ClickHouse v26.9 Release Webinar

Alexey Milovidov

The ClickHouse v26.9 release webinar explores how the database is expanding across analytical, observability, data-lake, and AI workloads. Alexey Milovidov demos new syntax for retrieving the rows surrounding an event, DISTINCT queries that spill to disk, scoped temporary tokens, incremental replication into Iceberg, and ClickHouse acting as a Prometheus-compatible server. There’s also a self-hosted SQL console, improved embedded documentation, new query dialects, and another substantial set of performance improvements. Bonus Q&A: is ClickHouse becoming a universal, multimodal DBMS? Alexey has some thoughts.

• Find the rows surrounding an event — new LIMIT AFTER and LIMIT UNTIL extensions locate a matching row and return the records before, after, or between conditions. Alexey demonstrates it by finding an error in the ClickHouse text log and retrieving the ten messages that followed it.

• Incremental refreshable materialized views — scheduled refreshes can track changes to a source MergeTree table and process only the accumulated delta. This makes it possible to replicate data incrementally and exactly once into systems such as Iceberg without inserting into the destination continuously.

• ClickHouse speaks Prometheus — expose a Prometheus-compatible HTTP API, connect Grafana or ClickStack directly, and query ClickHouse using PromQL. PromQL can be used as the active query dialect or embedded inside SQL through a table function.

• Large DISTINCT queries can spill to disk — ClickHouse can move intermediate data into external memory when a query reaches its RAM threshold, then merge the results at the end. The demo shows only a small performance difference between the in-memory and on-disk executions.

• Scoped, temporary credentials — individual authentication methods can now receive their own subset of a user’s grants. The new CREATE TOKEN syntax generates an expiring credential, valid for 30 minutes by default, which can be restricted to specific operations.

• Documentation that matches your server — the new system.statements table contains syntax and Markdown documentation for SQL statements. Use HELP SELECT in the terminal or browse the reference from ClickHouse’s embedded web interface, with content corresponding to the exact server build.

• More ways to run ClickHouse locally — turn ClickHouse Local into a server using SYSTEM START LISTEN HTTP, inspect recently executed query IDs through system.session_query_ids, and use the ClickHouse Cloud SQL console with any ClickHouse installation.

• Performance work across the engine — min/max aggregations can be answered from column statistics, empty columns can be omitted from storage, and native-protocol queries need fewer network round trips. The release also improves ARM execution, joins, sorted-stream merging, recursive CTEs, aggregation, object-storage access, and text-index merges and phrase searches.

• More capable text and JSON queries — the text index now accelerates prefix and suffix searches, indexes can still be applied after ARRAY JOIN, and JSON subcolumns can be accessed using array-style subscripts as well as dot notation. The new arrayFlattenedLength function counts the leaf values in nested arrays without materializing a flattened copy.

• Data-lake improvements — create new Delta Lake tables directly from ClickHouse, process Iceberg manifests concurrently, prune unnecessary manifests using summary metadata, and write boundary values that help subsequent queries eliminate more data.

• Compatibility for existing query workloads — expanded experimental Kusto support adds more faithful syntax and function behaviour, while the experimental Trino dialect maps hundreds of functions for applications with an existing collection of Trino or Presto queries.

• More goodies — limits on table size and the number of tables in a database, parallel replicas for Merge tables, SQL-standard regression aggregates, an exclusive single-server mode for S3Queue, AI functions promoted beyond experimental status, and On-Demand Compute for scaling individual queries across additional workers.