Yarrow.dev · speaker landing

"The shape of a slow query — a forensic walkthrough."

A 35-minute talk on diagnosing the kind of postgres queries that look fine in EXPLAIN and still ruin your afternoon. Live demos, real production traces, no slides longer than three lines.

Event: PGConf Europe, Tallinn
Date: 17 October 2026 — 11:30 EET
Room: Old Stage
Length: 35 min talk + 10 min Q&A
Recording: Yes (CC BY) — link will replace this page after the event.
Slides: Posted to this URL, in plain HTML, 24h after the talk.
Rehearsal: Open evening on the day before, in the Old Stage foyer. Bring questions.

"The hard part isn't the bad plan. It's the plan that's good on average and ruinous for the one query that matters." — rehearsal notes, May 2026

Three things we will cover

Why pg_stat_statements is only the start of the conversation, not the end of it.
The pattern of "fast in staging, slow in prod" — and how to actually pin it down without ssh-ing into a replica at 3am.
The art of a one-line index that fixes everything, and the discipline of not believing it for at least a week.
What auto_explain quietly tells you, if you let it.
And one war story about a query that turned out to be perfectly correct, only running 41 million times an hour.

Speaker

A. Yarrow

Background

11 years of platform engineering, 3 of them on systems running postgres harder than they should.

Previously

"How fsync ate my lunch", FOSDEM 2024

Not sponsored by

any cloud database vendor

For the curious: the talk repository will appear at this same URL after October 17th. Until then, all you'll find here is this page. If you wrote in during May with a follow-up question — yes, I'll work an answer into the Q&A.

This page is the speaker's personal landing for the talk. Slides and the demo repo will appear here within 48 hours of the session. srv-yarrow-7.eu-west.internal