On Wed, Oct 30, 2024 at 10:45 AM Peter Geoghegan <pg@xxxxxxx> wrote:
It sounds like you have no updates and deletes. Right? So the only
thing that could be different is the way that the pages are being
split (aside from variations in the width of index tuples, which seems
highly unlikely to be the only factor).
Correct, the table only sees inserts as far as DML goes.
The heuristics used to trigger the relevant behavior are fairly
conservative. I wonder if we should be more aggressive about it.
> I should have also included that this is on PG 15 (currently 15.8 but we created the indexes when it was still 15.7) on Ubuntu 22.04 LTS.
That shouldn't matter, as far as this theory of mine is concerned.
Anything after 12 could be affected by the issue I'm thinking of.
Why would last month's index be so much smaller?
Both indexes were created using CONCURRENTLY, as each was created during its month when we started testing. The September index was created toward the end of the month (Sep 26), whereas the October one was created Oct 1. Both table partitions are getting regularly autovacuum/autoanalyze work.
Don.
-- Don Seiler
www.seiler.us
www.seiler.us