Postgres: stories that will scare you
summary
A Postgres operations team that runs about 10–15k instances describes real failure cases and their causes: a FreeBSD replica using a C collation that corrupts indexes requiring full reindex and manual deduplication; a misconfigured connection pooler plus hourly jobs opening thousands of connections; workloads creating 900 subtransactions that exhaust a fixed cache in pre‑Postgres‑17 builds; an accidental deployment that reduced shared_buffers by 4096× causing extreme CPU/system time; unusable SQL dump backups; and a firmware/network bug that crashed primary servers.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:06Nox context
- 2:29Collation problem
- 10:33Connection overload
- 15:11Misconfigured parameter
- 21:33Subtransactions
- 27:13Cluster crash
- 32:26Moral and advice
- 35:16Q&A
related talks
51mHow Postgres is misused and abused in the wild
01h 00mThe 10 Deadly PostgreSQL Sins
36mTales from the Postgres Front, and What We Can Learn
36mPostgres at any Scale
47mPostgreSQL worst practices
13mBuilding reliable infrastructure with Postgres
37mWhen it all goes Wrong (with Postgres)
51mScaling Postgres to the next level at OpenAI