ClickHouse v26.10 Release Webinar

Alexey Milovidov

The ClickHouse v26.10 release webinar covers the biggest release yet: 104 new features (the first release to pass 100), 225 performance optimizations and 739 bug fixes. Alexey Milovidov demos the new obfuscate table function generating an endless stream of anonymized test data from a query, training an XGBoost model as a dictionary and predicting with it in SQL, the new columns cache, LogsQL as a query dialect, and reading an Elasticsearch index straight from ClickHouse. There's also a guessing game with the chat on how much faster each new optimization makes a live query.

• Half the memory for hash tables — large hash tables switch to Robin Hood hashing, halving memory for GROUP BY, uniqExact, DISTINCT and JOIN while also getting faster (uniqExact over 100M values: 13.5 GB → 4.6 GB, 1.33 s → 0.31 s)
• Caching, everywhere — the query condition cache now works for now() and today() filters by snapping them to a time grid, the query result cache can live on disk and survive restarts (shared across replicas via the distributed cache in ClickHouse Cloud), and the experimental columns cache keeps decompressed, deserialized columns in memory for MergeTree and Parquet
• Reading less data — string filters like LIKE and startsWith are applied while strings are read (275 MB → 5 MB peak memory), and Top-K dynamic filtering pushes the ORDER BY … LIMIT boundary into MergeTree scans (9.45 GB → 144 MB read) and Parquet and Iceberg readers
• Faster joins and concurrency — merge joins split into parallel ranges of the primary key (9.1 s → 0.78 s), JOIN ON can use any condition, IN subqueries can spill to disk, query pipelines narrow under high load, and workloads gain fair, priority and short-queries-first scheduling
• New SQL — LATERAL JOIN for top-N-per-group queries, a function composition operator for lambdas (arrayMap(lower | reverse, …)), materialized CTEs inside recursive queries, order-insensitive query normalization, and an opt-in setting for SQL-standard integer division
• New engines and formats — Cluster and Overlay database engines for federating and shadowing databases, an Elasticsearch table engine, SQLite files as an import/export format, a Freeform format that infers the structure of any tabular text, and the PCO codec for numeric columns
• Data lakes — mutations and hard links on s3_plain_rewritable disks, VARIANT columns from Parquet and Iceberg read as Dynamic, Unity Catalog v2 for both Delta Lake and Iceberg tables, and Iceberg v3 deletion vectors
• More goodies — written_rows quotas, IPCrypt format-preserving IP encryption, bit-packing functions for embedding quantization, Trino and Kusto dialects in Beta, chdig with a Keeper browser, psql query cancellation, logrotate-friendly FileLog, PromQL at 100% on the Prometheus compliance suite, and notebooks in the built-in Web UI. In ClickHouse Cloud: executable UDFs are generally available, plus JWT authentication and SCIM 2.0 provisioning