Alexey Milovidov
The ClickHouse v26.4 release brings 39 new features, 45 performance optimizations, and 238 bug fixes — a new record. Alexey and Tyler, presenting together in person for the first time across 40+ release calls, walk through the highlights: float parsing gets a 2x speedup by swapping the DragonBox library for Zme, the stem function graduates from experimental to production across all compound types, and LIKE expressions now automatically use text indexes without any query changes. Column statistics are enabled by default, with the overhead deferred to background merges so inserts stay fast — the final piece needed for automatic join reordering with the cost-based optimizer to work out of the box.
The bigger story this release is AI landing directly in ClickHouse. New functions — ai_generate, ai_classify, ai_extract, and ai_translate — connect to any LLM endpoint via a simple config block with API key and model name, with built-in token-limit controls and usage introspection. Alongside that, a new chctl CLI manages multiple ClickHouse installations and can install agent "skills" that teach LLMs how to design schemas and write efficient queries. There's also a live benchmark showing ClickHouse Postgres hitting 28,000 TPS — 2-3x faster than Aurora, Neon, and Supabase.
This session covers:
- obfuscate_query, highlight, array_transpose, and autocorrelation — new utility functions for safe bug reporting and data processing
- Joins that exceed memory now spill to disk automatically via Grace hash join (experimental, enabled with one setting)
- Arrow Flight protocol extended from table reads to arbitrary SELECT queries
- Postgres compatibility additions: NATURAL JOIN, compound intervals, OVERLAY syntax, and Postgres-style VALUES
- Play UI updates — parameterized queries, multi-query result sets, and totals now visible in the built-in web interface
- Data lake improvements for Iceberg orphan file cleanup, Delta Lake partition pruning, and Unity Catalog GA preparation
- Quotas keyed by normalized query hash for per-query-structure rate limiting without an API layer
