Josh Berkus <josh@agliodbs.com> writes:
>> select count(1), sum(relpages) from pg_class where relkind in
>> ('r','i','t')
> Well, if you do that for all databases in the cluster, it's the number you
> start with. However, setting FSM_pages to that would be assuming that you
> excpected 100% of the rows to be replaced by UPDATES or DELETEs before you
> ran VACUUM. I generally run VACUUM a little sooner than that.
Not at all. What it says is that you expect 100% of the pages to have
useful amounts of free space, which is a *much* weaker criterion.
I think you can usually get away with setting max_fsm_pages to less than
your actual disk footprint, but I'm not sure how much less. It'd
probably depend a lot on your usage pattern --- for instance,
insert-only history tables don't need any FSM space.
regards, tom lane
