Alexey Milovidov
The ClickHouse v26.3 release — a long-term supported release — brings 27 new features, 40 performance optimizations, and more than 200 bug fixes, a new record. Alexey walks through the highlights: the sharded Map data type ships production-ready and enabled by default, splitting map keys into subcolumns during background merges so reads of a single key pull 20x less data without touching insert performance. Async inserts are now on by default, with consistent deduplication across regular inserts, async inserts, and materialized views unified for the first time. On the compatibility front, ClickHouse now passes 97%+ of SQLite's five-million-query SQL logic test suite — and the queries that fail are ones where ClickHouse actually behaves more correctly.
The standout experimental feature is WebAssembly user-defined functions — write functions in Rust, Go, C, or any language that compiles to Wasm, then call them as native ClickHouse functions. A Rust crate makes the integration ergonomic, and a community contributor has already built a full PostGIS-compatible API on top of it. There's also a new ALP codec for adaptive lossless floating-point compression that outperforms Gorilla and FPC for industrial time series and telemetry workloads.
Guest speaker Oxton presents chDB v4, which achieves zero-copy reading and writing between ClickHouse and pandas DataFrames — no serialization, no Parquet round-trips. The result is a one-line import swap (import chdb.dataframe as pd) that runs existing pandas code on the ClickHouse engine, typically 4x faster on a laptop and up to 200x faster on multi-core machines, with dramatically lower memory usage.
This session covers:
- Materialized CTEs — subqueries now run once and reuse results, demonstrated at 2x speedup in a multi-join query
- Natural sort key function for sorting version strings and filenames numerically rather than lexicographically
- Vertical merge algorithm now supported for TTL deletes, reducing memory pressure during compaction
- Iceberg metadata prefetching with configurable staleness, and S3Queue ordered mode cutting list API calls with start-after
- Data lake parallelization re-enabled — up to 100x improvement on remote Parquet queries
- Full join reordering extended to anti-joins and semi-joins
- 31 alternative SQL dialects (Snowflake, Teradata, and more) via the dialect setting
