I see a lot of people want to talk about #PostgreSQL things that go wrong with stuff I wrote. When autovacuum goes wrong; how multixacts require extra table freezing; bgworkers handle signals wrong.
Today: "subtransactions -- can go wrong badly".
Conversation
Replying to
People often overattribute problems in this area to an underlying general design. The implementation itself may well be improved within the confines of an old design, though. My argument is usually harder to make, and less pithy, but that tells us nothing about what really works.
1
In this particular context (Subtrans SLRU), if we just allow configuring its size and increase it (from 32 pages to 1 MiB or so), it will hardly help... gitlab.com/postgres-ai/po
But maybe there are other ways to improve without changing design
cc
3
1
The patch to have SLRUs be in the normal buffer pool will address some part of this in a more fundamental way than just changing a small constant sized buffer to a slightly less small constant sized buffer...
1
3
Bigger (and faster) SLRU is a horse when we want an automobile. Yet we could have this horse long long time ago without tremendous effort.
2
2
That definitely seems possible. I am not sure either way.