Performance

The bug went away when I added logging. I know. Here is what was going on.

Defaults are fine for toys. Here is what I changed to make pgvector work on a real dataset.

The default JSONB index in Postgres is GIN, and there's a very common case where that's the wrong choice.

In Postgres 12 the optimization fence around CTEs came down. For us, that was mostly bad.

I used to be scared of perf record in prod. Then I wasn't.

A query that should have been a boring index scan turned into a full-table shuffle because the planner guessed wrong.

We flipped our edge servers from CUBIC to BBR and measured the actual win for our traffic, which was smaller than the marketing suggests

A weekend of measuring, swapping, and tuning reduced my rack's idle draw from 220W to 180W, with surprises

Replacing an epoll loop with io_uring gave a 1.8x throughput boost in a naive benchmark, and a 0.9x loss in a realistic one

A staging VM's OOM killer reliably picked postgres over a misbehaving test process, and the fix was understanding oom_score_adj

Every non-cluster DNS lookup in our pods was paying for five failed attempts first, and lowering ndots cut tail latency significantly

My Proxmox host kept ballooning to 60 GB of ARC and starving VMs, and the fix was not what I expected

Tail call optimization, manual trampolines, and a crash that took me three days to trace to the optimizer

Portable SIMD is real and it's good; a walkthrough of a checksum I vectorized and what surprised me

Dynamic linking on ELF is a two-step dance through the PLT and the GOT, and once you see it you cannot unsee it