Alexey Milovidov
The ClickHouse v26.2 "winter" release brings 25 new features and the largest set of performance optimizations of any release to date. Alexey walks through the highlights: right and full joins are up to 3.5x faster after parallelizing the non-joined-records phase, JSON parsing for the native JSON data type is 2.7x faster, and primary key expressions now propagate automatically to WHERE clauses — meaning a table keyed by cityHash64(user_id) no longer requires spelling out the hash expression in every query. OpenTelemetry tracing lands in ClickHouse Keeper with automatic sampling so the overhead stays negligible.
Two features that have been in the works for years both cross into production this release. The full-text search index — prototyped externally in 2022, rewritten in 2025 — is now stable, with a live demo showing orders-of-magnitude speedups for token-based log and text searches that make Elasticsearch and Splunk replacements genuinely practical. And QBit, the vector embedding data type that lets you trade precision for speed at query time (from 16-bit full precision down to 5-bit fast), is also production-ready. Together they make ClickHouse a credible single store for observability, text search, and vector workloads.
This session covers:
- Embedded ClickStack ships inside ClickHouse by default — the full observability UI with automatic filters, no separate install required
- TOTP (time-based one-time passwords) in the ClickHouse client for two-factor auth
- Per-database lazy table loading, useful for servers with millions of external or file-backed tables
- Google BigLake (Iceberg catalog over BigQuery) support in beta, with a live 1-billion-row demo
- Streaming insert improvements: per-second flushing and incomplete-batch handling for Kafka-free ingestion
- Agent skills package to teach AI coding agents ClickHouse-specific schema and query patterns
- primes table function, OKLab perceptually uniform color space, and new introspection system tables
