Chasing that 32 ms tail latency in ClickHouse

T1Covered by 1 source · first reported by Databricks / ClickHouse / DuckDB / Supabase / Neon / PlanetScale at 13 Oct, 08:49 UTC

A customer saw an extra ~32 ms on simple SELECT queries after idle periods. We traced it to unbatched Native protocol metadata packets overflowing a reset TCP congestion window, reproduced it with network namespaces, and fixed it by batching them.

Sources

• T1Chasing that 32 ms tail latency in ClickHouseDatabricks / ClickHouse / DuckDB / Supabase / Neon / PlanetScale13 Oct, 08:49