Conversation

Amazing article, exactly mirrors what I saw w another PG user. Very important: wasn't savepoints alone - was their combination with long running transactions - that caused big system perf stalls. Happiness hint to sample pg_stat_activity isn't just for waits; shows long xacts too

Quote Tweet

Dammian Miller

@DammianMiller

Sep 30, 2021

Why we spent the last month eliminating PostgreSQL subtransactions disq.us/t/41jn7w3

Peter Geoghegan

@petervgeoghegan

Replying to

@jer_s

Right. There is no logical reason why any long running transaction should slow down checking if a certain XID committed on a standby. We're just answering a question about a physical XID, which has little to do with tuple visibility. Not the same thing at all.

8:55 PM · Sep 30, 2021Twitter Web App

Replying to

and

A simpler way of saying almost the same thing: it bothers me that the problem is *only* on replicas, since it is "caused" on the primary. As a general rule, transactions should not be able to create debt that unrelated transactions must pay - tends to destabilize whole system.