #103284 (Sema Checherinda).
• Fix Parquet ColumnIndex stats min_value > max_value for String columns. #103334 (Saurabh Kumar Ojha).
• Check for malformed flattened Dynamic data in Native format. #103392 (Pavel Kruglov).
• Populate _time column from url table function. #103437 (Nikita Taranov).
• Fix SVE detection using SVE instructions when unavailable. #103568 (Raúl Marín).

​

Build/Testing/Packaging Improvement

• Libstemmer (Snowball) dependency now uses newer version v3.0.1. #99256 (Jimmy Aguilar Mena).
• Randomize settings in clickhouse-test: use_skip_indexes_for_top_k, use_top_k_dynamic_filtering, query_plan_max_limit_for_top_k_optimization. #91782 (Nikita Fomichev).
• Implement a stress test for functions, checking the soundness of their various properties. #93543 (Michael Kolupaev).
• Provide own CMake configuration for llvm-project instead of importing upstream one. #97453 (Konstantin Bogdanov).
• Randomize more optimize_* settings in the test infrastructure to improve coverage of query optimization passes. #97547 (Alexey Milovidov).
• Use Rust toolchain nightly-2026-03-22. #98602 (Konstantin Bogdanov).
• Use wasmtime v42.0.1. #98603 (Konstantin Bogdanov).
• Use llvm-project 22.1.1. #98882 (Konstantin Bogdanov).
• Replace SANITIZE_COVERAGE (custom sanitizer callbacks, symbol-level granularity) with LLVM source-based coverage (WITH_COVERAGE, -fprofile-instr-generate -fcoverage-mapping) for the nightly per-test coverage pipeline. The server now reads its own coverage mapping from ELF sections at startup and collects (file, line_start, line_end) tuples per test via a new SYSTEM SET COVERAGE TEST 'name' command. Test selection in targeted CI checks uses line-range queries against a new checks_coverage_lines CIDB table and ranks candidate tests by how many changed diff lines they cover. #99513 (Nikita Fomichev).
• Fix llvm-libc link error with -O0 build. #100023 (Zheguang Zhao).
• Fixed two error messages on container startup in case the configuration did not contain the logger.log and logger.errorlog settings (e.g. when all log messages are supposed to go to STDOUT/STDERR). #100239 (Simon).
• Prefer ld64.lld over Apple’s ld (cctools-port ld64) for macOS builds. This should significantly reduce Darwin link times, as cctools-port’s ld64 is very slow with -ffunction-sections and -dead_strip. Falls back to ld if ld64.lld is not available. #100275 (Alexey Milovidov).
• Tests to ensure that directories do not register any secrets (REST + Glue). #100307 (Konstantin Vedernikov).
• Repeat recently modified tests with different randomized settings. #100385 (Alexey Milovidov).
• Track #embed file dependencies in CMake and enable ccache depend mode for correct rebuilds. #100411 (Alexey Milovidov).
• Add correctness tests for TPC-H queries. #100580 (Raufs Dunamalijevs).
• Use aws-sdk-cpp 1.11.771. #100582 (Konstantin Bogdanov).
• Allow performance test XMLs to reference external SQL query files and settings via file attribute. #100747 (Raufs Dunamalijevs).
• Fix compilation with -march=x86-64-v4 by adding missing TargetSpecific.h include in LowerUpperImpl.h. #100932 (Alexey Milovidov).
• Fix build on gentoo by adding --no-default-config to cxxflags. #100973 (Isak Ellmer).
• Added utils/auto-bisect/ — a shell-based bisection framework that downloads pre-built CI binaries and runs a user-supplied test script to find the first commit that introduced a regression, without requiring a local build. #100989 (Nikita Fomichev).
• Disable ThinLTO by default in CMake so that local developer builds no longer implicitly enable it. CI release builds are unaffected since they explicitly pass -DENABLE_THINLTO=1. #101041 (Alexey Milovidov).
• Add TPC-DS SF1 benchmark to performance tests. #101209 (Raufs Dunamalijevs).
• Added stateless tests for MergeTree lightweight deletes covering: has_lightweight_delete flag lifecycle, COUNT(*) correctness with optimize_trivial_count_query, MATERIALIZED/DEFAULT column integrity, ReplicatedMergeTree flag recovery, read_in_order with deleted rows, multiple delete/merge cycles, _row_exists column hiding, predicate variety, and ALTER DELETE RBAC enforcement. #101792 (Nikita Fomichev).
• Distroless Docker images are now published with ubuntu and alpine variants for tagged releases. #101941 (Rahul Nair).
• Stack traces now show clean bare relative paths (e.g. src/Common/Exception.cpp) instead of paths polluted with the build directory (e.g. ./ci/tmp/fast_build/./src/Common/Exception.cpp). #102000 (Raúl Marín).
• Add a CI style check that rejects files larger than 5 MB committed to the repository, with a whitelist for existing legitimate test data. Remove unused 14 MB zookeeper_log.parquet. #102080 (Raúl Marín).
• Remove ~400 unused #include directives from headers to reduce compilation times. #102585 (Raúl Marín).
• Use wasmtime v43.0.1. #102603 (Konstantin Bogdanov).
• Use openssl 3.5.6. #102606 (Konstantin Bogdanov).
• Use xz 5.8.3. #102607 (Konstantin Bogdanov).
• Upgrade distroless Docker image base from Debian 12 (glibc 2.36, OpenSSL 3.0) to Debian 13 (glibc 2.41, OpenSSL 3.5), reducing CVE surface to zero reachable vulnerabilities. #101678 (Rahul Nair).

​

ClickHouse release 26.3 LTS, 2026-03-26. Presentation, Video

​

Backward Incompatible Change

• Downgrading after upgrading may cause data loss. Propagate data types serialization versions to nested data types. For example, String serialization version with_size_stream previously was applied only on top-level String columns and Tuple elements. Now it’s applied to any String type inside any nested type like Array/Map/Variant/JSON/etc. This behaviour is controlled by MergeTree setting propagate_types_serialization_versions_to_nested_types that is now enabled by default. After this change, newly created data parts cannot be read by older versions, but old parts can be read on new version with no problems. Upgrade is safe, but downgrade is not — if you need to roll back after upgrading to 26.3, data written by 26.3 in columns with nested types will be unreadable! See #101429 for details. #94859 (Pavel Kruglov).
• Remove the hypothesis skip index type. It was an obscure, experimental feature with limited practical use. Creating tables with INDEX ... TYPE hypothesis will now produce an error. #96874 (Alexey Milovidov).
• Remove the experimental detectProgrammingLanguage function. #99567 (Alexey Milovidov).
• Fix NOT operator precedence to match the SQL standard: NOT now binds looser than IS NULL, BETWEEN, LIKE, and arithmetic operators. For example, NOT (x) IS NULL is now parsed as NOT (x IS NULL) instead of (NOT x) IS NULL. This may change the result of queries that relied on the previous (non-standard) behavior. #97680 (Alexey Milovidov).
• Corrects the metadata of normal projections so that projections with multi-column sorting keys are properly recognized. Builds on top of #90429. #91352 (Amos Bird).
• Fixed skip index files not respecting replace_long_file_name_to_hash setting, causing “File name too long” errors and broken index reads for indices with long names. Skip index filenames are now hashed when they exceed max_file_name_length, similar to column files. This is backward compatible (new servers read old parts), but downgrading (or old servers during a rolling upgrade) may cause long-named indices to be ignored. #97128 (Raúl Marín).
• Turn on async insert by default. ClickHouse will be batching all small inserts by default now. This setting is set under compatibility. If you set compatibility=<version less than 26.2> then default value would be the previous one, false. You can turn off/on async inserts at several levels: in the config in users profiles, for the session, for the query, or for the MergeTree table. #97590 (Sema Checherinda).
• Change the default value of mysql_datatypes_support_level from empty to decimal,datetime64,date2Date32, enabling proper mapping of MySQL DATE to Date32, DECIMAL/NUMERIC to Decimal, and DATETIME/TIMESTAMP with precision to DateTime64 by default. Previously, MySQL DATE columns were mapped to Date which cannot represent dates before 1970-01-01, causing data corruption. #97716 (Alexey Milovidov).
• Accept array of part names instead of regexp for mergeTreeAnalyzeIndexes{,UUID} since regexp is slow (Experimental feature). #98474 (Azat Khuzhin).
• Change default stderr_reaction from throw to log_last for executable UDFs. UDFs that write warnings to stderr no longer fail when exit code is 0. Exit code exceptions now include stderr content. #99232 (Xu Jia).

​

New Feature

• Added bucketed serialization for Map columns in MergeTree (map_serialization_version = 'with_buckets'). Keys are split into hash-based buckets so that reading a single key (m['key']) only reads one bucket instead of the entire column, providing 2-49x speedup for single-key lookups depending on map size. The number of buckets and the bucketing strategy can be controlled by new MergeTree settings: map_serialization_version, max_buckets_in_map, map_buckets_strategy, map_buckets_coefficient, and map_buckets_min_avg_size. #99200 (Pavel Kruglov).
• Support materialized CTE. Allow evaluating CTEs only once during query execution and store their results in temporary tables. Closes #53449. #94849 (Dmitry Novik).
• Allow certain SQL-standard functions without parentheses for compatibility, such as NOW. Closes #52102. #95949 (Aly Kafoury).
• You can now use the natural sort key function as naturalSortKey(s). #90322 (Nazarii Piontko).
• You can now use native JSON/Object input for JSONExtract functions. Closes #88370. #96711 (Fisnik Kastrati).
• If a query parameter has Nullable type and is not specified, we will assume that its value is NULL. #93869 (Vikash Kumar).
• Support auxiliary ZooKeeper for Replicated database. #95590 (RinChanNOW).
• Support has function for JSON type to check path existence, similar to Map. #96927 (DQ).
• Added the mergeTreeTextIndex(database, table, index) table function, which allows reading data directly from a text index. This function can be used for introspection or for performing aggregations on top of text index data. #97003 (Anton Popov).
• Add table_readonly MergeTree setting to mark tables as read-only, preventing inserts and modifications. #97652 (Alexey Milovidov).
• Add new setting use_partition_pruning and alias use_partition_key. Set it to false to disable partition pruning based on the partition key. #97888 (Nihal Z. Miaji).
• Implements ALTER TABLE ... EXECUTE expire_snapshots('<timestamp>') for Iceberg tables. #97904 (murphy-4o). #99130
• Allow each type=http entry in <protocols> to specify a custom <handlers> key pointing to a separate <http_handlers_*> config section, enabling different HTTP routing rules per port. #98414 (Amos Bird).
• Add pretty=1 option to EXPLAIN for tree-style indented output and compact=1 to collapse Expression steps, making query plans more readable. #98500 (Kirill Kopnev).
• Add restore_access_entities_with_current_grants server setting. When enabled, restored users/roles from backups have their grants limited to what the restoring user is allowed to grant (same semantics as GRANT CURRENT GRANTS), instead of failing with ACCESS_DENIED. #98795 (pufit).
• Add caseFoldUTF8 and removeDiacriticsUTF8 functions for Unicode case folding and diacritical mark removal. #98973 (George Larionov).
• Add normalizeUTF8NFKCCasefold string function for NFKC_Casefold Unicode normalization, which combines NFKC normalization with case folding. #99276 (George Larionov).
• Add unicode_word tokenizer for full-text indexes and the tokens function. It splits text using Unicode word boundary rules: ASCII words are formed with connector characters (underscore, colon, dot, single quote), while non-ASCII Unicode characters become single-character tokens. Note: 26.3 accepts this tokenizer only under the name unicode_word; it was renamed to asciiCJK in 26.4. #99357 (Amos Bird).
• Added max_skip_unavailable_shards_num and max_skip_unavailable_shards_ratio settings to limit how many shards can be silently skipped when skip_unavailable_shards is enabled. If the number or ratio of unavailable shards exceeds the configured threshold, an exception is thrown instead of returning silently incomplete results. #99369 (Alexey Milovidov).
• Users can now use SOME keyword for subquery expressions. It behaves identically to ANY. #99842 (Artem Kytkin).
• Add output_format_trim_fixed_string setting to strip trailing null bytes from FixedString values in text output formats. #97558 (NeedmeFordev).
• Support parenthesized table join expressions in FROM clause, e.g. SELECT * FROM (t1 CROSS JOIN t2). #97650 (Alexey Milovidov).
• Implement function toDaysInMonth: it returns the number of days in the month of the specified date. #99227 (Vitaly Baranov).

​

Experimental Feature

• Add experimental support for WebAssembly-based user-defined functions (UDFs), allowing custom function logic to be implemented in WebAssembly and executed within ClickHouse. Special thanks to Alexey Smirnov for contributing the Wasmtime backend support. #88747 (Vladimir Cherkasov). Incremental improvements for WASM UDF support. #99373 (Vasily Chekalkin).
• Add support for external SQL dialects using polyglot library. #99496 (Alexey Milovidov).
• Add ALP floating-point compression codec (without ALP_rd fallback for non-compressible doubles). #91362 (Nazarii Piontko).
• Add experimental lazy type hints for JSON columns. When enabled via allow_experimental_json_lazy_type_hints, ALTER TABLE ... MODIFY COLUMN json JSON(path TypeName) that only adds or modifies type hints completes instantly as a metadata-only operation, without rewriting historical data. Type hints are applied at query time for old parts and materialized during INSERTs and background merges. #97412 (tanner-bruce).
• Enable parallel reads from YTsaurus table engine. #97343 (MikhailBurdukov).

​

Performance Improvement

• Improve the performance of data lakes. In previous versions, reading from object storage didn’t resize the pipeline to the number of processing threads. This gives orders of magnitude improvements (~40x) on multi-core machines. #99548 (Alexey Milovidov).
• Now, the relationship between enable_parallel_replicas and automatic_parallel_replicas_mode is as follows. A query can use parallel replicas only if enable_parallel_replicas > 0. Additionally, if automatic_parallel_replicas_mode=1, the decision whether to use parallel replicas is made during planning based on the previously collected statistics. If automatic_parallel_replicas_mode=0, parallel replicas will be used for all supported queries regardless of any statistics. One notable exemption is distributed insert-select with parallel replicas: in this case, queries will always be executed as if automatic_parallel_replicas_mode=0. #97517 (Nikita Taranov).
• Allow partition pruning when the predicate contains any comparison operator (=, <, >, !=) and the partition key is wrapped in a deterministic function chain (e.g. PARTITION BY x and predicates like cityHash64(x) % 5 > 2, toYYYYMM(x) < 2026, toYYYYMM(x) = 2026, or toYYYYMM(x) != 2026 will all use the partition key for pruning). Closes #28800. #98432 (Nihal Z. Miaji).
• Allow read-in-order optimization and primary-key pruning when the CAST target type is Nullable and the conversion is monotonic; for example, with PRIMARY KEY x, ClickHouse can use read-in-order optimization for ORDER BY x::Nullable(UInt64) and can apply primary-key pruning for predicates such as WHERE x::Nullable(UInt64) > 500000. #98482 (Nihal Z. Miaji).
• Allow index pruning and filter pushdown when an integral column is compared with a float literal; for example, predicates like WHERE x < 10.5 can now use the primary key for pruning, and filters such as prime < 1e9 or number < 1e5 are now pushed down for the primes() and numbers() table functions instead of causing unbounded execution. Closes #85167. #98516 (Nihal Z. Miaji).
• Added a new SLRU cache for Parquet metadata to improve read performance by removing the need to re-download files just to read metadata. #98140 (Grant Holly).
• Support swapping sides of ANTI, SEMI and FULL joins based on optimizer statistics. #97498 (Hechem Selmi).
• Optimize granules skipping for pointInPolygon for large polygons and fix pointInPolygon index analysis throwing during primary key pruning. #91633 (Nihal Z. Miaji).
• Improve performance of levenshteinDistance function. #94543 (Joanna Hulboj).
• Optimize batch decimal type conversions by avoiding per-element function calls. #95923 (Konstantin Bogdanov).
• Iceberg tables now support asynchronous metadata prefetching via iceberg_metadata_async_prefetch_period_ms table setting, which periodically pre-populates the metadata cache. Additionally, the iceberg_metadata_staleness_ms query setting allows SELECT queries to use cached metadata if it is fresher than the specified staleness, eliminating calls to the Iceberg catalog during request processing. #96191 (Arsen Muk).
• S3Queue ordered mode uses S3 ListObjectsV2 StartAfter to avoid re-listing full prefix history, reducing ListObjects calls. #96370 (Venkata Vineel ).
• Lower memory usage for inserts deduplication. In general the original block is needed for deduplication, but for sync insert we could omit it and save the memory for good. #96661 (Sema Checherinda).
• Use an arch-specific value for cache line size instead of a hardcoded value of 64. #97357 (Nikita Taranov).
• Slightly optimized reading from the text index dictionary, improving the overall performance of text index analysis. #97519 (Anton Popov).
• Speed up LZ4 decompression of 16 byte blocks in ARM. #97774 (Raúl Marín).
• Refactor tokenization to a new high-performance interface, replacing the old iterator-style API to support SIMD and stateful tokenizers. Part of #90268. #97871 (Amos Bird).
• Improved performance of text index analysis for queries with combined conditions involving both indexed and non-indexed columns. Previously, early exit optimization during index analysis was incorrectly disabled in such cases. #98096 (Anton Popov).
• Improve the performance of queries with constant expressions that generate very long arrays or maps. #98287 (Alexey Milovidov).
• Fix key condition analysis for DateTime64 primary keys compared with integer constants, which previously resulted in no granule pruning. #98410 (Amos Bird).
• The setting optimize_syntax_fuse_functions is enabled by default. #98424 (Alexey Milovidov).
• Optimize avgWeighted aggregate function by using local accumulators instead of per-row store-forwarding through aggregate state, improving performance by up to 27% for Nullable inputs. #98793 (Antonio Andelic).
• Improves performance and reduces memory usage for parallel window functions in certain scenarios, and for arrayFold workloads with large arrays. This can also reduce page-fault pressure and improve stability under tight memory limits for affected queries. #98892 (filimonov).
• Improve performance of sorted merges. #99013 (Artem Zuikov).
• Optimize INTERSECT ALL and EXCEPT ALL. #99097 (Raufs Dunamalijevs).
• Support the read_in_order_use_virtual_row optimization for reverse-order reads. #99198 (Vladimir Cherkasov).
• Reduce cache contention in RIGHT and FULL JOINs by checking if JoinUsedFlags is set before writing. #99274 (Hechem Selmi).
• Optimize PrefetchingHelper::calcPrefetchLookAhead by replacing floating-point math with pure integer arithmetic, improving instruction cache layout and reducing cycle overhead during aggregation loops. #99327 (Riyane El Qoqui).
• Reduce memory consumption of Keeper by replacing absl::flat_hash_set with CompactChildrenSet for storing node children. The new container stores 0–1 children inline without heap allocation, which covers the majority of Keeper nodes. This reduces KeeperMemNode size from 144 to 128 bytes. #99860 (Antonio Andelic).
• Aggregate projections are now correctly supported in views. Fixes #32753. #88798 (Amos Bird).
• Support OUTER to INNER join conversion optimization with join_use_nulls. Closes #90978. #95968 (Vladimir Cherkasov).
• Improve subcolumns reading by correct sizes calculation before reading. It reduces memory usage and speeds up the subcolumns reading. #96251 (Pavel Kruglov).
• Make mark, uncompressed and page caches to use separate jemalloc arena to avoid memory fragmentation when short lived allocations, i.e., for queries and requests are mixed with longer lived allocations for caches. #96812 (Seva Potapov). #98812. #99021
• Tables with DELETE TTL rules can now use the vertical merge algorithm. #97332 (murphy-4o).
• Apply data skipping indexes during distributed index analysis. #97767 (Azat Khuzhin).
• Marks of secondary indexes are now prewarmed when the prewarm_mark_cache setting is enabled (loaded into the index mark cache during data part fetches and table startup). #97772 (Anton Popov).
• Reduced locking during access control. #97894 (Nikita Taranov).
• When apply_row_policy_after_final or apply_prewhere_after_final is enabled, compound AND conditions in row policies and PREWHERE are now decomposed to extract sorting-key atoms for primary key index analysis. Previously, if a deferred filter contained a mix of sorting-key and non-sorting-key predicates (e.g. x > 1 AND y != ‘foo’), the entire expression was excluded from index analysis. Now sorting-key atoms (like x > 1) are extracted and used for granule pruning, even from nested AND expressions. #98513 (Yarik Briukhovetskyi).
• Reduce lock contention in MergeTreeBackgroundExecutor by making task resources release without acquiring the lock. Closes #93620. #98604 (Dmitry Novik).
• Fix excessive memory usage (~514 MiB) during format auto-detection when reading non-Arrow data (e.g. JSON from url or file without explicit format), caused by the ArrowStream reader misinterpreting the first bytes as a huge metadata length. #98893 (Konstantin Bogdanov).

​

Improvement

• Makes it possible to parse GeoParquet files that contain different Geo type in the same column. #97851 (Mark Needham).
• Introduce tokensForLikePattern SQL function that tokenizes LIKE patterns while respecting wildcard semantics: % and _ are treated as wildcards, escaped wildcards (\%, \_) are treated as literals, and tokens adjacent to unescaped wildcards are discarded. #97872 (Amos Bird).
• Add a {_schema_hash} placeholder for the S3 table engine that inserts a hash of the table’s column definitions into the S3 path. #98265 (Miсhael Stetsyuk).
• SymbolIndex, addressToSymbol, system.symbols, and buildId now work on macOS by parsing Mach-O symbol tables. #99014 (Alexey Milovidov).
• The system.stack_trace table now works on macOS, allowing introspection of all server thread stack traces. #98982 (Alexey Milovidov).
• Add per-server LDAP config option <follow_referrals> (default false) to control whether the LDAP client follows referrals. Disabling referral chasing avoids timeouts and hangs when searching from an Active Directory domain-root base DN. Referral-related log messages are moved from warn to trace. #96765 (paf91).
• We now track any data skipping indices that were used during query execution in the query_log table in a new column named skip_indices. Fixes #78676. Original author @pheepa. #87862 (Grant Holly).
• ACCESS_DENIED hints no longer reveal column names unless the user can show all required columns; database/table names remain visible in the hint. #91067 (filimonov).
• Add a dedicated cleanup thread for MergeTree to prevent cleanup delays under heavy merge load. This resolves #86181. #91574 (Amos Bird).
• Reload cluster config if IPs of local server’s hostname changed, instead of IPs of any host. Fixes #81215, #70156 and #65268. #93726 (Zhigao Hong).
• Allow optimize_aggregators_of_group_by_keys to correctly optimize aggregate functions in GROUPING SETS queries. #93935 (Xiaozhe Yu).
• Keeper-bench: report errors in metrics and generate json metrics file for --input-request-log mode. #95748 (Mohammad Lareb Zafar).
• Add new clause ROLE to CREATE USER. #97074 (Vitaly Baranov).
• You can now set internal_replication settings for a cluster created by the Replicated database. #97228 (Pervakov Grigorii).
• New setting allow_nullable_tuple_in_extracted_subcolumns controls whether extracted Tuple(...) subcolumns from Tuple, Variant, Dynamic and JSON are returned as Nullable(Tuple(...)) (NULL for missing rows) or as Tuple(...) (default tuple values for missing rows). Disabled by default, this setting can only be changed by restarting the server. #97299 (Nihal Z. Miaji).
• Add information about deferred filters as a separate item to EXPLAIN query output (when using Row Policies/PREWHERE with FINAL). Related: #91065. #97374 (Yarik Briukhovetskyi).
• Enable type_json_allow_duplicated_key_with_literal_and_nested_object by default. It allows to avoid errors about duplicated keys during parsing of JSONs like {"a" : 42, "a" : {"b" : 42}} that can be formatted by ClickHouse from original JSON data {"a" : 42, "a.b" : 42}. #97423 (Pavel Kruglov).
• Keeper improvement: find_super_nodes is a very useful command for debugging unexpected growth in the node count in Keeper. Unfortunately, if there are multiple super nodes, it’s almost impossible to find more than one, because the command gets stuck forever traversing the children of the first encountered super node. This PR forbids traversing the children of super nodes. #97819 (pufit).
• Initial completion support for clickhouse-keeper-client. #97828 (Konstantin Bogdanov).
• Flush async logging buffers in case of crash. #97836 (Azat Khuzhin).
• Enable the impersonate feature by default (see EXECUTE AS target_user). #97870 (Vitaly Baranov).
• Improve canceling queries with SQLite table engine by KILL QUERY and cancel query (Ctrl+C) in clickhouse-client. #97944 (Roman Vasin).
• Add server setting jemalloc_profiler_sampling_rate to control jemalloc’s lg_prof_sample and expose it as jemalloc.prof.lg_sample asynchronous metric. #97945 (Antonio Andelic).
• Support weights in concurrent bounded queue implementation. #97962 (Daniil Ivanik).
• Add sslmode to the allowed keys for PostgreSQL dictionary sources. Previously, sslmode was not in the dictionary_allowed_keys allowlist in PostgreSQLDictionarySource.cpp, making it impossible to configure SSL mode for PostgreSQL dictionary connections. This prevented dictionaries from connecting to PostgreSQL servers that require SSL (e.g., AWS RDS, which enforces SSL by default), as the connection would fail TLS negotiation and the server would reject the unencrypted fallback. #98014 (mcalfin).
• Show a clear “no such file” error when passing a non-existent file path to clickhouse or clickhouse-local, instead of a confusing generic message. #98048 (Raúl Marín).
• Text indexes can now be built on top of Nullable([Fixed]String) and Array(Nullable([Fixed]String)) columns. #98118 (Jimmy Aguilar Mena).
• Avoid dropping named collections that are dependencies of dictionary sources. #98127 (Pablo Marcos).
• Enable grace_hash join algorithm for queries with totals. #98144 (János Benjamin Antal).
• Cancel background merges early in DROP DATABASE for ordinary shared merge tree. #98161 (Shaohua Wang).
• Improve canceling queries with MongoDB and MySQL by KILL QUERY and cancel query (Ctrl+C) in clickhouse-client. #98187 (Roman Vasin).
• Remove NetlinkMetricsProvider and use procfs exclusively for per-thread taskstats metric collection. Netlink-based collection is problematic in containerized environments and has worse tail latency under contention. #98229 (Amos Bird).
• Refactor Iceberg manifest file handling to fix issues with manifest file caching. #98231 (Daniil Ivanik).
• Now we take into account cases where the sorting key of the table can be an expression like toDate(time), and can make the decisions about not-deferring expressions like this if they are part of filters. #98237 (Yarik Briukhovetskyi).
• Add a new MaxAllocatedEphemeralLockSequentialNumber metric for the maximum sequential number allocated for ephemeral lock znodes in ZooKeeper. #98243 (Miсhael Stetsyuk).
• Update clickstack to version 2.20.0. #98252 (Aaron Knudtson).
• Added a new profile event KeeperRequestTotalWithSubrequests that counts each subrequest within a multi-request individually, providing better visibility into the actual Keeper workload. The existing KeeperRequestTotal event continues to count each multi-request as a single request. #98348 (Antonio Andelic).
• SYSTEM RELOAD DICTIONARIES now reloads dictionaries in topological order so that dictionaries sourcing from other dictionaries see fresh data after reload. #98356 (Alexey Milovidov).
• Restart the statistics cache after changing the MergeTree setting. #98520 (Han Fei).
• Only “alive” replicas (that are connectable) participate in distributed index analysis. #98521 (Azat Khuzhin).
• Add a setting access_control_improvements.disallow_config_defined_profiles_for_sql_defined_users (disabled/allowed by default) that disallows using config-defined settings profiles (except for the default profile) for SQL-defined users. #98662 (Alexander Tokmakov).
• Cap the number of nodes used in the automatic parallel replicas heuristic to the actual number of nodes in the cluster (instead of only the max_parallel_replicas setting). #98668 (Nikita Taranov).
• Implement hedged requests and asynchronous reading for distributed index analysis. #98724 (Azat Khuzhin).
• Deserialization of binary AggregateFunction states now requires consuming the full input. If extra redundant trailing bytes are present, ClickHouse throws an exception instead of accepting malformed state data. #98786 (Nihal Z. Miaji).
• Make TRUNCATE DATABASE respond to query cancellation. #98828 (Shaohua Wang).
• Improve keeper-bench with request pipelining, warmup period, per-operation stats, reproducible seeds, and better error handling. #98906 (Antonio Andelic).
• Support SAMPLE clause in distributed index analysis. #98931 (Azat Khuzhin).
• Show chart title in dashboard even when query returns empty result or encounters an error. #98975 (Yash ).
• Analyzer error messages no longer dump all columns of a table (which could produce 150KB+ exceptions). Column lists are now capped at 10 entries. #99002 (Yash ).
• Properly return column stats from sub-queries with joins so that parent query can use them for join reordering. #99096 (Alexander Gololobov).
• Mark ZooKeeper session as expired immediately when finalization starts, instead of waiting for the send thread to exit. This allows other threads to establish a new session without delay. #99102 (Raúl Marín).
• Use more math functions from LLVM-libc: exp, exp2, expm1, fabs, fabsl, floor, fmodl, log, log2, logf, pow, scalbn, scalbnl, copysignl, nan, nanf, nanl, and the explogxf shared constants. #99118 (Konstantin Bogdanov).
• Reduce memory usage and fix potential duplicate output in system.jemalloc_profile_text collapsed format. #99121 (Antonio Andelic).
• Add is_subrequest column to system.aggregated_zookeeper_log to separate standalone requests from sub-requests inside Multi/MultiRead requests. Previously, sub-requests were aggregated into the same buckets as standalone requests, and since each sub-operation was logged with the total multi-request duration, the average latency became misleading. Sub-requests now have zero latency. #99169 (Miсhael Stetsyuk).
• Allow ALTER TABLE MODIFY COLUMN x TTL ... command without specifying column type. #99208 (Nikolay Degterinsky).
• Skip stale Keeper requests for sessions that have already disconnected, avoiding unnecessary Raft round-trips. The number of tracked finished sessions is capped by the max_finished_sessions_cache_size coordination setting. #99246 (Antonio Andelic).
• Support text index built on mapValues(map) with IN operator. #99286 (Anton Popov).
• Shell-like completion support in clickhouse keeper-client (handle completions of arguments in quotes, i.e. 'foo ba', handle escaped arguments, i.e. foo\ ba, make ls print quoted nodes, if they have whitespaces). #99312 (Azat Khuzhin).
• Prevent Keeper mntr command from getting stuck because of lock contention. #99472 (Antonio Andelic).
• Reduce lock contention in Keeper dispatcher by invoking callbacks and dispatching read requests outside mutex scope, and add profiled lock guards for observability. #99751 (Antonio Andelic).
• Tolerate missing padding at the end of the last block of parquet files. #99857 (Seva Potapov).

​

Bug Fix (user-visible misbehavior in an official stable release)

• Fixes how an Alias table target is saved as a DDL dependency when not fully qualified: it’s now saved with the Alias table database instead of the session database. #95175 (Enric Calabuig).
• Fix wrong result or exception during reading subcolumns of ALIAS columns. #95408 (Pavel Kruglov).
• Fix missing column when non-standard identifier alias in JOIN with old analyzer. Fixes #25594, #47288 and #53263. #95679 (Zhigao Hong).
• Fixes a crash in Kusto dialect functions bin(), bin_at(), extract(), and indexof() when empty arguments are provided. #95736 (NeedmeFordev).
• Forbids mounting local_object_storage (which is used by datalakes above local filesystem and maybe by LocalDisk) anywhere apart from user_files_path in clickhouse-client. #96201 (Daniil Ivanik).
• In DeltaLake table engine fix logical race on snapshot version change, remove redundant heavy snapshot reloads. #96226 (Kseniia Sumarokova).
• Fix logical error on attaching a part in MergeTree if there were several chained renames between detaching and attaching. #96351 (Alexey Milovidov).
• Fix a bug where explicit settings sent alongside compatibility in the same request could be silently ignored when their value matched the server default. #97078 (Raufs Dunamalijevs).
• Fix the client reporting NETWORK_ERROR instead of the actual parsing error (with the correct row number) when an INSERT with parallel parsing encounters invalid data. #97339 (Alexey Milovidov).
• Fix sumCount aggregate function not being able to read older serialized states after introduction of Nullable(Tuple). Closes #97370. #97502 (Nihal Z. Miaji).
• Fix exception in tuple comparison involving Nothing type elements (e.g., comparing with NULL tuple elements) when used with GROUPING SETS and ORDER BY. #97509 (Alexey Milovidov).
• Fix non-deterministic uncompressed_hash computation for Compact MergeTree parts when multiple compression codecs are used, which could cause incorrect deduplication behavior. #97522 (Alexey Milovidov).
• Fix logical error about missing stream during INSERT SELECT with JSON and buckets in shared data. Closes #97331. #97523 (Pavel Kruglov).
• Fix MEMORY_LIMIT_EXCEEDED exceptions being incorrectly reported as CORRUPTED_DATA during SummingMergeTree and CoalescingMergeTree merges. #97537 (János Benjamin Antal).
• Fix “Context has expired” exception for correlated subqueries containing table functions like url(). #97544 (Alexey Milovidov).
• Fix exceptions and incorrect behavior in optimize_syntax_fuse_functions with aggregate projections, Date types, and column name preservation. #97545 (Alexey Milovidov).
• Remove incorrect replaceRegexpOne to extract query rewrite that produced wrong results when the regexp didn’t match; also fix exception when replaceRegexpOne is used with GROUP BY ... WITH CUBE and group_by_use_nulls=1. #97546 (Alexey Milovidov).
• Fix DROP DATABASE with database_atomic_wait_for_drop_and_detach_synchronously hanging indefinitely when the query is killed. #97586 (Alexey Milovidov).
• Fix KILL QUERY not being able to terminate queries stuck in WITH FILL generation, dictionary loading via dictGet, or ALTER DELETE with mutations_sync=1 on ReplicatedMergeTree. #97589 (Alexey Milovidov).
• The loop table function was calling inner_storage->read() directly, bypassing the interpreter layer where row policies, column-level grants, and other security checks are applied. This allowed a user restricted by row policies to read all rows via loop(table) even when a direct SELECT returned zero rows. #97682 (pufit).
• Fixes incorrect partition pruning in cases of using pre-epoch DateTime64 with toDate() function. #97746 (Yarik Briukhovetskyi).
• After this patch, hasPartitionId will return false if another partition with a higher partition ID exists in the data part set. #97748 (Mikhail Artemenko).
• Fix possible crashes during reading of empty granules in advanced shared data in JSON. Closes #97563. #97778 (Pavel Kruglov).
• Fix Cannot schedule a file LOGICAL_ERROR on INSERT into Distributed due to race between DROP and INSERT. #97822 (Azat Khuzhin).
• Fixed ClickHouse server crash/assert in call to mapContainsKey/mapContainsKeyLike with tokenbf_v1 skip index. #97826 (Shankar Iyer).
• Fix LOGICAL_ERROR exceptions caused by LowCardinality inside compound types (Variant, Dynamic, Tuple) in concatWithSeparator, format, IN subqueries, GLOBAL IN, and joins with runtime filters. #97831 (Raúl Marín).
• Fix LOGICAL_ERROR exception Chunk info was not set for chunk in MergingAggregatedTransform when using ARRAY JOIN with merge() table function over multiple Distributed tables combined with GROUP BY. #97838 (Raúl Marín).
• Fix server crash (std::terminate) caused by an uncaught exception in the HTTP connection pool destructor when the connection group hard limit is reached under high concurrency. The exception HTTP_CONNECTION_LIMIT_REACHED could escape ~PooledConnection when recycling a connection back to the pool, leading to SIGABRT. #97850 (Antonio Andelic).
• Fix incorrect result when grace_hash algorithm is used with non-equi joins and left block cannot be processed completely because of the size constraints on the join result. #97866 (János Benjamin Antal).
• Fix performance inefficiency in DeltaLake metadata scanning introduced in #96686. #97880 (Kseniia Sumarokova).
• Fix data race in ZooKeeper client between sendThread and receiveThread. #97887 (Pablo Marcos).
• Fixes a bug where it was not possible to use CTE with distributed insert selects. Continuation of https://github.com/ClickHouse/ClickHouse/pull/87789. Closes #95837. #97889 (Yarik Briukhovetskyi).
• Fix exception from CachedOnDiskReadBufferFromFile::readBigAt. Closes #97325. #97890 (Kseniia Sumarokova).
• Fix LOGICAL_ERROR exception in Alias engine with materialized columns due to column mismatch. Closes #97907. #97921 (Kai Zhu).
• Fix Keeper data loss after restart when using Azure Blob Storage with s3_plain metadata for log storage. #97987 (Antonio Andelic).
• Fix JIT miscompilation of sign function for integer types wider than Int8 — values outside the -128..127 range could produce incorrect sign. #98012 (Alexey Milovidov).
• Fixed DUPLICATE_COLUMN exception and silent NULLs when reading Delta Lake tables that use column mapping “name” mode with struct fields whose names contain dots (e.g. STRUCT<`a.foo`: STRING, `b.foo`: STRING>). #98013 (Caio Ishizaka Costa).
• Fix mutation after lightweight update and secondary indices. #98044 (Raúl Marín).
• Fix incorrect result of FINAL queries when mixing primary key and non primary key skip indexes. #98097 (Raúl Marín).
• Enforce READ ON FILE checks for scalar file() and DESCRIBE TABLE file(). #98115 (Nikolay Degterinsky).
• Fixes a crash where querying files with a glob pattern (e.g., file('dir/**', 'LineAsString')) would throw an unhandled filesystem exception (STD_EXCEPTION) if the directory contained a dangling symlink. Dangling symlinks are now silently skipped, and the query returns results from all valid files. #98143 (Mark Andreev).
• Fix segfault in query plan optimization when converting outer join to inner join with arrayJoin in filter expression. #98147 (Alexey Milovidov).
• Fix ProtobufList format not working with Kafka engine due to read state not being reset between messages. #98151 (Alexey Milovidov).
• Fix logical error with analyzer_compatibility_join_using_top_level_identifier and ARRAY JOIN, closes #98164. #98179 (Vladimir Cherkasov).
• Set Watch component for watch responses in aggregated_zookeeper_log instead of leaving it empty. #98202 (Antonio Andelic).
• If the partition key columns are not covered by the sorting key, then partition pruning could incorrectly skip partitions containing rows that should “win” during FINAL deduplication. #98242 (Yarik Briukhovetskyi).
• Fix logical error “Bad cast from type DB::ColumnConst to DB::ColumnArray” in kql_array_sort_asc/kql_array_sort_desc when called with constant array arguments. #98251 (Alexey Milovidov).
• Fix out-of-bounds access in ColumnConst::getExtremes that could cause a crash when extremes = 1 is enabled. #98263 (Alexey Milovidov).
• Fix potential deadlock when two concurrent MOVE PARTITION operations work with the same pair of tables in opposite directions. #98264 (Alexey Milovidov).
• HTTP server now returns an error message in the response body for 400 Bad Request responses caused by malformed headers, instead of an empty body. #98268 (Alexey Milovidov).
• Fix wrong results with distributed index analysis (experimental feature) and query condition cache. #98269 (Azat Khuzhin).
• Fixed LOGICAL_ERROR exception “Invalid binary search result in MergeTreeSetIndex” triggered by toDate conversion on key columns with data crossing the 65535 boundary. #98276 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception when a RIGHT JOIN wrapped in a CROSS JOIN is swapped by the query_plan_join_swap_table optimization in the legacy join step code path. #98279 (Alexey Milovidov).
• Validate corrupted data during DDSketch deserialization to prevent segfaults, exceptions, infinite loops, and OOM when reading corrupted quantilesDD aggregate function states. #98284 (Alexey Milovidov).
• Fix LOGICAL_ERROR “Trying to execute PLACEHOLDER action” when correlated columns from outer queries are referenced inside lambda functions such as arrayMap. #98285 (Alexey Milovidov).
• Fix logical error exception in caseWithExpression when CASE expression involves materialize(NULL) or other Nullable(Nothing) arguments. #98290 (Alexey Milovidov).
• Fix bad cast exception when filtering _table virtual column in merge table function. #98291 (Alexey Milovidov).
• Fix sporadic deduplication failure where re-inserts were incorrectly deduplicated due to inconsistent cleanup ordering between blocks/ and deduplication_hashes/ ZooKeeper directories. #98293 (Alexey Milovidov).
• Fix exception when ORDER BY ... WITH FILL is used together with LIMIT BY. #98361 (Alexey Milovidov).
• Fix silent data corruption when inserting a Parquet/Arrow Date column into an Enum column — now properly rejects the incompatible type conversion instead of storing invalid enum values. #98364 (Alexey Milovidov).
• Fix exception when reading an Arrow file with Array column into a table with Nested column. #98365 (Alexey Milovidov).
• Fix MATERIALIZE INDEX and MATERIALIZE PROJECTION mutations getting stuck when the index or projection is dropped before the mutation finishes. #98369 (Alexey Milovidov).
• Fix exception when reading from Nullable(Tuple(...)) where a Tuple element name collides with the Nullable null subcolumn. #98372 (Alexey Milovidov).
• Fix exception “Column … query tree node does not have valid source node” when joining a Merge table (wrapping a Distributed table) with another table. #98376 (Alexey Milovidov).
• Fix incorrect Parquet Bool to FixedString conversion in native V3 reader that produced raw bytes instead of string representation. #98378 (Alexey Milovidov).
• Fix tryGetColumnDescription to filter subcolumns by parent column kind, consistent with other column lookup methods. #98391 (Alexey Milovidov).
• Accept base64 credentials without padding in HTTP Basic Auth. Some HTTP clients omit trailing = padding in the Authorization: Basic header, which previously caused authentication failures. #98392 (Amos Bird).
• Fix incorrect partition pruning results after merging parts with Nullable partition key columns, caused by wrong min-max index bounds. #98405 (Amos Bird).
• Fix rare exception in the pipeline executor, that could manifest as a Received signal 6 (only in debug builds), when pipeline expansion races with query cancellation. #98428 (Alexey Milovidov).
• Fix exception “Column identifier is already registered” when count_distinct_optimization is used with a QUALIFY clause. #98433 (Alexey Milovidov).
• Fix exception “cannot be inside Nullable type” when using IN/NOT IN with LowCardinality column arguments (e.g. a NOT IN (b) where a is LowCardinality(String)). #98443 (Alexey Milovidov).
• Fix “Pipeline stuck” exception in full_sorting_merge joins caused by a deadlock in PingPongProcessor when the FilterBySetOnTheFly optimization created a circular dependency with MergeJoinTransform. #98454 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception “Projection cannot increase the number of rows in a block” when merging parts with TTL that deletes all rows and an aggregate projection with a constant GROUP BY key. #98458 (Alexey Milovidov).
• Fix logical error exception when CROSS JOIN is used together with INNER JOIN USING. #98459 (Alexey Milovidov).
• Fix null pointer dereference in dictGetOrDefault when the key argument is Nullable. #98460 (Alexey Milovidov).
• Fix exception in DISTINCT queries when using aggregate projections and materialize causes LowCardinality type differences between the query and the projection. #98462 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception when arrayJoin is used in a filter expression with OUTER JOIN and join_use_nulls enabled. #98464 (Alexey Milovidov).
• Fix logical error exception “Replica decided to read in WithOrder mode, not in ReverseOrder” when using parallel replicas with optimize_aggregation_in_order. #98467 (Alexey Milovidov).
• Fixed ClickHouse Keeper disconnecting Java ZooKeeper clients after addWatch request. The Java client expects a 4-byte ErrorResponse body in the addWatch response, but Keeper was sending an empty body, causing EOFException and session disconnect. This broke Apache Curator’s CuratorCache and any Java application using persistent watches. Fixes #98079. #98499 (Antonio Andelic).
• Fix zk_followers and zk_synced_followers Keeper metrics not decreasing when a follower goes down. Add new zk_learners and zk_synced_non_voting_followers metrics to the mntr four-letter-word command. Fixes #54173. #98504 (Antonio Andelic).
• Fix a LOGICAL_ERROR exception in renameAndCommitEmptyParts that could occur when TRUNCATE TABLE runs concurrently with OPTIMIZE TABLE using MergeTree transactions. #98508 (Alexey Milovidov).
• Fixed Keeper’s secure raft port ignoring cipherList and dhParamsFile from openSSL configuration, always using defaults instead of user-specified values. Close #51188. #98509 (Antonio Andelic).
• Fixed misleading Keeper log messages like “Receiving request for session X took 9963 ms” where the reported time was actually spent waiting idle in poll() between heartbeats, not performing the operation itself. Fixes #79026. #98510 (Antonio Andelic).
• Fix unexpected result with read_in_order_use_virtual_row and monotonic functions, close #97837. #98514 (Vladimir Cherkasov).
• Fix LOGICAL_ERROR: Not-ready Set is passed as the second argument for function 'in' when using PREWHERE with IN subquery on MergeTree tables. #98522 (Alexey Milovidov).
• Fix Keeper TCP connections preventing graceful server shutdown by not responding to shutdown signal. #98525 (Alexey Milovidov).
• Fix exception “Sorting column wasn’t found in the ActionsDAG’s outputs” when query_plan_convert_join_to_in is enabled with query_plan_merge_expressions = 0. #98526 (Alexey Milovidov).
• Fix MongoDB dictionary source failing with named collections. Closes #97840. #98528 (Pablo Marcos).
• Fixed LOGICAL_ERROR when Identifier is empty after parameter substitution. #98530 (Pervakov Grigorii).
• Fix pipeline deadlock when using sort_overflow_mode = 'break' together with window functions. #98543 (Alexey Milovidov).
• Fix column rollback in Buffer engine during handling an exception during appending a new block. Old logic could lead to corrupted in-memory state of columns. #98551 (Pavel Kruglov).
• Fixed exception Bad cast from type ColumnConst to ColumnDynamic in null-safe comparison (<=> / IS NOT DISTINCT FROM) with const Dynamic or Variant columns and NULL. Also fixed IS DISTINCT FROM with Dynamic/Variant vs NULL always incorrectly returning 0. #98553 (Alexey Milovidov).
• Fixed usage of the text index with other skip indexes. Previously, logical errors such as “Trying to get non-existing mark” could be thrown when a query filter utilized a text index and other regular skip indexes simultaneously. #98555 (Anton Popov).
• Fix logical error “TABLE_FUNCTION is not allowed in expression context” when a table function with an alias appears multiple times in the same query scope (e.g. in both PREWHERE and QUALIFY clauses). #98557 (Alexey Milovidov).
• Fix distributed index analysis with expressions (not just columns) in PK (leads to zero filtering of redundant granules on remote replicas). #98561 (Azat Khuzhin).
• Disallow dropping column when its subcolumns are used in other columns default/alias expressions and use analyzer for default expressions on alter drop column. #98569 (Nikita Mikhaylov).
• Fix S3 requests being incorrectly retried on non-retryable errors (including HTTP_CONNECTION_LIMIT_REACHED) in the HTTP client. #98598 (Sema Checherinda).
• Fixes a decimal overflow when partition pruning with DateTime64. #98628 (Yarik Briukhovetskyi).
• Fix two bugs in JIT expression compilation: a copy-paste error in nativeCast type checking that made integer-to-integer and float-to-float cast branches unreachable, and incorrect nullptr TargetMachine passed to LLVM PassBuilder preventing target-specific optimization passes from being registered. #98660 (Alexey Milovidov).
• Fix RBAC bypass that allowed users to DESCRIBE any table via remote(), remoteSecure(), cluster(), or clusterAllReplicas() pointed at localhost, without requiring SHOW_COLUMNS privilege. #98669 (pufit).
• Fix BAD_GET exception and incorrect query results when a non-boolean expression (e.g. sin(col)) is used in both WHERE and SELECT with a JOIN, due to filter push-down optimization corrupting shared DAG nodes. #98681 (Alexey Milovidov).
• Fix LOGICAL_ERROR “Replica decided to read in Default mode, not in WithOrder” when using read_in_order_through_join with parallel replicas. #98685 (Alexey Milovidov).
• Fix exception “Bad cast from type DB::TableFunctionNode to DB::QueryNode” when using input table function as an argument of remote. #98694 (Alexey Milovidov).
• Fix outdated data parts resurrection caused by incorrectly cleaning up empty covering parts. #98698 (Shaohua Wang).
• Fix exception in LogicalExpressionOptimizerPass when a boolean function in an equals comparison returns a Variant type. #98712 (Alexey Milovidov).
• Fix parseDateTimeBestEffort incorrectly parsing words starting with month/weekday prefixes. Closes #97965. #98742 (Pavel Kruglov).
• Fix UNKNOWN_IDENTIFIER exception when querying merge() table function or Merge engine over tables with JSON columns that have different parameters (e.g. different SKIP fields) and ALIAS columns referencing JSON sub-paths, with the new analyzer enabled. Closes #97812. #98753 (Pavel Kruglov).
• Fix optimize_skip_unused_shards optimization with the analyzer in case Distributed storage is used in a View. #98754 (Nikolai Kochetov).
• Fix tuple subcolumn access by name (e.g. SELECT x.a for Tuple(a UUID, b Int32)) for external tables passed via --external in clickhouse-client. Closes #96925. #98755 (Pavel Kruglov).
• Fix reverseUTF8 exception on invalid (truncated) UTF-8 input. #98770 (Alexey Milovidov).
• Fix detecting set skip index usefulness with OR with false (i.e. or(x, 0)) predicate. #98776 (Azat Khuzhin).
• Fix a LOGICAL_ERROR exception (Block structure mismatch in removeUnusedColumns) that could occur with FINAL + PREWHERE + constant WHERE expression + column-independent aggregates like count(). #98778 (Alexey Milovidov).
• Make system.trace_log entries for ClickHouse dictionaries’ auto-reloads have non-empty query IDs. #98784 (Miсhael Stetsyuk).
• Fixes a crash where we could de-reference a null pointer in system tables created between the time when we snapshot the tables in the IDatabaseTablesIterator::table() call and the tables changing in another thread during later iteration. #98792 (Grant Holly).
• Fix SYSTEM START REPLICATED VIEW not waking up the refresh task. #98797 (Pablo Marcos).
• Fix exception “Inconsistent table names” when using view() table function containing JOINs inside another JOIN (only with old analyzer). #98809 (Alexey Milovidov).
• Fix adjusting RLIMIT_SIGPENDING (via pending_signals). #98829 (Azat Khuzhin).
• Fix exception when composing loop with cluster table functions. #98860 (Konstantin Bogdanov).
• LEFT ANTI JOIN with multiple join key columns returned wrong results when enable_join_runtime_filters=1 (which is default). #98871 (Alexander Gololobov).
• Fix WITH FILL STALENESS producing extra filled rows when data is read in multiple chunks (e.g., with small index_granularity). #98895 (Alexey Milovidov).
• Fix “RPNBuilderFunctionTreeNode has A arguments, attempted to get argument at index B” LOGICAL_ERROR. #98900 (Azat Khuzhin).
• Fix memory tracking drift caused by failed allocations not being rolled back, nallocx(0) undefined behavior, and off-by-one in global peak tracking. Extend tracking to cover io_uring ring buffers. #98915 (Antonio Andelic).
• Forbid attaching to local data lake tables outside user paths, not only creating them. #98936 (Daniil Ivanik).
• Fix a race condition that could cause a “ReadBuffer is canceled” exception in queries using urlCluster or similar cluster table functions. #98955 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception in financial functions (financialNetPresentValue, financialInternalRateOfReturn, etc.) when BFloat16 type arguments are passed. #98958 (Alexey Milovidov).
• Fix skip indexes (and primary key conditions) not being applied for ALIAS columns when query plan expression merging is disabled (query_plan_merge_expressions = 0 or query_plan_enable_optimizations = 0). #98960 (Peng).
• Increment InsertQuery ProfileEvent for async inserts. Closes #98626. #98962 (Narasimha Pakeer).
• Fix exception “Inconsistent KeyCondition behavior” in debug builds when primary key contains NaN float values, by making accurateLess and accurateEquals handle NaN consistently with ClickHouse sort order. Closes #98075. #98964 (Alexey Milovidov).
• SummingMergeTree no longer sums Bool (and other domain type) columns. Bool values are kept as-is instead of being arithmetically summed. #98976 (Yash ).
• Fixes an exception Scalar doesn’t exist that occurred when querying a remote shard with optimize_const_name_size set and enable_scalar_subquery_optimization = 0. Large constants replaced with __getScalar references in the remote query were not being sent to the shard, causing the query to fail. #98979 (andriibeee).
• Fix NOT_FOUND_COLUMN_IN_BLOCK for some queries with GROUP BY and expressions that include inverse dictionary lookup, Date/DateTime conversion comparisons, and tuple comparisons. Closes #98888. #98980 (Nihal Z. Miaji).
• Fixed undefined behavior (null pointer dereference) when altering a version/sign/is_deleted column to EPHEMERAL or ALIAS in MergeTree engines. Such alterations are now properly rejected. #98985 (Alexey Milovidov).
• Fixed an issue where system.grants omitted the regular expression parameters for URL and S3 grants in the access_object column. #98987 (DQ).
• Fixed Iceberg BigLake reads: ADC credentials are now forwarded to the GCS S3 client (fixing 403 errors), OAuth2 credentials are URL-encoded before sending (fixing auth failures for tokens with special characters), and namespace traversal no longer aborts on BigLake HTTP 400 responses. #98998 (Nikita Fomichev).
• Fixed clickhouse-client failing to switch timezone when TZ environment variable uses POSIX file path syntax (e.g. TZ=:/etc/localtime). #99000 (Yash ).
• Fix incorrect or less pruning when startsWith, LIKE, NOT LIKE used with FixedString column. Additionally, FixedString to String cast function can now prune granules when wrapped around key column. Closes #98940. #99001 (Nihal Z. Miaji).
• Fixed windowFunnel with strict_deduplication returning incorrect level when a duplicate event was encountered. #99003 (Yash ).
• Fixes a bug where EXISTS would ignore LIMIT and OFFSET clauses in subqueries, causing incorrect results when the subquery returned no rows due to an offset or a zero limit. Closes #88722. #99005 (andriibeee).
• Fix “Block structure mismatch” exception when filter push-down optimization encounters an AND expression that short-circuits to a constant with GROUPING SETS. #99010 (Alexey Milovidov).
• Fix exception when reading patch parts (lightweight updates) without _part_offset column in the query plan. #99023 (Alexey Milovidov).
• A query like SELECT * FROM table WHERE pk_id = '' where pk_id is the primary key and of String type will now correctly use the primary key index for filtering granules. #99027 (Shankar Iyer).
• Fix DEPENDENCIES_NOT_FOUND exception in Kafka engine when materialized view is detached while the background thread is streaming data. #99028 (Alexey Milovidov).
• Fix exception when creating a table with an EPHEMERAL column that has the same name as a virtual column (e.g., _part_offset). #99031 (Alexey Milovidov).
• Fix misleading “inflate failed: buffer error” when reading non-existent compressed files via url() table function with glob patterns. Now returns empty result as expected when http_skip_not_found_url_for_globs is enabled. #99034 (Alexey Milovidov).
• Fix server crash (std::terminate) when executing ALTER TABLE ... DROP PART on a patch part after a schema change (e.g. ADD COLUMN). The crash was caused by missing system columns (_part) in the empty coverage part metadata, leading to an uncaught exception inside a NOEXCEPT_SCOPE. #99036 (Peng).
• ClickHouse server process could crash if there was a memory limit exceeded exception thrown during a cached disk read. That is now fixed. #99042 (Shankar Iyer).
• Fix LOGICAL_ERROR when querying a table that has both a ROW POLICY and an ALIAS column using dictGet. The issue was caused by premature access to the table expression during ALIAS column resolution in the new analyzer. #99065 (Peng).
• Fix out of bounds error when user tries to query only virtual columns from Iceberg table with Avro format for data. It’s super rare scenario so not marking it as critical. Fixes #88238. #99080 (alesapin).
• Fix segfault in recursive CTE with remote() + view(). #99081 (Konstantin Bogdanov).
• Skip unnecessary extra index analysis when read-in-order optimization is applied. #99084 (Vladimir Cherkasov).
• Fixed a crash triggered by a memory limit exception thrown during patch part application. #99086 (Anton Popov).
• Fix debug assertion in DDLWorker caused by stale first_failed_task_name after a ZooKeeper entry is deleted during reinitialization recovery. #99099 (Antonio Andelic).
• Fixed rebuild of text indexes on merges with TTL. #99107 (Anton Popov).
• Fix crash in ALTER TABLE ... REMOVE SETTINGS query for Iceberg table engine. Fixes #86330. #99108 (alesapin).
• Fixes a bug in query_plan_convert_any_join_to_semi_or_anti_join optimization, returning an incorrect result for unmatched rows. Related: https://github.com/ClickHouse/ClickHouse/pull/95995. #99112 (Yarik Briukhovetskyi).
• Fix LOGICAL_ERROR exception in ASTColumnsExceptTransformer::transform. #99119 (Pablo Marcos).
• Fix RBAC bypass that allowed users to obtain table structure via DESCRIBE TABLE or CREATE TABLE AS on table functions (mysql(), postgresql(), sqlite(), arrowFlight(), jdbc(), odbc(), etc.) without the required source access privileges. For functions that infer schema from remote servers, this also allowed triggering outbound connections (SSRF) without authorization. #99122 (pufit).
• Fix Keeper crash (segfault in NuRaft) during dynamic reconfiguration and leadership transfer. #99133 (JIaQi Tang).
• Fix crash on usage of Buffer table with SAMPLE when destination does not support it. #99141 (Kseniia Sumarokova).
• Fix LOGICAL_ERROR due to patch parts column order mismatch. #99164 (Pablo Marcos).
• Fix very rare crash when Iceberg table contains files of mixed format (ORC and Parquet). Fixes #88126. #99168 (alesapin).
• Fix max_execution_time not being applied for backup/restore. #99205 (Kseniia Sumarokova).
• Fix insert_deduplication_token being silently ignored for INSERT SELECT queries without ORDER BY ALL. Previously, deduplication was disabled entirely for unsorted INSERT SELECT, even when an explicit user token was provided. Now, providing insert_deduplication_token is sufficient to enable deduplication regardless of ORDER BY ALL. #99206 (Desel72).
• Fix excessive access checks during InverseDictionaryLookupPass optimization by checking CREATE_TEMPORARY_TABLE grant once before the pass instead of for every visited node. #99210 (Mikhail Artemenko).
• Fix clickhouse format --obfuscate producing invalid SQL by obfuscating skip index types, compression codec names, database engine names, and dictionary layout/source definitions. #99260 (Raúl Marín).
• Fixed a bug where, in some cases, comparing between Time[64] and DateTime[64] types was confusing; now, in cases like this, Time[64] values are promoted to DateTime[64] by adding the 1970-01-01 as a date part. #99267 (Yarik Briukhovetskyi).
• Clamp settings constraints in DDL worker for distributed DDL queries. #99317 (Pablo Marcos).
• Fix minor issues with TOTP authentication: the --one-time-password CLI option with empty password, validation of <digits> and <period> configuration values. #99322 (Vladimir Cherkasov).
• Fix logical error unordered_map::at: key not found in Avro output format when serializing Enum8/Enum16 columns with values not present in the enum definition. #99332 (Desel72).
• Fix CHECK TABLE with sparse serialization inside Tuple with Dynamic. Closes #96588. #99351 (Pavel Kruglov).
• Fixed too strict validation of text index preprocessor. #99359 (Anton Popov).
• Fix compatibility when upgrading replicated tables with implicit minmax indices from 25.10 to newer versions. #99392 (Raúl Marín).
• Removed support for negated functions (notEquals, notLike, notIn) in text index analysis. These functions could never skip any granules, so analyzing the index for them only added overhead without any benefit. #99393 (Anton Popov).
• Fix optimize_skip_unused_shards with a new analyzer for the case when a Distributed table was used inside an IN subquery. #99436 (Nikolai Kochetov).
• Fix heap-use-after-free in INTERSECT/EXCEPT when the query produces duplicate column names. #99471 (Alexey Milovidov).
• Fix logical error in ALTER TABLE ... DROP PART when a typed query parameter is used for the part name. #99489 (Alexey Milovidov).
• Fixed NOT_FOUND_COLUMN_IN_BLOCK exception when a text index predicate (e.g. hasAllTokens) is referenced in both SELECT and WHERE clauses via an alias. #99504 (Anton Popov).
• Fix incorrect results when using hasAllTokens with OR across columns that have separate text indexes. #99505 (Anton Popov).
• Initialize page cache in clickhouse-local so that page_cache_max_size setting takes effect. #99510 (Alexey Milovidov).
• Fixed rare incorrect marking of a data part as broken and detaching it after DETACH/ATTACH TABLE query. #99529 (Anton Popov).
• Fix std::length_error exception when querying empty system tables with Pretty format via HTTP interface. #99541 (Alexey Milovidov).
• Fix LOGICAL_ERROR when using ALTER TABLE ADD COLUMN to create an EPHEMERAL column with the same name as a virtual column (e.g. _part_offset). #99549 (Alexey Milovidov).
• Fix VectorSimilarityIndexCache entries never being evicted after part removal due to mismatched cache keys. #99575 (Seva Potapov).
• Forbid reading Google credentials from a local file. This setting is insecure because it allows reading other credentials if the file path is known. #99584 (Konstantin Vedernikov).
• Fix performance degradation in the analyzer. Prune unused columns from ARRAY JOIN. #99587 (Dmitry Novik).
• Fixed reading of text index in table with existing lightweight deletes and row policies. #99661 (Anton Popov).
• Fix nullptr dereference in Parquet reader when filter-in-decoder path encounters filtered-out pages. Closes #99676. #99677 (Alexey Milovidov).
• Fix incorrect seek in AsynchronousReadBufferFromFileDescriptor with O_DIRECT. Closes #99358. #99678 (Pavel Kruglov).
• Fixed heap-buffer-overflow in CompressionCodecT64 and process abort in CompressionCodecMultiple when decompressing malformed compressed data. Both issues were found by new libFuzzer targets. The codecs now throw an exception instead of crashing. #99680 (Rahul).
• Delay processing until server has finished loading all the tables. #99700 (Seva Potapov).
• Fix MySQL dictionary source bypassing RemoteHostFilter for inline DDL params. #99720 (Shaohua Wang).
• Fix logical error when iterating over data lake tables in system.tables. #99739 (Konstantin Vedernikov).
• Fixed analysis of predicates with the IN function by text index with the preprocessor. Fixed the collision of searched tokens in the text index, which could lead to incorrect results. #99755 (Anton Popov).
• Fix infinite loop when reading Npy format files with negative shape dimensions. #99812 (Desel72).
• Fix global-buffer-overflow in CRC32 function on FixedString arguments when evaluated with zero rows during query plan header computation. #99835 (Alexey Milovidov).
• Fix crash (null pointer dereference) when executing ALTER TABLE ... MODIFY COLUMN ... COMMENT on Iceberg tables. #99838 (Desel72).
• Fix aggregate_functions_null_for_empty setting to work with aggregate functions returning non-Nullable types such as Array or Map (e.g., groupArray, sumMap). #99839 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception in midpoint function when called with mixed signed/unsigned integer types. #99867 (Alexey Milovidov).
• Fix “Block structure mismatch” exception in queries with HAVING clause where the filter expression contains both an aggregate wrapped in a NULL-producing function and materialize(0). #99915 (Alexey Milovidov).
• Fix assertion failure in sipHash128Keyed (and similar keyed hash functions) when the data argument is a Map with array keys or other nested array types. #99921 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception “Not-ready Set” in IN function during query plan optimization with convertAnyJoinToSemiOrAntiJoin. #99939 (Alexey Milovidov).

​

Build/Testing/Packaging Improvement

• Reduce compilation time by removing heavy header includes and moving expensive template instantiations out of headers. #97893 (Raúl Marín).
• Reduce compile time of arithmetic functions and related headers by shrinking template dispatch matrices and removing heavy includes. #98204 (Raúl Marín).
• Use mongo-c-driver 2.2.2. #98304 (Konstantin Bogdanov).
• Use postgres REL_18_3. #98306 (Konstantin Bogdanov).
• Enable jemalloc allocator for UBSan builds to avoid RSS accumulation from glibc malloc’s poor memory reclamation behavior. #98444 (Alexey Milovidov).
• Use Rust v0 symbol mangling and strip internal symbols from PRQL library to reduce symbol name bloat from parser combinator libraries. #98446 (Alexey Milovidov).
• Add TPC-H benchmark suite and TPC-DS README to tests/benchmarks. #98495 (Raufs Dunamalijevs).
• Add correctness tests for all 99 TPC-DS queries. #99204 (Raufs Dunamalijevs).
• Add integration test reproducing DDL CREATE TABLE + ALTER with offline replica bug (#44070), marked as expected failure. #99259 (Raufs Dunamalijevs).
• Integrate jemalloc with je_ prefix and remove usage of linker’s --wrap. #99342 (Azat Khuzhin).

​

ClickHouse release 26.2, 2026-02-26. Presentation, Video

​

Backward Incompatible Change

• Deduplication is turned ON for all inserts by default. It was OFF before for async inserts and for MV’s, but it was ON for sync inserts. The goal is to have the same defaults for both ways of inserts. If you have deduplication explicitly disabled on your cluster, you have to explicitly set deduplicate_insert='backward_compatible_choice' to keep the old behavior. The same with deduplicate_blocks_in_dependent_materialized_views. #95970 (Sema Checherinda).
• Improved storage format of statistics. All statistics are now stored in a single file. #93414 (Anton Popov). If you didn’t explicitly enable table statistics, you can ignore this item.
• Limit S3(Azure)Queue in-memory metadata. System tables are renamed from azure_queue to azure_queue_metadata_cache and system.s3queue to s3queue_metadata_cache. #95809 (Kseniia Sumarokova).
• Previously, applying a function to a Variant column silently returned NULLs when a variant sub-type was incompatible with the function; now it throws an exception, which may break queries that relied on the silent NULL behavior. #95811 (Bharat Nallan).
• DATE columns from PostgreSQL are now inferred as Date32 in ClickHouse (in previous versions they were inferred as Date, which led to overflow of the values outside of a narrow range). Allow inserting Date32 values back to PostgreSQL. Closes #73084. #95999 (Alexey Milovidov).
• The semantics of the do_not_merge_across_partitions_select_final setting were made more obvious. Previously, the feature could be automatically enabled when the setting was not explicitly set in the configs. It caused confusion repeatedly and, unfortunately, led to some issues in production. Now, the rules are simpler: do_not_merge_across_partitions_select_final=1 enables the functionality unconditionally. If do_not_merge_across_partitions_select_final=0, then automatic is used only if the new setting enable_automatic_decision_for_merging_across_partitions_for_final=1 and not used otherwise. To preserve the old behaviour as much as possible, the defaults were set to do_not_merge_across_partitions_select_final=0 and enable_automatic_decision_for_merging_across_partitions_for_final=1. #96110 (Nikita Taranov).
• When creating an S3 table with explicitly specified columns, ClickHouse now validates that those column names actually exist in the remote file’s schema. Queries that previously worked with mismatched column names will now fail at table creation time. This closes #96089. #96194 (Konstantin Vedernikov).
• Forbid using subqueries in ORDER BY and other table key expressions. #96847 (Alexey Milovidov).
• Enable apply_row_policy_after_final by default. Initially, when optimize_move_to_prewhere_if_final=0, both ROW POLICY and PREWHERE respect FINAL and were applied after FINAL. This was broken by #87303, which ignored the optimize_move_to_prewhere_if_final for the ROW POLICY filter. To fix this, this PR enables the setting apply_row_policy_after_final introduced in #91065. With apply_row_policy_after_final enabled, ROW POLICY would continue to respect FINAL by default, as previously. This PR is an incompatible change because it changes the behaviour for optimize_move_to_prewhere_if_final=1. Now, to get the ROW POLICY applied before FINAL, apply_row_policy_after_final should be used instead of optimize_move_to_prewhere_if_final. #97279 (Nikolai Kochetov).
• The Date type is now serialized as Arrow’s native date32 type in Arrow/ArrowStream formats, instead of uint16. Tools like PyArrow will now correctly see the column as a date type. The old behavior can be restored with the output_format_arrow_date_as_uint16 setting. Reading old Arrow files that used uint16 for Date columns is still supported. #96860 (Alexey Milovidov).

​

New Feature

• Users can now use ClickStack (an observability UI) directly from ClickHouse, useful for debugging and local development. #96597 (Aaron Knudtson).
• Support time-based one-time password (TOTP) as an authentication method. #71273 (Vladimir Cherkasov).
• Add lazy_load_tables database setting. When enabled, tables are not loaded during database startup — a lightweight StorageTableProxy is created instead and the real table engine is materialized on first access. #96283 (xiaohuanlin).
• Added input_format_max_block_wait_ms setting to emit data blocks by timeout and allowed processing of remaining data when an HTTP connection is closed unexpectedly. #94509 (Mostafa Mohamed Salah).
• Google BigLake catalog integration. This closes #95339. #97104 (Konstantin Vedernikov).
• Added system table system.tokenizers which shows all available tokenizers. #96753 (Robert Schulze).
• Add new system table system.user_defined_functions to monitor UDF loading status and configuration. #90340 (Xu Jia).
• Add system.jemalloc_stats table exposing jemalloc memory allocator statistics (via malloc_stats_print) for diagnosing memory usage on servers built with jemalloc. Also add a /jemalloc.html HTTP endpoint on the ClickHouse HTTP interface for interactive visualization of these statistics. #97077 (Antonio Andelic).
• Added system.jemalloc_profile_text table for reading and analyzing jemalloc heap profiles. The output format is controlled by the jemalloc_profile_text_output_format setting (raw, symbolized, or collapsed; default collapsed). Inline frame resolution is controlled by jemalloc_profile_text_symbolize_with_inline (when enabled, inline frames are included at the cost of slower symbolization; when disabled, they are skipped for faster output). For the collapsed format, jemalloc_profile_text_collapsed_use_count controls whether stacks are weighted by live allocation count (true) or live bytes (false, default). This enables easier memory profiling and flame graph visualization of jemalloc heap profiles. Fixes #93248. #97218 (Antonio Andelic).
• Add the default_dictionary_database setting, which lets ClickHouse resolve external dictionaries referenced without a database qualifier in a specified default database. This simplifies migration from XML-defined global dictionaries to SQL-defined per-database dictionaries—allowing existing dictionary queries (e.g. dictGet(‘name’, …)) to continue working without modification. #91412 (Dmitrii Plotnikov).
• Support auxiliary zookeeper for DatabaseReplicated. #91683 (RinChanNOW).
• Implement new table function primes and new system table system.primes that contains prime numbers in ascending order. Closes #90839. #92776 (Nihal Z. Miaji).
• Async inserts support parallel quorum. The inserted data is replicated to the quorum. If duplicates are found, query waits until previously inserted data is replicated as well. #93356 (Sema Checherinda).
• Added functions colorOKLABToSRGB, colorSRGBToOKLAB to convert value from sRGB to OKLAB and vice versa. #93361 (Pranav Tiwari).
• A new deduplicate_insert setting which overrides insert_deduplicate and async_insert_deduplicate. #94413 (Sema Checherinda).
• Server setting insert_deduplication_version makes it possible to migrate on unified deduplication hash. #95409 (Sema Checherinda).
• Add xxh3_128 hashing function. #96055 (Raúl Marín).
• Added OPTIMIZE <table> DRY RUN PARTS <part names> query to simulate merges without committing the result part. It may be useful for testing purposes: verifying merge correctness in the new version, deterministically reproducing merge-related bugs, and reliably benchmarking merge performance. #96122 (Anton Popov).
• Add a new check enabled by default via setting check_named_collection_dependencies to avoid dropping named collections used by tables. #96181 (Pablo Marcos).
• Added system.fail_points to inspect existing failpoints in the server and whether they are enabled or not. This is going to help automate testing. #96762 (Pedro Ferreira).
• Add role-based access to Glue catalog. Use settings aws_role_arn and, optionally, aws_role_session_name. #90825 (Antonio Andelic).
• Added a setting add_minmax_index_for_temporal_columns that, when enabled, automatically creates minmax indexes for all Date, Date32, Time, Time64, DateTime, and DateTime64 columns. #93355 (Michael Jarrett).
• Support for extended table aliases for JOINs (queries like SELECT * FROM (SELECT 1) AS t(a) JOIN (SELECT 1) AS u(b) ON a = b). Closes #95131. #95331 (Yarik Briukhovetskyi).
• Added support for ALTER TABLE RENAME COLUMN for Iceberg tables. Previously only ADD COLUMN, DROP COLUMN, and MODIFY COLUMN were supported. #97455 (murphy-4o).

​

Experimental Feature

• The text index is now GA. #96794 (Robert Schulze).
• The QBit data type for quantized bit-packed vector storage (used for approximate nearest-neighbor search) is now generally available and no longer requires enabling an experimental setting. #95358 (Raufs Dunamalijevs).
• Vector search in ClickHouse can now use replicas in the cluster to distribute the load and search of vector index parts. This enables ClickHouse to support large vector indexes that exceed the memory capacity of a single VM. #95876 (Shankar Iyer).
• Add server-side AST fuzzer controlled by ast_fuzzer_runs and ast_fuzzer_any_query settings. When enabled, the server runs randomized mutations of each query after its normal execution, discarding the results. #97568 (Alexey Milovidov).
• Add iif function to the experimental KQL dialect. #94790 (happyso).
• Schema inference now respects allow_experimental_nullable_tuple_type. When enabled, it allows inferred tuple types to be Nullable(Tuple(...)), so missing nested objects can become NULL instead of a tuple of NULL elements. #95525 (Nihal Z. Miaji).
• The use_statistics_cache setting is now enabled by default, so column statistics are cached in memory to speed up query optimization without needing to reload them from each part. #95950 (Han Fei).

​

Performance Improvement

• Allow any deterministic expression in Primary Key to be used for data skipping (e.g. ORDER BY cityHash64(user_id)/ ORDER BY length(user_id)). For deterministic expressions, ClickHouse can apply the expression to query constants and use the result in the primary key index for predicates like =, IN, and has. If the expression is also injective (e.g. ORDER BY hex(p) or ORDER BY reverse(tuple(reverse(p), hex(p)))), we can effectively use the index for the negated forms: !=, NOT IN, and NOT has. Closes #10685. Closes #82161. #92952 (Nihal Z. Miaji).
• Improved storage format of statistics. All statistics are now stored in a single file. #93414 (Anton Popov).
• Allow parallelized read for remote table engines/functions in the filesystem cache. #71781 (Kseniia Sumarokova).
• Allow using userspace page cache with local files and object storage table functions. #77874 (Michael Kolupaev).
• Avoid unnecessary memcpy in userspace page cache. #77884 (Michael Kolupaev).
• The default for concurrent_threads_scheduler is now max_min_fair instead of fair_round_robin. This improves fairness under high load by prioritizing queries with fewer allocated slots, so short-running queries aren’t penalized by long-running ones. #95300 (Sergei Trifonov).
• If a FINAL query used primary key condition for filtering followed by skip indexes for other conditions, the PrimaryKeyExpand processing step will now only check the initial shortlisted primary key ranges for intersection. #94903 (Shankar Iyer).
• When using parallel replicas with table functions like s3(...), queries with a single subquery wrapping the table function are now automatically parallelized across replicas, whereas previously only direct table function references were parallelized. Closes #92264. #96332 (phulv94).
• Enable splitting data and system files in cache into separate segments. #87834 (MikhailBurdukov).
• Speed up some hash join operations by implementing dynamic dispatch for ColumnVector::replicate. #79573 (Raúl Marín).
• Performance improvement for parallel hash join in cases of complex predicates. Previously, we were processing non-joined rows in one thread, which is suboptimal the idea of the optimization is to parallelize the processing of the non-joined rows across multiple threads. Can be toggled by the parallel_non_joined_rows_processing setting. Enabled by default. #92068 (Yarik Briukhovetskyi).
• Slightly optimize parsing of JSON type. #93614 (Pavel Kruglov).
• Improve memory footprint of AST. Optimization makes sense as fields are not used when highlighting is not used and there is no VALUES parsing. #93974 (Ilya Yatsishin).
• Optimize memory consumption of named Tuple AST objects. Place column names as strings in tuple object instead of having them in generic AST literal nodes. #94704 (Ilya Yatsishin).
• Devirtualization is improved with additional linker options. #94737 (Nikita Taranov).
• Improve clone replica performance for ReplicatedMergeTree tables with many parts by batching ZooKeeper requests. #94847 (c-end).
• When read step already has PREWHERE filters a new filter couldn’t be added. This change postpones PREWHERE optimization until after JOIN runtime filter optimization so that runtime filters can be also pushed to PREWHERE. #95838 (Alexander Gololobov).
• Speedup T64 codec compression by using dynamic dispatch on x86. #95881 (Raúl Marín).
• Speed up uniq over numeric types by batching inserts when possible (not null, not -If, no GROUP BY, no IPv6 or String). #95904 (Raúl Marín).
• Low-level optimizations for Keeper: ZooKeeper::observeOperations has been found to account for >20% of the ZooKeeper receive thread CPU consumption. This change addresses that by: 1. For AggregatedZooKeeperLog::stats, use CityHash64 instead of SipHash which is >10x faster. 2. For Coordination::ErrorCounter, use std::array<std::atomic<UInt32>, N> instead of std::unordered_map and std::mutex. #95962 (Miсhael Stetsyuk).
• Remove 64-byte alignment for ProfileEvents::Counter to save memory. #96097 (Azat Khuzhin).
• Memory optimization: trim size of CachedOnDiskReadBufferFromFile structure 50x. #96098 (Azat Khuzhin).
• Don’t copy old data on hash table resizing if it’s empty. #96180 (Raúl Marín).
• Support JOIN runtime filters for RIGHT OUTER JOINs. #96183 (Hechem Selmi).
• The optimization enable_join_runtime_filters is now the default. #89314 (Alexey Milovidov).
• Previously, text index direct read optimization was applied only when all parts had a materialized text index. This PR adds partial support: if some parts have a materialized text index, those parts will use it, while parts without a materialized text index will fall back to executing the original filter expression. #96411 (Anton Popov).
• Added minmax secondary indexes on time columns and bloom_filter indexes on query_id/initial_query_id columns to system log tables for faster filtering. #96712 (Alexey Milovidov).
• Lazy materialization optimization is now applied to all branches of a UNION ALL query, not just the first one. Queries that combine multiple sorted and limited reads from different MergeTree tables via UNION ALL will now benefit from deferred column reading on every branch, reducing I/O. #96832 (Federico Ginosa).
• Optimize minmax skip index computation during INSERT by removing an unnecessary data copy and enabling vectorized min/max calculation for numeric columns. #97392 (Raúl Marín).
• Storage DeltaLake now takes count() result from delta lake metadata and shows correct table stats in system.tables (total bytes/rows). #96190 (Kseniia Sumarokova).
• The unused columns are removed also from the reading step in case of reading from a MergeTree. It is especially useful when a filter is pushed down into PREWHERE. #89982 (János Benjamin Antal).
• Improved processing SHOW TABLES query by fetching only names of tables and improved getLightweightTablesIterator to return structure containing only table names. resolves #93835. #94467 (Smita Kulkarni).
• Improve assumeNotNull, coalesce, ifNull to enable primary key and skip index pruning for range predicates when key columns are wrapped in these functions. Closes #94689. #94754 (Nihal Z. Miaji).
• Add with_data & with_stat extension to getChildren Keeper request. This allows fetching not only the list of children, but also their stat and/or data in a single operation. #94826 (Nikolay Degterinsky).
• The index analysis is done only once (in most cases) regardless of whether we end up executing a local plan or a plan with parallel replicas. #94854 (Nikita Taranov).
• Allow to enable distributed index analysis based on amount of parts (distributed_index_analysis_min_parts_to_activate) and indexes size (distributed_index_analysis_min_indexes_size_to_activate). #95216 (Azat Khuzhin).
• Enable PREWHERE optimization for Iceberg tables. #95476 (Konstantin Vedernikov).
• Reduce the memory footprint of some AST classes. #95514 (Raúl Marín).
• Limit the number of pipeline streams generated with split_intersecting_parts_ranges_into_layers enabled. Helps to avoid excessive memory consumption. #96478 (Nikita Taranov).
• Implement equivalent sets optimization for multiple joins. Queries with multiple consecutive INNER JOIN operations now benefit from improved filter pushdown optimization. When tables are joined on equivalent columns (e.g., t1 JOIN t2 ON t1.id = t2.id JOIN t3 ON t2.id = t3.id WHERE t1.id > 10), filters applied to any table in the chain are automatically pushed down to all tables. Close #96550. #96596 (Vladimir Cherkasov).
• Optimize delta lake metadata scan. Uses changes from delta-kernel PR https://github.com/delta-io/delta-kernel-rs/pull/1827. #96686 (Kseniia Sumarokova).
• In Replicated database, don’t update the cached cluster for every dummy query. #96897 (Tuan Pham Anh).
• Use the primary key index when filtering with startsWithUTF8 if the prefix contains only ASCII characters. #97055 (vkcku).

​

Improvement

• Add OpenTelemetry tracing for Keeper requests. #91332 (Miсhael Stetsyuk).
• New configuration options: logger.startup_console_level & logger.shutdown_console_level to allow for overriding the console log level during the startup & shutdown of ClickHouse respectively. #95919 (Garrett Thomas).
• Respect command-line overrides when reloading configuration. Closes #80294. #80295 (Alexey Milovidov).
• Allow key-value overrides for named collection parameters in mongodb table function. #89616 (vanchaklar).
• The read-in-order optimization for Iceberg tables now works with complex sorting functions like icebergBucket and icebergTruncate, not just simple column references. #90256 (Konstantin Vedernikov).
• Add a new column named parts_postpone_reasons in system.mutations to improve diagnostics, which shows the parts postpone reasons. #92206 (Shaohua Wang).
• Track changes in the number of rows to read (due to inserts/deletes or query condition cache usage) in DataflowStatisticsCache. #93636 (Nikita Taranov).
• Support SYSTEM RESET DDL WORKER [ON CLUSTER] query. It requests to reset the state of DDLWorker in its main thread. It is helpful to refresh the replica active when host IDs are updated. #93780 (Tuan Pham Anh).
• Support mutation_ids in system.part_log for event type of MUTATE_PART and MUTATE_PART_START. #93811 (Shaohua Wang).
• Background operations (Mutate, Merge) can now be configured independently via ‘background’ profile. Previously such operations shared settings with regular queries via ‘default’ profile. #93905 (Arsen Muk).
• Add more information to system.crash_log. #94112 #95857 (Miсhael Stetsyuk).
• Added new QueryNonInternal metric to track the number of executing non-internal queries. This metric is exposed as ClickHouseMetrics_QueryNonInternal and helps operators monitor query concurrency against the max_concurrent_queries limit, which only applies to non-internal queries. #94284 (Ashwath Singh).
• Support input bytes statistics collection for columns from compact parts in RuntimeDataflowStatisticsCacheUpdater. #94626 (Nikita Taranov).
• Add a check for Keeper misconfiguration leading to cluster assembly failures. Closes #60932. #94682 (Konstantin Bogdanov).
• Improve JSON prefixes deserialization during part loading. #94848 (Pavel Kruglov).
• Refactor the write using full INSERT pipeline, which triggers materialized views on the target table. #94890 (Kai Zhu).
• Use vector similarity search plan optimizations only if the index exists for the search column. #94998 (Eduard Karacharov).
• Check for the total memory limit before the user authentication and throw (total) memory limit exceeded if the total limit is more than allowed. #95003 (Nikolai Kochetov).
• Added the throw_on_unmatched_row_policies configuration option which, when enabled, throws an exception if a user queries a table that has row policies but none of them apply to that user — preventing the ambiguous behavior of returning all rows due to access control misconfiguration. #95014 (Vitaly Baranov).
• Dynamic update s3 access tokens in long queries with unity catalog. This closes #93981. #95069 (Konstantin Vedernikov).
• Disable jemalloc’s dirty page decay if ClickHouse is under sustained memory pressure for memory_worker_decay_adjustment_period_ms milliseconds. Enable jemalloc’s dirty page decay back if ClickHouse is working under normal conditions for same amount of time. #95145 (Antonio Andelic).
• S3Queue auxiliary Zookeeper support using keeper_path setting from s3Queue. #95203 (Diego Nieto).
• Respect max_parts_to_merge_at_once in TTL drop part merges. #95315 (Kseniia Sumarokova).
• Add connection_address and connection_port to query_log to reflect physical connection (address and port are replaced when connected through proxy and auth_use_forwarded_address=1). #95471 (Yakov Olkhovskiy).
• Fix incorrect memory accounting for the query conditions cache. The key problem was that it didn’t take into account the cache key that composed of several strings (like part_name, the table id and the whole SQL condition). #95478 (Nikita Mikhaylov).
• Server started with the embedded configuration will allow to manipulate users and grants, saving them to the access directory, as the regular configuration does. This improves testing. Also enabled all access_control_improvements in the embedded config and in clickhouse-local. #95481 (Alexey Milovidov).
• Improved S3 authentication error messages to include a hint to check credentials when access is denied. #95648 (Gerald Latkovic).
• Enable statistics cache and set the update period of cache to 300s. #95841 (Han Fei).
• Add component name to system.aggregated_zookeeper_log. #95882 (Antonio Andelic).
• Skip object storage reads when querying DeltaLake tables from system.tables. #95899 (Antonio Andelic).
• Enable enable_max_bytes_limit_for_min_age_to_force_merge by default if the compatibility setting is 26.2 or higher. #95917 (Christoph Wurm).
• Delta Lake is now available on macOS. Closes #95979. #95985 (Alexey Milovidov).
• In previous versions, when combining conflicting ALTER expressions with UPDATE and RENAME COLUMN, a logical error was thrown instead of a proper exception. Closes #70678. #96022 (Alexey Milovidov).
• Improve the help output for all ClickHouse applications and add a --no-sudo option with a few fixes. This is a continuation of #58244 from Ilya Yatsishin. #96025 (Alexey Milovidov).
• Add distanceCosine alias for cosineDistance because all other distance functions already have alias of this form. #96065 (Raufs Dunamalijevs).
• Add support for the with_data Keeper extension to improve table fetching in Database Replicated. #96090 (Nikolay Degterinsky).
• Update chdig to v26.2.1 (new features and MacOS support). #96113 (Azat Khuzhin).
• Improve filter pushdown for numbers and primes. ClickHouse can now derive conservative value bounds from WHERE conditions when exact bounds cannot be derived, and restrict sequence generation accordingly (for example, for WHERE number % 5 < 2 AND number > 100 AND number < 300, ClickHouse will only generate numbers between 100 and 300, then apply the predicate), avoiding unbounded scans. Closes #84853. Closes #93913. #96115 (Nihal Z. Miaji).
• The formatter previously wrapped SELECT in parentheses when a COMMENT clause was present to disambiguate parsing. Instead, output COMMENT before AS SELECT, which eliminates the ambiguity without parentheses. #96293 (Alexey Milovidov).
• The allow_impersonate_user config setting is now located inside the access_control_improvements section rather than being a standalone server setting. #96451 (Vitaly Baranov).
• Make core_dump.size_limit configuration setting hot-reloadable, to avoid having to restart servers for configuration changes to take place. #96524 (Miсhael Stetsyuk).
• Improves CPU and real-time profiler interoperability with socket timeouts. #96601 (Sergei Trifonov).
• Prevent the resurrection of dropped data if ADD COLUMN is run quickly after the DROP COLUMN mutation. #96713 (Alexey Milovidov).
• Change function_id type in system.instrumentation from LowCardinality(Int32) to Int32. #96726 (Copilot).
• Synchronous waiting for mutations will respect query cancellation and time limits. #96756 (Alexey Milovidov).
• Added system command SYSTEM RELOAD DELTA KERNEL TRACING <level> to be able to change delta-kernel logging, which can be useful for debugging. #96763 (Kseniia Sumarokova).
• Filtering by IP address family, i.e. dns_allow_resolve_names_to_ipv4/ipv6 settings, are applied even if DNS cache is disabled. #96810 (c-end).
• Better jemalloc introspection. #96840 (Azat Khuzhin).
• Fix /play Web UI throwing QUERY_CACHE_USED_WITH_SYSTEM_TABLE when querying system tables. #96869 (Alexey Milovidov).
• Improve Web UI: change favicon to indicate running query state; display errors from auxiliary queries (loading databases and tables) instead of silently ignoring them. Closes #85055. #96883 (Alexey Milovidov).
• Make the left panel in /play UI clickable to toggle the database list. #96884 (Alexey Milovidov).
• DROP DATABASE now drops tables in reverse dependency order, improving crash-safety when the database contains tables with loading dependencies (e.g. Distributed tables using joinGet). #97057 (Alexey Milovidov).
• Bump yaml-cpp to prevent skipping invalid YAML. #97333 (Azat Khuzhin).
• Show a loading indicator in the play.html sidebar while tables are being fetched. #97531 (Alexey Milovidov).
• Add a copy-to-clipboard button for raw query results in the built-in web UI (play.html). #97532 (Alexey Milovidov).
• Fix query obfuscator (clickhouse-format --obfuscate) to produce parseable SQL in more cases. #97584 (Alexey Milovidov).

​

Bug Fix (user-visible misbehavior in an official stable release)

• After metadata-only ALTERs, such as extending the elements of Enums, the optimization of aggregation with projection may end up producing an exception. #84143 (Alexey Milovidov).
• Materialized views now use the database where they were created as execution context, meaning that: - it is possible to omit explicit database qualification on names referred in view’s select query - if no explicit database qualification is given, the same database where the materialized view was created is assumed. #88193 (Dmitry Kovalev).
• Fix query parameter substitution in CREATE USER authentication methods when using ON CLUSTER. Query parameters in authentication methods (e.g., password) were not being replaced, causing UNKNOWN_QUERY_PARAMETER errors on remote nodes. #92777 (xiaohuanlin).
• Fixed inconsistencies in text index analysis for has, mapContainsKey, and mapContainsValue functions. Previously, queries using these functions could return different results depending on whether the expression was evaluated with or without a text index. #93578 (Anton Popov).
• Fix crash when attaching a table to a MaterializedPostgreSQL database if dropReplicationSlot throws during stack unwinding. #96871 (Alexey Milovidov).
• Backups could crash the server if you do many concurrent backups clashing over the same files. #93659 (Alexey Milovidov).
• Fixes queries with parallel replicas and JOIN with non-MT table. Closes #92056. #93902 (Igor Nikonov).
• Fixes an issue when Iceberg columns with dot in names returned NULL as values. #94335 (Mikhail Koviazin).
• Fixed handling of UTF8 strings in stringJaccardIndexUTF8 and improve performance. #94613 (Joanna Hulboj).
• Fix possible overflows in WITH FILL STALENESS (that leads to UB or/and endless loops). Fix possible endless loop due to big jumps. Add old analyzer support (mostly for stress tests). #94663 (Azat Khuzhin).
• Fix possible hung distributed queries when hostnames resolve to multiple addresses and a remote replica freezes. #94726 (c-end).
• Fix invalid result on joining multiple table expressions, when leftmost table expression is a -Cluster table function. Resolves #89996. #94748 (Konstantin Bogdanov).
• Fix incorrect primary key and skip index pruning for predicates involving toWeek, toYearWeek, toStartOfWeek, toLastDayOfWeek, and toDayOfWeek, and fix exceptions in some of these functions for valid queries with LowCardinality(String). #94816 (Nihal Z. Miaji).
• Remove unnecessary skip permissions check in ATTACH queries for a view with SQL Security. This prevents potential privilege escalation when a user attaches a view with a definer without validating required access. #94865 (pufit).
• Fixes a crash during ReplicatedMergeTree startup caused by concurrent removal of delete_tmp_* directories. #94892 (myeongjun).
• Fix INSERT into Iceberg tables with materialized views losing deduplication information, which caused an exception. #94938 (Daniil Ivanik).
• Fix a bug where SYSTEM DROP QUERY CACHE TAG 'TAGNAME' ON CLUSTER <CLUSTERNAME> would drop the full cache on the cluster. #94978 (Rory Crispin).
• Preserve constant index granularity (use_const_adaptive_granularity) after Vertical merges (v2 with a fix for Nested, and in general). #95013 (Azat Khuzhin).
• Fixes race in filesystem cache in version 26.1 after [ClickHouse/ClickHouse#82764](https://github.com/ClickHouse/ClickHouse/pull/82764). #95042 (Kseniia Sumarokova).
• Fix postgresql() table function canceling by KILL QUERY and cancel query (Ctrl+C) in clickhouse-client. #95136 (Roman Vasin).
• Fixed type inference for qualified columns from source tables when multiple joins are used with USING clause. Previously, subsequent joins incorrectly updated types of underlying source columns to a common supertype even when the column was not involved in that join (e.g., in SELECT t2.a FROM t1 LEFT JOIN t2 USING (a) LEFT JOIN t3 USING (a), the t2.a column is only used by the first join, so its type should be the supertype of t1.a and t2.a, excluding t3.a). This could lead to logical errors or crashes when functions expected different column types than what actually appeared in the execution plan. #95157 (Vladimir Cherkasov).
• Make column transform only once during getting content of manifest .avro list and files. #95164 (Daniil Ivanik).
• Fix incorrect calculation of JSON column sizes that could lead to excessive memory usage or wrong column statistics. #95207 (Azat Khuzhin).
• Fixed inaccurate memory accounting when applying large patch parts after lightweight updates. Previously, applying large patches could cause excessive memory usage and result in the server process being killed by the OOM killer. #95231 (Anton Popov).
• Fix undefined behavior that could cause incorrect results or an exception when a distributed query with max_parallel_replicas fell back to a local replica during index analysis. #95263 (Azat Khuzhin).
• Fix aggregation of sparse columns for sum and timeseries when group_by_overflow_mode is set to any. #95301 (Mikhail Koviazin).
• Fix a reliability issue in plain_rewritable disk policy where a network error mid-way through unlinking a metadata file could leave the storage in an inconsistent state. #95302 (Mikhail Artemenko).
• Replace Date with Date32 for iceberg. #95322 (Konstantin Vedernikov).
• The password argument of the redis table function now will be masked in the logs and system tables (e.g.: query_log). #95325 (János Benjamin Antal).
• Fix a bug where tables could be dropped or altered while a distributed query was still executing against them, potentially causing exceptions or incorrect results. #95356 (Azat Khuzhin).
• Fix a logical error in some cases when negative LIMIT/OFFSET is used in distributed queries. #95357 (Nihal Z. Miaji).
• Fix a bug where clickhouse-client would ask for password twice when connecting using ssh. #95372 (Isak Ellmer).
• Fix a data race in storage S3(Azure)Queue. #95385 (Kseniia Sumarokova).
• Fix the prewhere filter error caused by lambda expressions in prewhere. #95395 (Xiaozhe Yu).
• Fix optimize_syntax_fuse_functions to not rewrite sum/count/avg into sumCount() when the aggregate argument is Nullable. Closes #95390. #95441 (Nihal Z. Miaji).
• Avoid possible crash for distributed queries in case of cancellation. #95466 (Aleksandr Musorin).
• Fix deduplication for streaming from S3(Azure)Queue engine. #95467 (Kseniia Sumarokova).
• Fix updating row policies assigned to the initial user in distributed queries. #95469 (Vitaly Baranov).
• Fix check for encrypted disks over plain_rewritable (Fixes possible It is not possible to register multiple plain-rewritable disks with the same object storage prefix). #95470 (Azat Khuzhin).
• The mergeTreeProjection table function was missing an access check, allowing users without SELECT permission on a table (but with permissions for table functions) to read data from its projections. This fix adds the same access check that mergeTreeIndex and mergeTreeAnalyzeIndexes already have. #95480 (Alexey Milovidov).
• Fix possible logical error during reading of size subcolumn from dynamic subcolumns of Dynamic/JSON types. #95573 (Pavel Kruglov).
• Fix regression in (experimental) zero‑copy replication introduced by #94262 where shared parts could be deleted before other replicas finished fetching them. #95597 (filimonov).
• Fix crash during tupleElement applied to arrays of JSON. Closes #95581. #95647 (Pavel Kruglov).
• Fix logical error exception when using a matcher (*) inside a lambda function within a VALUES clause in a JOIN with USING. Close #93675. #95661 (Vladimir Cherkasov).
• Fixed There was an error: Cannot obtain error message logical error when waiting for a distributed DDL and dropping the Replicated database concurrently. Fixes #95539. #95664 (Alexander Tokmakov).
• Fix IN function returning incorrect results with NULL values when transform_null_in is enabled. Closes #65776. #95674 (Nihal Z. Miaji).
• Correctly handle LowCardinality Nullable types in CAST when the setting cast_keep_nullable is enabled. Closes #95670. #95747 (Alexey Milovidov).
• Fix squashing partitioned delta lake data. #95773 (Kseniia Sumarokova).
• Fix race condition for Nullable join column in runtime filters. #95775 (Hechem Selmi).
• Fix possible logical error in query with matcher (*, table.*) and analyzer_compatibility_join_using_top_level_identifier when USING column has different types in tables and select list. Close #90477. #95808 (Vladimir Cherkasov).
• Fix memory safety bugs in parallel thread pool operations (backups, aggregation, distributed queries) that could cause exceptions when an error occurred during task scheduling. #95818 (Raúl Marín).
• Fixes a crash on the DROP WORKLOAD while running concurrently with queries using the workload being dropped. #95856 (Alexey Milovidov).
• Fix slow performance when querying system tables with a user that has limited grants on many databases. Closes #89371. #95874 (pufit).
• Fix executing tupleElement on JSON with nested paths, previously it could lead to wrong query result. #95907 (Pavel Kruglov).
• Fixed a NOT_SUPPORTED error that could occur when using the direct join algorithm with an empty MergeTree table. #95935 (Vladimir Cherkasov).
• Fix the client not suggesting and auto-completing alias names for settings, closes #92190. #95945 (phulv94).
• Fix event_date in system.asynchronous_metric_log. #95947 (Raúl Marín).
• Fix skipping paths in JSON data type. Previously with JSON(SKIP path) all JSON keys with prefix path were skipped, even keys like "pathpath", so it could lead to data loss for these paths during insert. Now it’s fixed and only key "path" is skipped. #95948 (Pavel Kruglov).
• Part with unknown projections should not be marked as lost forever. #95952 (Mikhail Artemenko).
• Fix empty string becoming NULL in Join table with Nullable(String) key. Closes #71414. #96002 (Alexey Milovidov).
• Now the PostgreSQL engine can correctly read BOOLEAN[]. Closes #72754. #96006 (Alexey Milovidov).
• Fix the ProtobufList format for the case of reading from an empty file. Closes #70059. #96007 (Alexey Milovidov).
• Fix ProtobufList format producing ghost record for empty tables. Closes #72596. #96010 (Alexey Milovidov).
• Fix if function type mismatch between UInt64 and Int32 in an unusual case of distributed queries and PREWHERE, with type inference. Closes #70017. #96012 (Alexey Milovidov).
• Fix JIT compiled queries involving Bool types. #96013 (Alexey Milovidov).
• Fix logical error when reading UUID column from SQLite TEXT column. Closes #71263. #96016 (Alexey Milovidov).
• Fix SQLite engine type conversion for DateTime, Date, UUID, and other types. Closes #73481. #96017 (Alexey Milovidov).
• FixedString values were escaped incorrectly in queries to external databases, SQLite and PostgreSQL. Closes #73519. Co-authored with @jh0x. #96019 (Alexey Milovidov).
• Fix assertion failure in WindowTransform with a large PRECEDING offset. Closes #75852. #96026 (Alexey Milovidov).
• Fix a bug with possible data corruption when concurrent async inserts are using the same parameter names but contain different values. #96035 (Seva Potapov).
• Fix period for global profilers (controlled by global_profiler_real_time_period_ns and global_profiler_cpu_time_period_ns). Instead of set value, a truncated value was used, causing profiler to wakeup more than intended. #96048 (Antonio Andelic).
• Earlier if reference data file inside iceberg manifest file for position delete was present in an entry but was null, we didn’t get correct bounds for a corresponding data files. This PR fixes this bug. #96061 (Daniil Ivanik).
• Fix revoking default roles. #96103 (Vitaly Baranov).
• Fix use-after-free in the index analysis in a rare combination of disabled use_primary_key and a very large number of disjunctions of conditions that use the index. #96112 (Alexey Milovidov).
• Fix a regression with the Gorilla codec when an explicitly specified size does not correspond to the data type size, and the buffer size is too small. In previous versions, it threw an exception on decompression. Closes #78253. #96118 (Alexey Milovidov).
• Avoid a deadlock in dictionaries loaded when one dictionary references a Merge table that references it recursively. Closes #78360. #96120 (Alexey Milovidov).
• Fix use-of-uninitialized-value in formatDateTime with non-fixed-width formatters, such as MySQL and JODA-style. #96133 (Alexey Milovidov).
• The combination of settings use_const_adaptive_granularity and index_granularity_bytes (which means “non-adaptive granularity”) led to a miscalculation of the number of rows to read and an exception. #96143 (Alexey Milovidov).
• Running an invalid ALTER UPDATE mutation on object storage file-like tables, such as S3 and Azure, could lead to a nullptr dereference. Closes #92994. #96162 (Alexey Milovidov).
• Fix AccessRights::contains returning incorrect results with partial revokes. #96170 (pufit).
• Fix query condition cache hash collision for CTE folded constants, which could lead to a wrong query result. Closes #96060. #96172 (Alexey Milovidov).
• Fix possible deadlock in ProcessList. It can happen because of possible lock inversion if memory overcommit tracker triggers when we are adding task to cancellation checker. #96182 (Antonio Andelic).
• Fixed a bug where queries involving outer joins (LEFT, RIGHT, or FULL) combined with multiple INNER JOINs could return incorrect results due to illegal join reordering. When an outer join’s ON condition referenced columns from multiple previously joined tables, the optimizer failed to account for all table dependencies and could reorder the joins incorrectly, producing missing rows. Close #95972. #96193 (Vladimir Cherkasov).
• When a table has no statistics defined, ClickHouse shouldn’t try to load them. This avoids some overhead (100+ms) for checking if the statistics files exist. (issue #96068). #96233 (Han Fei).
• Fix optimize_syntax_fuse_functions to not rewrite sum/count/avg into sumCount() when the aggregate argument is LowCardinality(Nullable). Closes #95390. #96239 (Nihal Z. Miaji).
• Fix incorrect partition pruning for not IN and not has function in some cases. #96241 (Nihal Z. Miaji).
• Fix stack-use-after-scope in the vector similarity index. #96259 (Alexey Milovidov).
• Fix test runner not recognizing error hint comments when a query is preceded by a SQL comment. #96336 (Yakov Olkhovskiy).
• Fix logical error in KeyCondition when a table has a nullable primary key, and the query uses the coalesce function, which has its first argument constant. #96340 (Alexey Milovidov).
• The interaction of GROUPING SETS, group_by_use_nulls, and the Tuple data type with LowCardinality inside it, could produce an unexpected block structure in the query pipeline, which led to a logical error. This appeared after the introduction of Nullable Tuples. #96358 (Alexey Milovidov).
• It was possible to create a table with an empty expression () as an index, which led to an invalid memory access. #96363 (Alexey Milovidov).
• Fixed crash in old analyzer if JOIN and duplicated aliases. #96405 (Ilya Golshtein).
• Fix Nested columns sizes are inconsistent with local_discriminators error due to a wrong in-place filtering optimization for Variant columns. #96410 (Alexey Milovidov).
• Fix CREATE TABLE ... CLONE AS ... ignoring full qualifier of source table. #96415 (Hasyimi Bahrudin).
• Fix mysql table function canceling by KILL QUERY and cancel query (Ctrl+C) in clickhouse-client. #96437 (Roman Vasin).
• Fixes livelock in the cancellation checker thread for queries with high max_execution_time values. #96450 (Sergei Trifonov).
• Fix a logical error in some cases when fractional LIMIT/OFFSET is used in distributed queries. #96475 (Nihal Z. Miaji).
• Fix null pointer dereference in certain expressions with lambda functions. #96479 (Alexey Milovidov).
• Fix incorrect results when LowCardinality columns are converted to Nullable. #96483 (Nihal Z. Miaji).
• Fix a crash when creating an Iceberg table with an ORDER BY clause referencing a non-existent column or using a positional argument. Closes #93280. #96484 (Konstantin Vedernikov).
• Fix runtime filter exception for Tuple columns with Nullable subfields. #96509 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception in Parquet V3 native reader when PREWHERE filter column contains non-boolean UInt8 values. #96594 (Alexey Milovidov).
• Fix implicit index regeneration in replicated tables during metadata changes. #96600 (Raúl Marín).
• Fixes a datarace on DROP WORKLOAD. #96614 (Sergei Trifonov).
• Fix a bug in Iceberg table writes where partitioned inserts could produce incorrect data distribution across partition files. #96620 (Konstantin Vedernikov).
• Fixed heap-use-after-free in CREATE TABLE with constraints. #96669 (Nikita Taranov).
• Validate witness version in bech32 to avoid buffer overflow. #96671 (Raúl Marín).
• Fix system.tables returning errors when a Data Lake REST catalog is created with an invalid auth_header setting. #96680 (Han Fei).
• Fix min(timestamp) returning epoch (1970-01-01) via _minmax_count_projection after TTL merge when all rows in a block are filtered out. #96703 (Raquel Barbadillo).
• Improve validation of the iceberg_metadata_file_path setting to prevent path traversal and ensure the specified metadata file is within the table directory. #96754 (Daniil Ivanik).
• Fix crash in ifNull with Variant argument used in GROUP BY. #96790 (Alexey Milovidov).
• Fixed cache key collisions between tables with table_disk=1 setting. #96818 (Raufs Dunamalijevs).
• Fix MemoryWorker’s purging thread being stuck because of a race condition. #96819 (Antonio Andelic).
• Don’t log data with credentials in iceberg catalogs. #96831 (Konstantin Vedernikov).
• Fix exit status of clickhouse-client after server error. #96841 (Vitaly Baranov).
• Queries with CROSS JOINs and enabled parallel replicas could return incorrect result. Fixes #74337. #96848 (Igor Nikonov).
• Fixed ALTER TABLE DROP COLUMN queries failing after a lightweight update was previously performed on the same column. #96861 (Anton Popov).
• Fix stack overflow (crash) when creating archive-based backups (.zip, .tzst) to a plain_rewritable object storage disk. #96872 (Alexey Milovidov).
• Fix server crash when backup fails due to full disk or other I/O errors on the destination filesystem. #96873 (Alexey Milovidov).
• Fix EXCEPT ALL and INTERSECT ALL ignoring row multiplicities and behaving like their DISTINCT counterparts. #96876 (Alexey Milovidov).
• Fix std::terminate exception in indexOfAssumeSorted when called with incompatible types (e.g., IPv4 array with integer search value). #96877 (Alexey Milovidov).
• Fix exception Bad cast from type DB::ColumnNullable to DB::ColumnString when using window functions with group_by_use_nulls = 1 and CUBE/ROLLUP/GROUPING SETS. #96878 (Alexey Milovidov).
• Fix incorrect results when JIT-compiled expressions convert DateTime to DateTime64 (e.g., in CASE/if/multiIf with mixed DateTime types). The value was reinterpreted instead of properly scaled, producing wrong timestamps after expression compilation kicked in. #96879 (Alexey Milovidov).
• Fix logical error exception in CoalescingMergeTree when a skip index expression produces a constant column (e.g., bloom_filter on ifNotFinite(1, c0) for an integer column). #96880 (Alexey Milovidov).
• Fix wrong port number in error message when accidentally connecting with HTTP to the TLS-enabled native protocol port. #96881 (Alexey Milovidov).
• Fix per-subquery SETTINGS not being applied to table functions like file in CTEs and subqueries. #96882 (Alexey Milovidov).
• Fix memory leak of BIO objects when reading X509 certificates. #96885 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception in query analyzer when a lambda expression is passed where a concrete value is expected (e.g., as the accumulator argument of arrayFold). #96892 (Alexey Milovidov).
• Fix ColumnNullable is not compatible with original exception when casting complex nested types (Array of Nullable Tuple containing Map with Nullable Enum values). #96924 (Alexey Milovidov).
• Fix a race condition in sharded HASHED dictionary parallel loading that could occasionally cause some rows to not be loaded. #96953 (Alexey Milovidov).
• Fix a race condition between REPLACE PARTITION and background mutations that could result in both old and new data being visible after the replace. #96955 (Alexey Milovidov).
• Fix arrayJoin function producing duplicate rows when used with INNER JOIN and WHERE clause, caused by the partial predicate push-down optimization incorrectly pushing filters containing arrayJoin below a JOIN. #96989 (Alexey Milovidov).
• Fix crash (SEGFAULT) in clearCaches caused by BlockIO::operator= not moving query_metadata_cache, leading to premature destruction of cached storage snapshots and use-after-free of MergeTreeData storage. #96995 (Alexey Milovidov).
• Fix assertion failure in IfTransformStringsToEnumPass when the if or transform function returns Nullable(String) (e.g. with GROUP BY ... WITH CUBE and group_by_use_nulls = true). #97002 (Alexey Milovidov).
• Fix incorrect data written during INSERT ... SELECT with UNION ALL and JOIN, where constant string columns could receive wrong values after block squashing. #97019 (Hasyimi Bahrudin).
• Fix assert_cast exception (or silent data corruption in release builds) when building column statistics after ALTER TABLE MODIFY COLUMN changes the column type. #97027 (Alexey Milovidov).
• Fix reads of uninitialized memory in Azure Blob Storage, SSH protocol, and Arrow Flight interfaces. #97053 (Alexey Milovidov).
• Fix cases where indexes were affecting the result for queries with row policy/PREWHERE and FINAL. #97076 (Yarik Briukhovetskyi).
• Fix remaining race condition between REPLACE PARTITION and background mutations in MergeTree tables that could cause old data to reappear. #97105 (Alexey Milovidov).
• Fix implicit indices with alias columns and do full validation before creating them. #97115 (Raúl Marín).
• Fix logical error in FunctionVariantAdaptor with functions requiring const arguments like arrayROCAUC. #97116 (Bharat Nallan).
• Fix stuck mutations when PartCheckThread re-enqueues a GET_PART for an already-mutated part, leaving phantom entries in parts_to_do. #97162 (Alexey Milovidov).
• Fix query plan row count estimation for subqueries with ORDER BY ... LIMIT, which could cause the optimizer to choose a suboptimal join order. #97193 (Alexander Gololobov).
• Fix LOGICAL_ERROR exception in FunctionVariantAdaptor when a function operating on Variant columns returns Nothing type, which can happen with empty arrays in UNION ALL queries. #97213 (Alexey Milovidov).
• Fix a data race during S3 multipart copy operations (e.g., during BACKUP/RESTORE to S3) that could cause exceptions under concurrent access. #97227 (Azat Khuzhin).
• Fix LOGICAL_ERROR exception when arrayJoin in WHERE clause references columns from both sides of a JOIN. #97239 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception when reading .size subcolumn of a sparse Nullable(String) in a Tuple with PREWHERE. #97264 (Alexey Milovidov).
• Fix exception “Number of rows in lazy chunk does not match number of offsets” in LazyMaterializingTransform when reading from tables with non-adaptive index granularity (index_granularity_bytes = 0) using ORDER BY ... LIMIT. #97270 (Alexey Milovidov).
• Fix SYSTEM RESTART REPLICA losing table from database when table re-creation fails with a non-ZooKeeper exception (e.g. memory limit), causing metadata digest mismatches in DatabaseReplicated. #97276 (Alexey Milovidov).
• Field readonly in system.merge_tree_settings now properly reflects that certain merge tree settings (e.g. index_granularity) are unconditionally readonly. #97277 (Robert Schulze).
• Fix a crash during count() optimization on MergeTree tables when the storage snapshot was created without data. #97281 (Pablo Marcos).
• Fix a possible crash when resolving function names from debug information for stack traces. #97294 (Azat Khuzhin).
• Fix logical error with analyzer_compatibility_join_using_top_level_identifier and ALIAS columns. Close #96228. #97297 (Vladimir Cherkasov).
• Fix LOGICAL_ERROR exception in applyOrder when using text-indexed columns with QUALIFY clause. #97313 (Alexey Milovidov).
• System table system.functions now shows for internal functions categories = 'Internal' instead of categories = ''. #97315 (Robert Schulze).
• Query with RIGHT JOIN chain and enabled parallel replicas can produce incorrect result. Fixes #74341. #97316 (Igor Nikonov).
• Fix spurious TABLE_UUID_MISMATCH errors that could occur with refreshable materialized views and other scenarios where tables are renamed. #97323 (Azat Khuzhin).
• Fix segfault in StorageKeeperMap backup due to use-after-free of dangling storage pointer in lazy backup batch. #97336 (Alexey Milovidov).
• Fix the exists function with a scalar subquery inside an ALTER UPDATE/DELETE when mutations_execute_subqueries_on_initiator is enabled. The scalar subquery was incorrectly evaluated, which could lead to an error or a corrupt mutation command that made the table unloadable on the next server restart. #97347 (Kirill Kopnev).
• Fix logical exception Unexpected return type from equals. Expected Nullable(UInt8). Got Const(LowCardinality(Nullable(UInt8))) when comparing NULL with a Variant column containing LowCardinality types. #97379 (Alexey Milovidov).
• Fix a possible race condition when EXCHANGE TABLES is executed in parallel with the sharded query cache enabled. #97411 (Konstantin Vedernikov).
• Fix LOGICAL_ERROR exception in Array-to-QBit conversion when nullable_source from an outer Tuple wrapper replaces the converted array column with a mismatched column type. Closes #97389. #97413 (Alexey Milovidov).
• Fix AST formatting roundtrip inconsistency for aliased tuple literals inside parentheses, e.g., (('a', 'b') AS x) was incorrectly reformatted as tuple(('a', 'b') AS x). #97418 (Alexey Milovidov).
• Fix an exception during asynchronous inserts with deduplication when a parsing failure produced an empty block with zero rows. #97460 (Sema Checherinda).
• Fix exception “Number of rows in lazy chunk does not match number of offsets” in LazyMaterializingTransform when reading from tables with non-adaptive index granularity (index_granularity_bytes = 0) using ORDER BY ... LIMIT. #97482 (Alexey Milovidov).
• Fix insert iceberg settings. Add alias for the allow_experimental_insert_into_iceberg setting. #97483 (Konstantin Vedernikov).
• Fix ACCESS_DENIED for users without CREATE TEMPORARY TABLE permission when optimize_inverse_dictionary_lookup optimization rewrites dictGet(...) predicates. ClickHouse now skips the rewrite and executes the original expression. Closes #97269. #97484 (Nihal Z. Miaji).
• Fix assertion failure (exception in debug/sanitizer builds) in Set and MergeTreeIndexSet when processing columns with inner sparse subcolumns (e.g., Tuple columns from MergeTree parts with different sparse serialization profiles). #97493 (Alexey Milovidov).
• Fix a possible use after free in StorageKafka2. #97520 (Bharat Nallan).
• Fix INTO OUTFILE with TRUNCATE and into_outfile_create_parent_directories setting when the output path contains directories. #97549 (Alexey Milovidov).
• Fix BAD_ARGUMENTS error when querying tables with lambda expressions inside ALIAS columns through the merge() table function with the analyzer enabled. #97551 (Alexey Milovidov).
• Fix system.zookeeper_info exception when Keeper zxid is 0. #97553 (Alexey Milovidov).
• Fix a possible logical error in ip_trie dictionary when key type is not String. #97555 (Bharat Nallan).
• Fix REST catalog OAuth authentication not working for the base RestCatalog (it only worked for derived catalogs like OneLakeCatalog). This broke the default REST catalog after BigLake catalog was introduced. #97561 (Konstantin Vedernikov).
• Geometry functions (perimeterSpherical, areaSpherical, etc.) now accept individual geometry subtypes (Polygon, Ring, Point, etc.) in addition to the Geometry variant type. #97571 (Alexey Milovidov).
• Fix LOGICAL_ERROR exception when using isNull/isNotNull on subcolumns of Nullable(Tuple(... Nullable(T) ...)) types. Closes #97224. #97582 (Alexey Milovidov).
• Fix null pointer dereference when applying patch parts during lightweight updates. #97583 (Alexey Milovidov).
• BaseSettings::readBinary passes the index from accessor.find to field_infos[] without checking for the not-found sentinel value (i.e., -1), which may cause a std::vector out-of-bounds access. The issue was caught thanks to libcxx hardening. This probably happened during query plan deserialization when a newer server sends a setting unknown to an older server. The string-based read method already handles this correctly; readBinary was missing the same check. #97585 (Miсhael Stetsyuk).
• Fix incorrect query results for UNION ALL queries where one branch had a constant-false predicate — the branch would incorrectly read data instead of returning nothing. #97620 (Bharat Nallan).
• Fix IN (col) with a single column reference failing with UNSUPPORTED_METHOD error. #97646 (Alexey Milovidov).
• Fix logical error exception during GROUP BY ... WITH ROLLUP/CUBE when keys include LowCardinality(Nullable(...)) inside Nullable(Tuple(...)). #97647 (Alexey Milovidov).
• Fix AST formatting inconsistency for NOT (1, 1, 1) that could cause LOGICAL_ERROR in debug builds. #97653 (Alexey Milovidov).
• Fix keeper-converter exception when encountering empty ZooKeeper transaction log files. #97673 (Alexey Milovidov).

​

Build/Testing/Packaging Improvement

• ClickHouse can be built with clang-23 (master). #95578 (Alexey Milovidov).
• Fix force is_local to false when bind_host is configured and replace with integration test. Follow-up for #74741. #93109 #96018 (Zhigao Hong).
• Stress tests: fix stress and upgrade tests in CI. ignore no-{build} tags. add compatibility randomization. #94693 (Nikita Fomichev).
• Publish parser_memory_profiler binary from build. Tool can be used to analyze AST memory consumption. #95826 (Ilya Yatsishin).
• Add --symbolize flag for parser_memory_profiler tool that produces .heap.sym files with resolved symbols in results. #96477 (Ilya Yatsishin).
• Pin third-party Docker images in integration tests to specific versions. #96500 (Alexey Milovidov).
• Restore the possibility to link OpenSSL dynamically. This is not recommended and not used by any production builds, but the option still exists for enthusiasts on the Internet. #96506 (Govind R Nair).
• Reduce magic_enum range from [-100, 1000] to default [-128, 127] by using a per-type specialization for Coordination::OpNum, improving build time. #96632 (Alexey Milovidov).
• Remove unnecessary C++ templates from Function classes to reduce build times. #96646 (Alexey Milovidov).
• Move StorageSystemLicenses generation to configure time to improve build parallelism. #96697 (Alexey Milovidov).
• Parallelize license scanning. #96727 (Raúl Marín).
• Add stateless functional test for SSH protocol support. #96996 (Alexey Milovidov).
• Add Kafka 3.9.0 to the stateless functional test infrastructure, enabling direct testing of Kafka and Kafka2 table engines using ClickHouse Keeper as ZooKeeper. Six new stateless tests cover basic produce/consume, virtual columns, INSERT, multiple formats, broken message handling, and Keeper-based offset storage. #96997 (Alexey Milovidov).
• Add a CI workflow to build PGO+BOLT optimized clang toolchain. #96991 (Alexey Milovidov).
• Use the PGO-optimized LLVM/Clang build in CI, which should give 20..30% build speed improvement. #97031 (Alexey Milovidov).
• Replace math functions from glibc with llvm-libc implementations. #90151 (Konstantin Bogdanov).
• Update Boost from 1.83 to 1.90, fixing a devector assertion failure in debug builds. #97037 (Alexey Milovidov).
• Update postgres to REL_18_1. #95189 (Konstantin Bogdanov).
• Use libexpat 2.7.3. #95218 (Konstantin Bogdanov).
• Use OpenSSL 3.5.5. #95345 (Konstantin Bogdanov).
• Use simdjson v4.2.4. #97129 (Konstantin Bogdanov).
• Use libarchive 3.8.5. #97131 (Konstantin Bogdanov).
• Use fast_float v8.2.3. #97133 (Konstantin Bogdanov).
• Use abseil-cpp 20260107.1, s2geometry to v0.13.1. #97134 (Konstantin Bogdanov).
• Bump libxml2 to 2.15.1. #95574 (Robert Schulze).
• Upgraded 7 Tier-3 integration test Docker images from EOL or removed base images to current supported versions. #97314 (Rahul).
• Add TPC-DS benchmark queries. #97349 (Raufs Dunamalijevs).
• Replace individual x86 instruction-set cmake options (ENABLE_SSSE3, ENABLE_AVX2, NO_SSE3_OR_HIGHER, ARCH_NATIVE, etc.) with a single numeric X86_ARCH_LEVEL option (1/2/3/4), matching the standard x86-64 microarchitecture levels already used by the runtime dispatch system. #97354 (Raúl Marín).
• Avoid instantiating division_by_nullable=true template variants for non-division operations in FunctionBinaryArithmetic, reducing compilation time and binary size. #97496 (Raúl Marín).
• Reduce the include footprint of Exception.h by removing it from high-fan-out headers like typeid_cast.h, assert_cast.h, Context_fwd.h, IDataType.h, and various Column headers. #97497 (Raúl Marín).
• Always use bundled compiler-rt headers (sanitizer and XRay interfaces) instead of the host compiler’s headers, and build compiler-rt libraries from source by default. #97499 (Raúl Marín).
• Avoid including boost/multiprecision headers in wide_integer_impl.h on platforms with adequate long double, improving build time. #96633 (Alexey Milovidov).
• Implement LLVM Code Coverage job and enable it initially for master branch. #90952 (Alexey Bakharew).
• Enable fast libcxx hardening for release builds. This is mostly needed for out-of-bounds checks. Given performance tests results, no noticeable performance impact is expected. #94757 (Miсhael Stetsyuk).

​

ClickHouse release 26.1, 2026-01-29. Presentation, Video

​

Backward Incompatible Change

• Fix inconsistent formatting caused by an incorrect substitution of aliases in the formatter. This closes #82833. This closes #82832. This closes #68296. This change is potentially backward incompatible: when the analyzer is disabled, certain CREATE VIEW queries with IN referencing an alias cannot be processed. To prevent the incompatibility, enable the analyzer (it is enabled by default since 24.3). #82838 (Alexey Milovidov).
• Codecs DEFLATE_QPL and ZSTD_QAT were removed. Users are advised to convert existing data compressed with DEFLATE_QPL or ZSTD_QAT to another codec before upgrade. Note that in order to use the codecs, settings enable_deflate_qpl_codec and enable_zstd_qat_codec had to be enabled. #92150 (Robert Schulze).
• Improve UDF debugging by enabling stderr capture in system.query_log.exception. Previously, UDF stderr was only logged to files and not exposed in query logs, making debugging impossible. Now stderr triggers exceptions by default and is fully accumulated (up to 1MB) before throwing, so complete Python tracebacks and error messages appear in system.query_log.exception for effective troubleshooting. #92209 (Xu Jia).
• Empty column list in JOIN USING () clause is now considered a syntax error. Previously it was supposed to be INVALID_JOIN_ON_EXPRESSION during query execution. In some cases such as joining with Join storage it led to LOGICAL_ERROR, close #82502. #92371 (Vladimir Cherkasov).
• Use partial match for SKIP REGEXP in JSON type by default. Closes #79250. #92847 (Pavel Kruglov).
• Revert “Allow INSERT into simple ALIAS columns” (Reverts ClickHouse/ClickHouse#84154). It does not work with custom formats, and is not guarded with a setting. #92849 (Azat Khuzhin).
• Setting to throw an error if a data lake catalog doesn’t have access to object storage. #93606 (Konstantin Vedernikov).
• The Lazy database engine is removed and no longer available. Closes #91231. #93627 (Alexey Milovidov).
• Remove the transposed_with_wide_view mode of the metric_log - it is unusable due to a bug. It is no longer possible to define system.metric_log with this mode. This partially reverts #78412. #93867 (Alexey Milovidov).
• CPU scheduling for workloads is now preemptive by default. See cpu_slot_preemption server setting. #94060 (Sergei Trifonov).
• Escape index filenames to prevent broken parts. With this change ClickHouse will fail to load indices with non-ascii characters in their name created by previous versions. To handle it you can use the merge tree setting escape_index_filenames. #94079 (Raúl Marín).
• Format settings exact_rows_before_limit, rows_before_aggregation, cross_to_inner_join_rewrite, regexp_dict_allow_hyperscan, regexp_dict_flag_case_insensitive, regexp_dict_flag_dotall and dictionary_use_async_executor were changed to be regular (non-format) settings now. This is a purely internal change without user-visible side effects except in the (unlikely) case that you specified any of these settings in Iceberg or DeltaLake or Kafka or S3 or S3Queue or Azure or Hive or RabbitMQ or Set or FileLog or NATS table engine definitions. In these cases, these settings were previously ignored, now such definitions throw an error. #94106 (Robert Schulze).
• The joinGet/joinGetOrNull functions now enforce SELECT privileges on the underlying Join table. After this change, executing joinGet('db.table', 'column', key) requires the user to have SELECT privilege on both the key columns defined in the Join table and the attribute column being retrieved. Queries lacking these privileges will fail with ACCESS_DENIED. To migrate, grant the necessary permissions using GRANT SELECT ON db.join_table TO user for full table access, or GRANT SELECT(key_col, attr_col) ON db.join_table TO user for column-level access. This change affects all users and applications relying on joinGet/joinGetOrNull where explicit SELECT grants were not previously configured. #94307 (Vladimir Cherkasov).
• Check SHOW COLUMNS for CREATE TABLE ... AS ... queries. Previously, it checked SHOW TABLES, which is an incorrect grant for this type of permission check. #94556 (pufit).
• Make the Hash output format independent of block sizes. #94503 (Alexey Milovidov). Note that this changes the output hash values compared to previous versions.

​

New Feature

• HTTP API and embedded Web UI for ClickHouse Keeper. #78181 (pufit and speeedmaster).
• Async insert deduplication now works with dependent materialized views. When collision by block_id occurs, the original block is filtered to remove rows associated with the block_id, and the remaining rows are transformed with all relevant materialized views select queries, this rebuilds original block without conflicting rows. #89140 (Sema Checherinda). It is allowed to use deduplication with async inserts when materialized views are involved. #93957 (Sema Checherinda).
• Introduced a new syntax and framework to simplify and extend projection index feature. This follows up https://github.com/ClickHouse/ClickHouse/pull/81021. #91844 (Amos Bird).
• Add text index support for Array columns. #89895 (Jimmy Aguilar Mena).
• Enable use_variant_as_common_type by default, which lets you use incompatible types inside an Array, in UNION queries, and in branches of if/multiIf/case. #90677 (Alexey Milovidov).
• New system table zookeeper_info. Implements #88014. #90809 (Smita Kulkarni).
• Support the Variant type in all functions. #90900 (Bharat Nallan).
• Adds a ClickHouse_Info metric to the Prometheus /metrics endpoint containing mainly version information so it’s possible to build charts tracking detailed version information over time. #91125 (Christoph Wurm).
• Introduce a new four letter rcfg command for keeper which allows to change cluster configuration. This command provides broader possibilities for configuration changes than standard reconfigure request. Command takes json string as an argument. The whole set of bytes sent to TCP interface should look like this: rcfg{json_string_length_big_endian}{json_string}. Some examples of command may look like this: {"preconditions": {"leaders": [1, 2], "members": [1, 2, 3, 4, 5]}, "actions": [{"transfer_leadership": [3]}, {"remove_members": [1, 2]}, {"set_priority": [{"id": 4, "priority": 100}, {"id": 5, "priority": 100}]}, {"transfer_leadership": [4, 5]}, {"set_priority": [{"id": 3, "priority": 0}]}]}. #91354 (alesapin).
• Add function reverseBySeparator which reverses the order of substrings in a string separated by a specified separator. Close #91463. #91780 (Xuewei Wang).
• Adds new setting max_insert_block_size_bytes which control the formation of inserted blocks in finer detail. #92833 (Kirill Kopnev).
• It is possible to execute DDL queries with ON CLUSTER clause for a Replicated database if the ignore_on_cluster_for_replicated_database setting is enabled. In this case, the cluster name will be ignored. #92872 (Kirill).
• Implement mergeTreeAnalyzeIndexes function. #92954 (Azat Khuzhin).
• Add new setting use_primary_key. Set it to false to disable granule pruning based on the primary key. #93319 (Nihal Z. Miaji).
• Add icebergLocalCluster table function. #93323 (Anton Ivashkin).
• Added cosineDistanceTransposed function that approximates the cosine distance between two points. #93621 (Raufs Dunamalijevs).
• Add files column to system.parts table that shows the number of files in each data part. #94337 (Match).
• Adds a max-min fair scheduler for concurrency control. Provides better fairness under high oversubscription, where many queries compete for limited CPU slots. Short-running queries are not penalized by long-running queries that have accumulated more slots over time. Enabled by the concurrent_threads_scheduler server setting max_min_fair value. #94732 (Sergei Trifonov).
• Added the ability for ClickHouse client to override TLS SNI when connecting to the server. #89761 (Matt Klein).
• Support temporary tables in joinGet function calls. #92973 (Eduard Karacharov).
• Support deletion vectors in DeltaLake table engine. #93852 (Kseniia Sumarokova).
• Support deletion vectors for deltaLakeCluster. #94365 (Kseniia Sumarokova).
• Google cloud storage support for data lakes. #93866 (Konstantin Vedernikov).

​

Experimental Feature

• Move QBit from Experimental to Beta. #93816 (Raufs Dunamalijevs).
• Add support for Nullable(Tuple). Set allow_experimental_nullable_tuple_type = 1 to enable it. #89643 (Nihal Z. Miaji).
• Support Paimon REST catalog, continuing from https://github.com/ClickHouse/ClickHouse/pull/84423. #92011 (JIaQi Tang).

​

Performance Improvement

• Setting use_skip_indexes_on_data_read is now enabled by default. This setting allows filtering in a streaming fashion, at the same time as reading, improving query performance and startup time. #93407 (Shankar Iyer).
• Improve performance of DISTINCT on LowCardinality columns. Closes #5917. #91639 (Nihal Z. Miaji).
• Optimize distinctJSONPaths aggregate function so it reads only JSON paths from data parts and not the whole JSON column. #92196 (Pavel Kruglov).
• More filters pushed down JOINs. #85556 (Nikita Taranov).
• Support more cases for push down from join ON condition when the filter uses inputs only from one side. Support ANY, SEMI, ANTI joins. #92584 (Dmitry Novik).
• Allow using equivalent sets to push down filters for SEMI JOIN. Closes #85239. #92837 (Dmitry Novik).
• Skip reading left side of hash join when right side is empty. Previously we were reading left side until first non-empty block, which might do a lot of work in case when there is heavy filtering or aggregation. #94062 (Alexander Gololobov).
• Using the “fastrange” (Daniel Lemire) method for partitioning data inside the query pipeline. This could improve parallel sorting and JOINs. #93080 (Alexey Milovidov).
• Improve performance of window functions when PARTITION BY matches or is a prefix of the sorting key. #87299 (Nikita Taranov).
• Outer filter is pushed down into views which allows applying PREWHERE on local and remote nodes. Resolves #88189. #88316 (Igor Nikonov).
• Implement JIT compilations for more functions. Closes #73509. #88770 (Alexey Milovidov with Taiyang Li).
• If a skip index used in a FINAL query is on a column that is part of the primary key, the additional step to check for primary key intersection in other parts is unnecessary and now not performed. Resolves #85897. #93899 (Shankar Iyer).
• Optimize performance and memory usage for fractional LIMIT and OFFSET. #91167 (Ahmed Gouda).
• Fix using of faster random read logic for Parquet Reader V3 prefetcher. Closes #90890. #91435 (Arsen Muk).
• Improve performance of icebergCluster. Closes #91462. #91537 (Yang Jiang).
• Don’t filter by virtual columns on constant filters. #91588 (c-end).
• Reduce INSERT/merges memory usage with wide parts for very wide tables by enabling adaptive write buffers. Add support of adaptive write buffers for encrypted disks. #92250 (Azat Khuzhin).
• Improved performance of full text search with text index and sparseGrams tokenizer by reducing the number of searched tokens in the index. #93078 (Anton Popov).
• Function isValidASCII was optimized for positive outcomes, i.e. all-ASCII input values. #93347 (Robert Schulze).
• The read-in-order optimization now recognizes when ORDER BY columns are constant due to WHERE conditions, enabling efficient reverse-order reads. This benefits multi-tenant queries like WHERE tenant='42' ORDER BY tenant, event_time DESC which can now use InReverseOrder instead of requiring a full sort.”. #94103 (matanper).
• Introduce Enum AST specialized class to store value parameters in (string, integer) pairs instead of ASTLiteral children to optimize memory consumption. #94178 (Ilya Yatsishin).
• Distributed index analysis on multiple replicas. Beneficial for shared storage and huge amount of data in cluster. This is applicable for SharedMergeTree (ClickHouse Cloud) and could be applicable for other types of MergeTree tables on a shared storage. #86786 (Azat Khuzhin).
• Reduce overhead of join runtime filters by disabling them in the following cases: - too many bits are set in the bloom filter - too few rows are filtered out at runtime. #91578 (Alexander Gololobov).
• Use an in-memory buffer for correlated subqueries input to avoid evaluating it multiple times. Part of #79890. #91205 (Dmitry Novik).
• Allow all replicas to steal orphaned ranges in parallel replicas reading. This improves load balancing and reduces long-tail latency. #91374 (zoomxi).
• External aggregation/sorting/join now respects query setting temporary_files_codec in all contexts. Fixed missing profile events for grace hash join. #92388 (Vladimir Cherkasov).
• Make query memory usage detection for spilling to disk during aggregation/sorting more robust. #92500 (Azat Khuzhin).
• Estimate total rows count and NDV (number of distinct values) statistics of aggregation key columns. #92812 (Alexander Gololobov).
• Optimize postings list compression with simdcomp. #92871 (Peng Jian).
• Refactor S3Queue Ordered mode processing with buckets. This should also improve performance, reducing the number of keeper requests. #92889 (Kseniia Sumarokova).
• Functions mapContainsKeyLike and mapContainsValueLike can now leverage a text index on mapKeys() or mapValues(), respectively. #93049 (Michael Jarrett).
• Reduce memory usage on non-Linux systems (enable immediate purging of jemalloc dirty pages). #93360 (Eduard Karacharov).
• Force purging of jemalloc arenas in case the ratio of dirty pages size to max_server_memory_usage exceeds memory_worker_purge_dirty_pages_threshold_ratio. #93500 (Eduard Karacharov).
• Reduce memory usage for AST. #93601 (Nikolai Kochetov).
• In some cases we’ve seen ClickHouse doesn’t respect a memory limit when reading from a table. This behaviour is fixed. #93715 (Nikita Mikhaylov).
• Enable CHECK_STAT and TRY_REMOVE Keeper extension by default. #93886 (Mikhail Artemenko).
• Parse lower and upper bounds of file names corresponding to position deletes from Iceberg manifest file entries for better selection of corresponding data files. #93980 (Daniil Ivanik).
• Add two more settings to control maximum number of dynamic subcolumns in JSON column. First is MergeTree setting merge_max_dynamic_subcolumns_in_compact_part (similar to already added merge_max_dynamic_subcolumns_in_wide_part) that limits number of dynamic subcolumns created during merge into a Compact part. Second is query level setting max_dynamic_subcolumns_in_json_type_parsing that limits number of dynamic subcolumns created during parsing of JSON data, it will allow to specify the limit on insert. #94184 (Pavel Kruglov).
• Slightly optimize squashing of JSON columns for some cases. #94247 (Pavel Kruglov).
• Lower the thread pool queue sizes based on the production experience. Add an explicit memory consumption check before reading any data from the MergeTree. #94692 (Nikita Mikhaylov).
• Make sure the scheduler would prefer MemoryWorker thread under the CPU starvation, because it protects ClickHouse process from an existential threat. #94864 (Nikita Mikhaylov).
• Run purging of jemalloc dirty pages in a different thread from main thread of MemoryWorker. If purging is slow, it could delay updates of RSS usage which could lead to out of memory kills of the process. Introduce new config memory_worker_purge_total_memory_threshold_ratio to start purging dirty pages based on ratio of total memory usage. #94902 (Antonio Andelic).

​

Improvement

• system.blob_storage_log is now available for Azure Blob Storage. #93105 (Alexey Milovidov).
• Implement blob_storage_log for Local and HDFS. Fix an error when S3Queue used something other than the disk name for logging in blob_storage_log. Add error_code column to blob_storage_log. Split the test configuration file to simplify local testing. #93106 (Alexey Milovidov).
• clickhouse-client and clickhouse-local will highlight digit groups (thousands, millions, etc.) inside numeric literals while typing. This closes #93100. #93108 (Alexey Milovidov).
• Adds support in clickhouse-client for command-line arguments with a space surrounding the equals sign. Closes #93077. #93174 (Cole Smith).
• With <interactive_history_legacy_keymap>true</interactive_history_legacy_keymap>, the CLI client can now fall back to Ctrl-R for regular search like before, while Ctrl-T does fuzzy search. #87785 (Larry Snizek).
• The statement to clear caches SYSTEM DROP [...] CACHE gave the false impression that the statement disables the cache. ClickHouse now supports statement SYSTEM CLEAR [...] CACHE which is more obvious. The old syntax remains available. #93727 (Pranav Tiwari).
• Support multiple columns as primary key in EmbeddedRocksDB. Closes #32819. #33917 (usurai).
• It is now possible to use non-constant IN for scalars (queries like val1 NOT IN if(cond, val2, val3)). #93495 (Yarik Briukhovetskyi).
• Prevent x-amz-server-side-encryption headers from being propagated to HeadObject, UploadPart & CompleteMultipartUpload S3 requests as they’re not supported. #64577 (Francisco J. Jurado Moreno).
• Tracking hive partitioning for ordered mode in S3Queue. Resolves #71161. #81040 (Anton Ivashkin).
• Optimize space reservation in filesystem cache. FileCache::collectCandidatesForEviction will be executed without unique lock. #82764 (Kseniia Sumarokova).
• Support composite rotation strategy (size + time) for server log. #87620 (Jianmei Zhang).
• CLI client can now specify <warnings>false</warnings> instead of the command line --no-warnings. #87783 (Larry Snizek).
• Add support for the avg aggregate function with Date, DateTime and Time values as arguments. Closes #82267. #87845 (Yarik Briukhovetskyi).
• The optimization use_join_disjunctions_push_down is enabled by default. #89313 (Alexey Milovidov).
• Support more table engines and data source kinds in the correlated subqueries. Closes #80775. #90175 (Dmitry Novik).
• If the schema of parameterized view is specified explicitly, it is shown. Close #88875, #81385. #90220 (Grigorii Sokolik).
• Correctly handle the gap in Keeper log entries if logs are before the last committed index. #90403 (Antonio Andelic).
• Improve min_free_disk_bytes_to_perform_insert setting to work correctly with JBOD volumes. #90878 (Aleksandr Musorin).
• Make it possible to specify storage_class_name setting in named collections for S3 table engine and s3 table function. #91926 (János Benjamin Antal).
• Support inserting auxiliary zookeeper by system.zookeeper. #92092 (RinChanNOW).
• Add new metrics for the keeper: KeeperChangelogWrittenBytes, KeeperChangelogFileSyncMicroseconds, KeeperSnapshotWrittenBytes and KeeperSnapshotFileSyncMicroseconds profile events as well as KeeperBatchSizeElements and KeeperBatchSizeBytes histogram metrics. #92149 (Miсhael Stetsyuk).
• Add a new setting, trace_profile_events_list, which limits tracing with trace_profile_event to the specified list of event names. This allows more precise data collection on large workloads. #92298 (Alexey Milovidov).
• Support SYSTEM NOTIFY FAILPOINT for pausable failpoints. - Support SYSTEM WAIT FAILPOINT fp PAUSE/RESUME. #92368 (Shaohua Wang).
• Add creation (implicit/explicit) column to system.data_skipping_indices. #92378 (Raúl Marín).
• Allow passing the description of columns for YTsaurus dyn tables to the dictionary source. #92391 (MikhailBurdukov).
• In #63985, we made it possible to specify all the parameters needed for TLS configuration on a per-port basis (see composable protocols), so we don’t have to rely on global TLS config. However, the implementation still implicitly requires a global openSSL.server config section to exist, which conflicts with setups where different TLS configurations are needed for different ports. For example, in keeper-in-server deployments, we need separate TLS configs for inter-keeper communication and clickhouse client connections. #92457 (Miсhael Stetsyuk).
• Introduce a new setting input_format_binary_max_type_complexity that limits the total number of type nodes that can be decoded in binary format to prevent malicious payloads. #92519 (Raufs Dunamalijevs).
• Reflect running tasks in system.background_schedule_pool{,_log}. Add documentation. #92587 (Azat Khuzhin).
• Execute current query in Ctrl+R search in client if no history match found. #92749 (Azat Khuzhin).
• Support EXPLAIN indices = 1 as an alias for EXPLAIN indexes = 1. Closes #92483. #92774 (Pranav Tiwari).
• Parquet reader now allows reading Tuple or Map columns as JSON: select x from file(f.parquet, auto, 'x JSON') works even if the type of column x in f.parquet is tuple or map. #92864 (Michael Kolupaev).
• Support empty tuples in parquet reader. #92868 (Michael Kolupaev).
• Fallback to read-write copy for Azure Blob Storage when native copy fails with BadRequest (e.g. invalid block list). Previously this was only done for Unauthorized error which was seen while copying blob to different storage accounts. But we also sometimes see “The specified block list is invalid” error. So now updated the condition to fallback to read & write for all native copy fails. #92888 (Smita Kulkarni).
• Fix EC2 metadata endpoint throttling when running many concurrent S3 queries with EC2 instance profile credentials. Previously, each query created its own AWSInstanceProfileCredentialsProvider, causing concurrent requests to the EC2 metadata service which could result in timeouts and HTTP response code: 403 errors. Now the credentials provider is cached and shared across all queries. #92891 (Sav).
• Rework insert_select_deduplicate setting to add an ability to keep backward compatibility. #92951 (Sema Checherinda).
• Log background tasks that are slower than average (background_schedule_pool_log.duration_threshold_milliseconds=30) to avoid excessive tasks logging. #92965 (Azat Khuzhin).
• In previous versions, some of C++ function names were displayed incorrectly (“mangled”) in the system.trace_log and system.symbols, and the demangle function didn’t process them well. Closes #93074. #93075 (Alexey Milovidov).
• Introduced the backup_data_from_refreshable_materialized_view_targets backup setting to skip back up of refreshable materialized views. RMVs with APPEND refresh strategy which are always backed up. #93076 (Julia Kartseva). #93658 (Julia Kartseva)
• Use minimal debug info instead of no debug info for heavy translation units, such as functions. #93079 (Alexey Milovidov).
• Added MinIO compatibility support to AWS S3 C++ SDK by implementing error code mapping for MinIO-specific errors. This change allows ClickHouse to properly handle and retry MinIO server errors when using MinIO deployments instead of AWS S3, improving reliability for users running object storage on self-hosted MinIO clusters. #93082 (XiaoBinMu).
• Write symbolized jemalloc profiles (eliminating the need for a binary during heap profile generation). #93099 (Azat Khuzhin).
• Resurrect clickhouse git-import tool - it was broken on large and invalid commits. See https://presentations.clickhouse.com/2020-matemarketing/. #93202 (Alexey Milovidov).
• Don’t show passwords from URL storage in query log. #93245 (Konstantin Vedernikov).
• Support Geometry type for flipCoordinates. #93303 (Bharat Nallan).
• Improve the UX of SYSTEM INSTRUMENT ADD/REMOVE: use String literals for function names, patch all functions that match and allow using function_name in REMOVE. #93345 (Pablo Marcos).
• Add a new setting materialize_statistics_on_merge which enables/disables materializing statistics during merge. The default value is 1. #93379 (Han Fei).
• ClickHouse can now parse SELECT without parentheses around DESCRIBE SELECT queries. Closes #58382. #93429 (Yarik Briukhovetskyi).
• Add randomization of cache correctness checks under probability. #93439 (Kseniia Sumarokova).
• Add setting type_json_allow_duplicated_key_with_literal_and_nested_object to allow duplicated paths in JSON where one is a literal and another is a nested object, e.g. {"a" : 42, "a" : {"b" : 42}}. Some data could be created before this restriction on duplicated paths was added in https://github.com/ClickHouse/ClickHouse/pull/79317 and further manipulation with this data can lead to errors now. With this setting, such old data cane still be used with no errors. #93604 (Pavel Kruglov).
• Don’t print values of simple types on separate lines in Pretty JSON. #93836 (Pavel Kruglov).
• When there are many alter table ... modify setting ... statements, it’s possible not to acquire lock for 5 seconds. Better to return timeout than logical error. #93856 (Han Fei).
• Prevent excessive output on a syntax error. Before this change, it output the whole SQL script, which could contain a lot of queries. #93876 (Alexey Milovidov).
• Do proper byte size calculation of the check request with stats in Keeper. #93907 (Mikhail Artemenko).
• Added use_hash_table_stats_for_join_reordering setting to control whether runtime hash table size statistics are used for join reordering. This setting is enabled by default, preserving the existing behavior of collect_hash_table_stats_during_joins. #93912 (Vladimir Cherkasov).
• Users can now partially view nested global server settings in the system.server_settings table (e.g. logger.level). This only covers settings with a fixed structure (no lists, enumerations, repetitions etc.). #94001 (Hechem Selmi).
• QBit can now be compared for equality. #94078 (Raufs Dunamalijevs).
• When Keeper detects broken snapshot or inconsistent changelogs, throw exception instead of manually aborting or cleaning up files automatically. This should lead to a safer behaviour of Keeper relying on manual intervention. #94168 (Antonio Andelic).
• Fix leaving possible leftovers in case of CREATE TABLE fails. #94174 (Azat Khuzhin).
• Fix uninitialized memory access (a bug in OpenSSL) when password protected TLS key is used. #94182 (Konstantin Bogdanov).
• Bump chdig to v26.1.1. #94290 (Azat Khuzhin).
• Support more generic partitioning for S3Queue ordered mode. #94321 (Bharat Nallan).
• Added alias use_statistics for setting allow_statistics_optimize. This is more consistent with existing settings use_primary_key and use_skip_indexes. #94366 (Robert Schulze).
• Enabled setting input_format_numbers_enum_on_conversion_error for conversion from Numbers to Enums to check whether the element exists. #94384 (Elmi Ahmadov).
• In S3(Azure)Queue ordered mode clean up failed nodes by tracking limits (before that was done only in Unordered mode for both failed and processed, so now this will also be done for Ordered but only for failed nodes). #94412 (Kseniia Sumarokova).
• Enable access management for default user in clickhouse-local. The default user in clickhouse-local was missing the access_management privilege, which caused operations like DROP ROW POLICY IF EXISTS to fail with ACCESS_DENIED error, even though the user should be unrestricted. #94501 (Alexey Milovidov).
• Enable named collection for YTsaurus dictionaries and tables. #94582 (MikhailBurdukov).
• Add support for SQL-defined named collections in BACKUP/RESTORE for S3 and Azure Blob Storage. Closes #94604. #94605 (Pablo Marcos).
• Support bucketing based on partition key for S3Queue in ordered mode. #94698 (Bharat Nallan).
• Add an asynchronous metric with the longest running merge elapsed time. #94825 (Raúl Marín).
• Add belonging file check before apply position delete using IcebergBitmapPositionDeleteTransform. #94897 (Yang Jiang).
• Now view_duration_ms shows the time when group was active, not the sum of the threads duration in it. #94966 (Sema Checherinda).
• Remove limit of the max number of search tokens in hasAnyTokens and hasAllTokens functions which was limited to 64. Example: SELECT count() FROM table WHERE hasAllTokens(text, ['token_1', 'token_2', [...], 'token_65']]); The query would result in a BAD_ARGUMENTS error because there are 65 search tokens. With this PR, the limit has been removed completely and the same query would run without an error. #95152 (Elmi Ahmadov).
• Add a setting input_format_numbers_enum_on_conversion_error for conversion from Numbers to Enums to check whether the element exists. Closes: #56144. #56240 (Nikolay Degterinsky).
• Share format parser resources between data file and position delete file reading in Iceberg tables to reduce memory allocations. #94701 (Yang Jiang).

​

Bug Fix (user-visible misbehavior in an official stable release)

• Fixes a bug where predefined query handlers would have trailing whitespace interpreted as data during inserts. #83604 (Fabian Ponce).
• Fix INCOMPATIBLE_TYPE_OF_JOIN error for Join storage and outer to inner join optimization applied. Resolves #80794. #84292 (Vladimir Cherkasov).
• Fix exception “Invalid number of rows in Chunk” when using hash join with allow_experimental_join_right_table_sorting enabled. #86440 (yanglongwei).
• Always replace file names to hash in MergeTree if filesystem is case insensitive. Previously on systems with case insensitive filesystem (like MacOS) it could lead to data corruption when several column/subcolumn names differs only in the case. #86559 (Pavel Kruglov).
• Add a full permissions check on the create stage for the underlying query inside a materialized view. #89180 (pufit).
• Fixed crash in icebergHash function on constant argument. #90335 (Michael Kolupaev).
• Fix logical error when mutation without transaction mutates parts in an active transaction, which is rolled back finally. #90469 (Shaohua Wang).
• Update system.warnings correctly after an ordinary database was converted to an atomic database. #90473 (sdk2).
• Fixes an assertion when reading from Parquet file, and part of a prewhere expression is used elsewhere in the query. #90635 (Max Kainov).
• Fix crash in a single-node cluster when reading from Iceberg in split-by-buckets mode. This closes #90913. #91553 (Konstantin Vedernikov).
• Fix possible logical error in Log engine during subcolumns reading. Closes #91710. #91711 (Pavel Kruglov).
• Fix Logical error: ‘Storage does not support transaction’ during ATTACH AS REPLICATED. #91772 (Shaohua Wang).
• Fix for runtime filters working incorrectly when LEFT ANTI JOIN has extra post-condition. #91824 (Alexander Gololobov).
• Fixes an error where we have a null-safe comparison involving the Nothing type. Closes #91834. Closes #84870. Closes #91821. #91884 (Yarik Briukhovetskyi).
• Fix DELTA_BYTE_ARRAY decoding bugs in native Parquet reader affecting highly repetitive string data. #91929 (Daniel Muino).
• Cache schema only for the file it was inferred from in globs instead of all files during schema inference. Closes #91745. #92006 (Pavel Kruglov).
• Fix the Couldn't pack tar archive: Failed to write all bytes error caused by an incorrect archive entry size header. Fixes #89075. #92122 (Julia Kartseva).
• Release request stream in insert select to prevent closing http connection. #92175 (Sema Checherinda).
• Fix logical error for queries with multiple JOINs with USING clause and join_use_nulls. #92251 (Vladimir Cherkasov).
• Fix logical error while join reordering with join_use_nulls, close https://github.com/clickhouse/clickhouse/issues/90795. #92289 (Vladimir Cherkasov).
• Fix inconsistent AST formatting of arrayElement with negate literal. Closes #92288 Closes #92212 Closes #91832 Closes #91789 Closes #91735 Closes #88495 Closes #92386. #92293 (Pavel Kruglov).
• Fix a possible crash with join_on_disk_max_files_to_merge setting. #92335 (Bharat Nallan).
• Related issue #https://github.com/ClickHouse/support-escalation/issues/6365. #92339 (Tuan Pham Anh).
• Fix missing access check in SYSTEM SYNC FILE CACHE. Closes #92101. #92372 (Kseniia Sumarokova).
• Fix count_distinct_optimization pass over window functions and over multiple arguments. #92376 (Raúl Marín).
• Fix “Cannot write to finalized buffer” error when using certain aggregate functions with window functions. Closes #91415. #92395 (Jimmy Aguilar Mena).
• Fix logical error with CREATE TABLE ... AS urlCluster() and database engine Replicated. Closes #92216. #92418 (Kseniia Sumarokova).
• Inherit source part serialization info settings during mutation in MergeTree. It fixes possible incorrect result of the query over mutated part after changes in data types serialization. #92419 (Pavel Kruglov).
• Fix possible conflict in column and subcolumn with the same name leading in using wrong serialization and query failures. Closes #90219. Closes #85161. #92453 (Pavel Kruglov).
• Fix a LOGICAL_ERRORs that caused by not wanted modification of query plan when converting outer join to inner join. Also relax the requirements of optimization to be able to apply it in cases when injective functions are applied to the aggregating keys during joins. #92503 (János Benjamin Antal).
• Fix possible error SIZES_OF_COLUMNS_DOESNT_MATCH during sorting of emty tuple column. Closes #92422. #92520 (Pavel Kruglov).
• Check for incompatible typed paths in JSON type. Closes #91577. #92539 (Pavel Kruglov).
• Fix deadlock for SHOW CREATE DATABASE for Backup database. #92541 (Azat Khuzhin).
• Use proper error code when validating hypothesis index. #92559 (Raúl Marín).
• Fix dynamic subcolumns resolution in column aliases in analyzer. Previously dynamic subcolumn in column alias was wrapped in getSubcolumn and in some cases could be not resolved at all. Closes #91434. #92583 (Pavel Kruglov).
• Prevent crash in tokens() with null second argument. #92586 (Raúl Marín).
• Fix potential crash caused by in place mutation of underlying const PREWHERE columns. This could’ve happened at column shrinking (IColumn::shrinkToFit) or filtering (IColumn::filter), which could’ve triggered concurrently from several threads. #92588 (Arsen Muk).
• Creating and materializing text indexes on tables containing large parts (over 4,294,967,295 rows) is temporarily disabled. This limitation prevents incorrect query results, as the current index implementation does not yet support such large parts. #92644 (Anton Popov).
• Fixes a logical error Too large size (A) passed to allocator while executing JOINs. Closes #92043. #92667 (Yarik Briukhovetskyi).
• Remove a bug that ngrambf_v1 indexes with ngram length (1st parameter) > 8 would throw an exception. #92672 (Robert Schulze).
• Fix uncaught exception during background named collections reload when zookeeper storage is used. Closes https://github.com/ClickHouse/clickhouse-private/issues/44180. #92717 (Kseniia Sumarokova).
• Reworks incorrect logic in access grant checks for wildcard grants. The previous attempt https://github.com/ClickHouse/ClickHouse/pull/90928 addressed a critical vulnerability but ended up being too restrictive, resulting in some wildcard GRANT statements failing due to unrelated revokes. #92725 (pufit).
• Fix bug in data skipping logic when not match(...) is used in WHERE causing incorrect results. Closes #92492. #92726 (Nihal Z. Miaji).
• Do not attempt to delete temporary directories at startup if a MergeTree table is created over a read-only disk. #92748 (Alexey Milovidov).
• Fix “Cannot add action to empty ExpressionActionsChain” for ALTER TABLE REWRITE PARTS (v2). #92754 (Azat Khuzhin).
• Avoid crash due to reading from a disconnected Connection. #92807 (Raufs Dunamalijevs).
• Fix logical error Failed to set file processing within 100 retries in storgae S3Queue in Ordered mode. It is now replaced with a warning. This error could happen before 25.10 version if keeper session expired, however it will still be a warning in 25.10+ versions, as it is still theoretically possible to get this error in case of high processing concurrency in Ordered mode. #92814 (Kseniia Sumarokova).
• Previously, some queries that used PK sharding with a false condition were failing. Now they’re not. Needed for https://github.com/ClickHouse/ClickHouse/pull/89313. #92815 (Yarik Briukhovetskyi).
• Fixed calculation of uncompressed sizes of text indexes in the system.parts table. #92832 (Anton Popov).
• Fixed usage of primary index in lightweight updates that have an IN clause with subqueries in the predicate of the WHERE clause. #92838 (Anton Popov).
• Fix creating type hint for path ‘skip’ in JSON. Closes #92731. #92842 (Pavel Kruglov).
• In the S3 table engine, we should avoid caching the partition key if there are non-deterministic functions. #92844 (Miсhael Stetsyuk).
• Fix possible error FILE_DOESNT_EXIST after mutation of a sparse column with ratio_of_defaults_for_sparse_serialization=0.0. Closes #92633. #92860 (Pavel Kruglov).
• Fix parquet schema inference in the old parquet reader (not used by default) when a JSON column comes after a Tupe column. Fix the old parquet reader (not used by default) failing on empty tuples. #92867 (Michael Kolupaev).
• Fix logical error with multiple joins on constant condition and join_use_nulls, close #92640. #92892 (Vladimir Cherkasov).
• Fix possible error NOT_FOUND_COLUMN_IN_BLOCK during insert into a table with subcolumn in partition expression. Closes #93210. Closes #83406. #92905 (Pavel Kruglov).
• Fix error NO_SUCH_COLUMN_IN_TABLE in Merge engine over tables with aliases. Closes #88665. #92910 (Pavel Kruglov).
• Fix NULL != NULL case for full_sorting_join on LowCardinality(Nullable(T)) column. #92924 (Vladimir Cherkasov).
• Fixed several crashes during merges of text indexes in MergeTree tables. #92925 (Anton Popov).
• Restore LowCardinality wrappers on SET expression results if needed during TTL aggregation to prevent exceptions during table optimization. #92971 (Seva Potapov).
• Fix logical error during index analysis when empty array is used in has function. Closes #92906. #92995 (Nihal Z. Miaji).
• Fix possible hung on terminating background schedule pool (may lead to server hungs on shutdown). #93008 (Azat Khuzhin).
• Fix possible error FILE_DOESNT_EXIST after sparse column mutation when setting ratio_of_defaults_for_sparse_serialization was changed to 1.0 via alter. #93016 (Pavel Kruglov).
• Fix bug in data skipping logic when not materialize(...) or not CAST(...) is used in WHERE causing incorrect results. Closes #88536. #93017 (Nihal Z. Miaji).
• Fix possible usage of outdated parts due to TOCTOU race for shared parts. #93022 (Azat Khuzhin).
• Fix crash when deserialising malformed groupConcat aggregate state with out-of-bounds offsets. #93028 (Raufs Dunamalijevs).
• Fix leaving connection in a broken state after preliminary cancellation distributed queries. #93029 (Azat Khuzhin).
• Fix join results when the right-side join key is a sparse column. This closes #92920. I can only reproduce the bug with set compatibility='23.3'. Not sure if it should be backported. #93038 (Amos Bird).
• Fix possible Cannot finalize buffer after cancellation in estimateCompressionRatio(). Fixes: #87380. #93068 (Azat Khuzhin).
• Fixed merges of text indexes built on top of the complex expressions (such as concat(col1, col2)). #93073 (Anton Popov).
• Fix applying projection when filter contains subcolumns. Closes #92882. #93141 (Pavel Kruglov).
• Fix logical error in some cases triggered when join runtime filters are added to query plan. It was caused by incorrectly returning duplicated const columns from one of join sides. #93144 (Alexander Gololobov).
• Special function __applyFilter used by join runtime filters was returning ILLEGAL_TYPE_OF_ARGUMENT in some valid cases. #93187 (Alexander Gololobov).
• Prevent different interpolated columns from collapse into the same column in a block when interpolated columns are effectively aliases of the same column. #93197 (Yakov Olkhovskiy).
• Do not add runtime filter when joining with already filled right table. #93211 (Alexander Gololobov).
• Fix keeper persistent watches cleanup after dead session. This closes #92480. #93213 (Konstantin Vedernikov).
• Fix order by tuple in iceberg. This closes #92977. #93225 (Konstantin Vedernikov).
• Fix bug with S3Queue setting s3queue_migrate_old_metadata_to_buckets. Closes #93392, #93196, #81739. #93232 (Kseniia Sumarokova).
• Remove unused columns when the projection is rebuilt during the merge. It reduces memory usage and creates fewer temporary parts. #93233 (Nikolai Kochetov).
• Fix unused columns removal from subqueries in the presence of a scalar correlated subquery. Before the fix column could have been removed if it was used only in the correlated subquery, and the query would fail with NOT_FOUND_COLUMN_IN_BLOCK error. #93273 (Dmitry Novik).
• Fix possible missing subcolumn in MV during alter of source table. Closes #93231. #93276 (Pavel Kruglov).
• Fix the Merge table engine query planning with the analyzer that could throw ILLEGAL_COLUMN for hostName() when merging local and remote/Distributed tables. Closes #92059. #93286 (Jinlin).
• Fixes a case where NOT IN with non-constant array arguments was returning the wrong value + Support for non-constant Array functions. Closes #14980. #93314 (Yarik Briukhovetskyi).
• Fix Not found column for use_top_k_dynamic_filtering optimization. Fixes #93186. #93316 (Nikolai Kochetov).
• Fixed rebuilding of text indexes created on top of subcolumns. #93326 (Anton Popov).
• Fixed handling of empty array as a second argument in hasAllTokens and hasAnyTokens functions. #93328 (Anton Popov).
• Fix logical error when runtime filters are used in a query with totals for right side table. #93330 (Alexander Gololobov).
• The server no longer crashes if function tokens is called with non-const tokenizer parameters (the 2th, 3rd, 4th parameter), e.g., SELECT tokens(NULL, 1, materialize(1)). #93383 (Robert Schulze).
• Fixed integer overflow vulnerability in groupConcat state deserialisation that could cause memory safety issues with crafted aggregate states. #93426 (Raufs Dunamalijevs).
• Fixed text index analysis on array columns when the index contains no tokens (all arrays are empty or all tokens are skipped by the tokenizer). #93457 (Anton Popov).
• Avoids oauth login in ClickHouse Client when username/password are within the connection string. #93459 (Krishna Mannem).
• Fix Azure ADLS Gen2 vended credentials support in DataLakeCatalog - parse adls.sas-token.* keys from Iceberg REST catalogs and fix ABFSS URL parsing. #93477 (Karun Anantharaman).
• Fix GLOBAL IN support with analyzer (previously set was created on the remote node again). #93507 (Azat Khuzhin).
• Fix extracting subcolumn during deserialization directly into Sparse columns. #93512 (Pavel Kruglov).
• Fixed direct reading from text index with duplicate search queries. #93516 (Anton Popov).
• Fix for NOT_FOUND_COLUMN_IN_BLOCK error when runtime filter is enabled and joined tables have the same column returned multiple times (e.g. SELECT a, a, a FROM t). #93526 (Alexander Gololobov).
• Fix a bug where clickhouse-client would ask for password twice when connecting using ssh. #93547 (Isak Ellmer).
• Make sure that zookeeper is finalized on shutdown (fix possible hung on shutdown in very unlikely cases). #93602 (Azat Khuzhin).
• Fix LOGICAL_ERROR when restoring ReplicatedMergeTree with deduplication race. #93612 (Pablo Marcos).
• Fix using Sparse column for TTL update during direct deserialization into Sparse columns in some input formats. It fixes possible logical error Unexpected type of result TTL column. #93619 (Pavel Kruglov).
• Fixed h3 index functions sometimes crashing or getting stuck when called on invalid inputs. #93657 (Michael Kolupaev).
• The usage of ngram_bf index on a non-UTF-8 data led to an uninitialized memory read, with values that could reside in the resulting index structure. Closes #92576. #93663 (Alexey Milovidov).
• Validate that the decompressed buffer size is as expected. #93690 (Raúl Marín).
• Prevent users to get the list of columns from a table without checking SHOW COLUMNS permission using the merge table engine. #93695 (János Benjamin Antal).
• Fixed materialization of skip indexes created on top of subcolumns. #93708 (Anton Popov).
• We store storages’ shared pointers in QueryPipeline::resources::storage_holders to make sure that the IStorage objects are not destroyed while PipelineExecutor is alive. #93746 (Miсhael Stetsyuk).
• Fix attaching Replicated DBs when the interserver host changed after restarting. #93779 (Tuan Pham Anh).
• Fix assert !read_until_position in ReadBufferFromS3 which happened when cache is enabled. #93809 (Kseniia Sumarokova).
• Fix logical error in a rare case when empty tuple is used with Map column. Closes #93784. #93814 (Nihal Z. Miaji).
• Fixed _part_offset corruption when projections are rebuilt during merges, and optimized projection processing by avoiding unnecessary reads of the _part_offset column and skipping unneeded columns in projection calculations. This continues the optimizations introduced in #93233. #93827 (Amos Bird).
• Remove ‘Bad version’ handling. #93843 (Anton Ivashkin).
• Fix optimize_inverse_dictionary_lookup not working with distributed query when key is signed integral type. Closes #93259. #93848 (Nihal Z. Miaji).
• Fix lag/lead not working with distributed remote() query. Closes #90014. #93858 (Nihal Z. Miaji).
• Fix system instrument dispatch bug. #93937 (Pablo Marcos).
• In https://github.com/ClickHouse/ClickHouse/pull/89173, we added an extra field to the structure that TraceSender sends through an internal pipe. However, the buffer size was not updated (here), therefore we are writing more data to buffer than buffer_size which results in multiple flushes. And because TraceSender::send is called from different threads, different threads’ flushes may interleave which breaks the invariant that the receiving end (TraceCollector) relies on. #93966 (Miсhael Stetsyuk).
• Fix type conversion to super type during the join operation of the storage Join with USING clause. Fixes #91672. Fixes #78572. #94000 (Dmitry Novik).
• Fix for FilterStep not properly added when join runtime filter is applied over Merge table. #94021 (Alexander Gololobov).
• A SELECT query containing a predicate on multiple columns with bloom filter skip indexes and both OR and NOT conditions are present could return inconsistent results. That is fixed now. #94026 (Shankar Iyer).
• Fix CLEAR column with dependent indices. #94057 (Raúl Marín).
• Fix use-of-uninitialized-value in ReadWriteBufferFromHTTP. #94058 (Alexey Milovidov).
• Fix bad check for typed paths in JSON. The check was introduced in https://github.com/ClickHouse/ClickHouse/pull/92842 and can lead to an error during existing tables startup. #94070 (Pavel Kruglov).
• Fix the crash during filter analysis in the presence of OUTER JOIN. Fixes #90979. #94080 (Dmitry Novik).
• Fix accuracy of uniqTheta when using UInt8 aggregation keys in parallel (max_threads > 1 - default). #94095 (Azat Khuzhin).
• Fix crash caused by exception thrown from a socket.setBlocking(true) call inside SCOPE_EXIT. #94100 (Miсhael Stetsyuk).
• Fix data loss when DROP PARTITION removes parts created by later log entries in ReplicatedMergeTree. #94123 (Tuan Pham Anh).
• Fixed parquet reader v3 incorrectly handling arrays that cross page boundaries. This happens e.g. for files written by Arrow without enabling page statistics or page index. Affects only columns of Array data type. Likely symptom is that one array every ~1 MB of data gets truncated. Before this fix, use this setting as workaround: input_format_parquet_use_native_reader_v3 = 0. #94125 (Michael Kolupaev).
• Fix too many watches in ReplicatedMergeTree while waiting for log entry. #94133 (Azat Khuzhin).
• Functions arrayShuffle, arrayPartialShuffle and arrayRandomSample to materialize const columns - so that different rows get different results. #94134 (Joanna Hulboj).
• Fix data race in evaluating table functions in materialized views. #94171 (Alexey Milovidov).
• Fix nullptr dereference in PostgreSQL database engines (when the query is incorrect). Closes #92887. #94180 (Alexey Milovidov).
• Fix memory leak in refreshable materialized views using SELECT queries with multiple subqueries. #94200 (Antonio Andelic).
• Fix data race in DataPartStorageOnDiskBase::remove vs system.parts. Closes #49076. #94262 (Alexey Milovidov).
• Remove the wrong noexcept specifier at HashTable copy assignment that may lead to crash (std::terminate) on memory exceptions. #94275 (Nikita Taranov).
• Previously, creating a projection with duplicate columns in GROUP BY (e.g., GROUP BY c0, c0) and inserting data caused a std::length_error if optimize_row_order is enabled. Closes #94065. #94277 (Alexey Milovidov).
• Fix obscure bug in ZooKeeper client on connect which leads to hungs and crashes. #94320 (Azat Khuzhin).
• Fix function to subcolumns optimization not applied to subcolumns. #94323 (Pavel Kruglov).
• Fix possibly incorrect result in nested RIGHT JOINs when enable_lazy_columns_replication is enabled. The bug caused all rows in replicated columns to incorrectly return the same value instead of their distinct values. Close #93891. #94339 (Vladimir Cherkasov).
• Fix filter pushdown for SEMI JOIN using equivalence sets. Do not push the filter down if argument types have changed. Fixes #93264. #94340 (Dmitry Novik).
• Fix usage of DeltaLake CDF with database DataLake database engine (delta lake catalogs integration). Closes #94122. #94342 (Kseniia Sumarokova).
• Fix incorrect value of current metric FilesystemCacheSizeLimit in case SLRU cache policy was used. #94363 (Kseniia Sumarokova).
• Creating a Backup database engine with less than two arguments now returns a more descriptive error message (Wrong number of arguments instead of std::out_of_range: InlinedVector::at(size_type) const failed bounds check.). #94374 (Robert Schulze).
• Ignores impossible revokes of global grants on the database level for grants with grant option. #94386 (pufit).
• Fix reading sparse offsets from compact parts. Closes #94385. #94399 (Pavel Kruglov).
• Don’t prevent ALTER of columns using implicit indexes, even if alter_column_secondary_index_mode’s throw mode is used. #94425 (Raúl Marín).
• Fix crash in TCPHandler when multiple receivePacketsExpectQuery calls read Protocol::Client::IgnoredPartUUIDs. #94434 (Miсhael Stetsyuk).
• Fix masking sensitive data in system.functions. #94436 (Vitaly Baranov).
• Fix nullptr dereference with disabled send_profile_events. This feature was introduced recently for the ClickHouse Python driver. Closes #92488. #94466 (Alexey Milovidov).
• Fix text index .mrk incompatibility during merges. #94494 (Peng Jian).
• When read_in_order_use_virtual_row is enabled, the code was accessing index columns based on the full primary key size without checking if the index was truncated, leading to use-after-free / uninitialized memory. Closes #85596. #94500 (Alexey Milovidov).
• Fix an error due to a type mismatch when sending external tables for subqueries with GLOBAL IN if the types are Nullable. Closes #94097. #94511 (Alexey Milovidov).
• In previous versions, queries with multiple index conditions over the same expression may erroneously throw an exception Not found column. Closes #60660. #94515 (Alexey Milovidov).
• Fix incorrect handling of Nullable join column in runtime filters. #94555 (Alexander Gololobov).
• Creating a workload in another workload that is currently in use no longer causes a crash. #94599 (Sergei Trifonov).
• Fix a crash during ANY LEFT JOIN optimization when isNotNull is evaluated on a missing column. #94600 (Molly).
• Fix default expression evaluation when referencing other columns with computed defaults. #94615 (Alexey Milovidov).
• Fix permission issues in BACKUP/RESTORE operations. #94617 (Pablo Marcos).
• Fix crash due to incorrect type cast when the data type is Nullable(DateTime64). #94627 (Miсhael Stetsyuk).
• Fixes a bug where certain distributed queries with ORDER BY could return ALIAS columns with swapped values (i.e., column a showing column b’s data and vice versa). #94644 (filimonov).
• Fix storing results of keeper-bench to file. #94654 (Antonio Andelic).
• Fix incorrect estimations with MinMax-type statistics when the column contains negative floating-point values. #94665 (zoomxi).
• Fix reading Parquet files when a map’s key is a struct. #94670 (Konstantin Vedernikov).
• Fix possibly incorrect RIGHT join result when using complex ON conditions. Close #92913. #94680 (Vladimir Cherkasov).
• Preserve constant index granularity (use_const_adaptive_granularity) after Vertical merges. #94725 (Azat Khuzhin).
• Fix mutation bug with scalar subqueries and table dependencies. If a table had dependencies (index or projections) over a column, scalar subqueries might be evaluated and cached without data and lead to incorrect changes. #94731 (Raúl Marín).
• Fix AsynchronousMetrics cpu_pressure fallback on error. #94827 (Raúl Marín).
• The getURLHostRFC function was missing bounds checks before dereferencing pointers. When an empty string was passed to domainRFC, it would read uninitialized memory, triggering MSan errors. #94851 (Alexey Milovidov).
• Fix readonlyness of encrypted disks. #94852 (Azat Khuzhin).
• Fix logical error in fractional LIMIT/OFFSET when using the old analyzer with Distributed tables. Closes #94712. #94999 (Ahmed Gouda).
• Fix crash under some conditions when join runtime filters are enabled by default. #95000 (Alexander Gololobov).
• Improve masking passwords in url used in table engine URL() and table function url(). #95006 (Vitaly Baranov).
• Function toStartOfInterval now works in the same way as toStartOfX, where X is Day, Week, Month, Quarter, Year when the enable_extended_results_for_datetime_functions is on. #95011 (Kirill Kopnev).
• Fix constant string comparisons not respecting the settings cast_string_to_date_time_mode, bool_true_representation, bool_false_representation, and input_format_null_as_default. Closes #91681. #95040 (Nihal Z. Miaji).
• Fix data race in the filesystem cache. #95064 (Alexey Milovidov).
• Fix a rare race condition in the Parquet reader. #95068 (Alexey Milovidov).
• Fix crash in the top K optimization when LIMIT is zero. Closes #93893. #95072 (Alexey Milovidov).
• Converting from DateTime/integers to Time64 extracts the time-of-day component using toTime, which is not monotonic. The ToDateTimeMonotonicity template incorrectly claimed this conversion was monotonic, causing “Invalid binary search result in MergeTreeSetIndex” exception in debug builds. #95125 (Alexey Milovidov).
• Recreated list of manifest file entries only if necessary (previously it was done on each iteration). #95162 (Daniil Ivanik).

​

Build/Testing/Packaging Improvement

• Add a set of tools for profiling memory allocations in the ClickHouse SQL parser using jemalloc’s heap profiling capabilities. #94072 (Ilya Yatsishin).
• Added a tool that simplifies debugging of memory allocations in parser. It uses jemalloc stats.allocated metric before and after we parse query to AST representation to show what is allocated. Also it supports memory profiling mode that dumps profile before and after to build reports where allocations occurred. #93523 (Ilya Yatsishin).
• Remove transitive libc++ includes. #92523 (Raúl Marín).
• Make some sequential tests parallel: https://github.com/ClickHouse/ClickHouse/pull/93030/changes#diff-c3a73510dae653c9bbfa24300b32f5d6ec663fd4e72cc4a3d5daa6e4342915df. #93030 (Nikita Fomichev).
• Cleanup some build flags. #93679 (Raúl Marín).
• Bump c-ares from v1.34.5 to v1.34.6. This addresses c-ares’ CVE-2025-62408, which is not relevant for ClickHouse. #94129 (Govind R Nair).
• Use curl 8.18.0. #94742 (Konstantin Bogdanov).