Josh Berkus <josh@agliodbs.com> writes:
> I'm currently looking at a database which has some extreme bloating of
> intarray GiST indexes. As in 1000% bloating in only a few months. This
> is not a particularly high-transaction-rate database, so the bloating is
> a little surprising; I can only explain it if vacuum wasn't cleaning the
> indexes at all, and maybe not even then.
> We're currently instrumenting the database so that we can collect a bit
> more data on update activity, but in the meantime, has anyone seen
> anything like this?
1. What PG version?
2. If new enough to have contrib/pgstattuple, what does pgstattuple() have to say about the index?
I'm suspicious that this might be bloat caused by a bad picksplit function,
not from having a lot of dead entries in the index. We've fixed several
other bogus picksplit functions in contrib in the past.
regards, tom lane