ClickHouse 26.5 Ships 38 New Features and 51 Performance Optimizations in Its Spring Release
ClickHouse 26.5 arrives on May 21 with 38 new features, 51 performance optimizations, a 27x Merge table speedup, Iceberg V3 geometry support, and new SQL compatibility aliases for PostgreSQL, Snowflake, and BigQuery.
Overview
ClickHouse shipped version 26.5 on May 21, 2026, delivering a release the team nicknamed its “Spring” edition. According to the official release presentation, the update contains 38 new features, 51 performance optimizations, and 224 bug fixes. The changelog confirms the release date as 2026-05-21, with version v26.5.1.882-stable now available on GitHub for AMD64 and ARM64 platforms.
Performance Improvements
The headline benchmark from the release presentation involves large Merge table queries: a 78 KB query on a Merge table over 500 tables that took 6.7 seconds in 26.4 completes in 0.25 seconds in 26.5. The improvement comes from the engine avoiding redundant structural examination across the constituent tables.
A new max_bytes_ratio_before_external_join setting, documented in the changelog, enables automatic hash join spilling to disk at a 0.5 memory ratio, helping workloads that previously exhausted memory during large joins. The release presentation describes the behavior: with 32 GiB free and 1 GiB per thread, ClickHouse allocates up to 32 threads; with only 1 GiB free, it falls back to 1 thread. This memory-aware parallelism caps thread counts under pressure rather than letting queries fail.
Grouping queries also benefit from a new optimization described in the presentation: aggregation now stops after collecting LIMIT + OFFSET distinct keys, avoiding full-table scans when only the top rows are needed. A related query_plan_top_k_through_join setting pushes ORDER BY…LIMIT before the join, so only the top N rows are joined rather than the whole table. Performance for UNION ALL queries is also improved, with new settings maintaining high parallelism but avoiding running every branch at once.
Lower-level improvements include data-parallel MD5 hashing multiple values at once and software prefetching for hash table probes, both detailed in the presentation. JIT compilation is now enabled on macOS, activating compile_expressions, compile_aggregate_expressions, and compile_sort_description on that platform.
New Features
The release adds a filesystem() table function that, as the presentation puts it, allows users to “list and analyze a directory as a queryable table.” A prettyPrintJSON function is included for log inspection, debugging, and report generation.
SQL compatibility receives several additions. New aliases, according to the presentation, include STRING_AGG as an alias for groupConcat (covering PostgreSQL, Snowflake, and BigQuery conventions) and STDDEV as an alias for stddevSamp for PostgreSQL compatibility. The CREATE OR REPLACE MATERIALIZED VIEW statement can now atomically swap a materialized view definition, and SYSTEM PAUSE VIEW temporarily stops a refreshable materialized view without dropping it.
The system.zookeeper_watches table, described in the presentation as revealing every active ZooKeeper watch, is added for operators debugging coordination state. An experimental web terminal is accessible at /webterminal, providing a browser-based ClickHouse client session.
Data Lake Integrations
For Apache Iceberg users, 26.5 introduces geometry and geography type support described in the presentation as mapping Iceberg V3 geometry and geography fields to the ClickHouse Geometry Variant type — an experimental feature marking ClickHouse’s first Iceberg V3 capability. An Iceberg query condition cache is also added, noted as useful for repeated dashboard queries.
Support for Apache Paimon arrives with new table engines that, per the presentation, support incremental reads with progress tracking in Keeper.
Backward Incompatible Changes
The changelog documents ten backward incompatible changes in 26.5. The most widely impactful is a default shift for date_time_input_format and cast_string_to_date_time_mode from basic to best_effort, broadening the range of datetime strings parsed automatically. The Arrow-based Parquet reader and writer are removed in favor of the native implementation. HTTP connection limits tighten: http_max_fields drops from 1,000,000 to 1,000, and http_max_field_name_size from 128 KB to 4 KB. The kql table function is removed; users should use SET dialect = 'kusto' instead. Window functions RANK and DENSE_RANK now reject arguments per the SQL standard.
What We Don’t Know
The ClickHouse team has not published a dedicated blog post for 26.5; the May 2026 newsletter focuses on the 26.4 release and the upcoming Open House conference. Performance numbers beyond the Merge table benchmark have not been independently reproduced.
Context
The release lands days before ClickHouse’s Open House user conference, described in the May 2026 newsletter as a free three-day event running May 26-28 at Convene in San Francisco. The newsletter names CEO Aaron Katz and CTO Alexey Milovidov among the speakers, alongside practitioners from Visa, Cisco, Shopify, and Zoox. Workshops are scheduled on real-time analytics, observability, AI agents, and database administration.