On Fri, Sep 4, 2009 at 2:48 PM, Tom Lane<tgl@sss.pgh.pa.us> wrote:
> Josh Berkus <josh@agliodbs.com> writes:
>>>> I have a client who uses temp tables heavily, hundreds of thousands of
>>>> creates
>>>> and drops per day. They also have long running queries. The only
>>>> thing that
>>>> keeps catalog bloat somewhat in check is vacuum full on bloated catalogs
>>>> a few times a day. With
>
>> Actually, this is a good point ... if we dropped VACUUM FULL, we'd need
>> to also be able to call CLUSTER (or VACUUM REWRITE) on the system catalogs.
>
> I don't think I believe the claim above that vacuum full is actually
> necessary. Reasonably aggressive regular vacuuming ought to do it.
>
> We used to have a bug that caused row deletions during backend shutdown
> to not get reported to the stats collector; which had the effect that
> dead catalog entries for temp tables didn't get counted, and so autovac
> didn't hit the catalogs often enough, and so you'd get bloat in exactly
> this scenario. I suspect the claim that manual vacuum full is necessary
> is based on obsolete experience from before that bug got stomped.
> It's hardly an ideal solution anyway given what an exclusive lock on
> pg_class will do to the rest of the system --- and a cluster-like
> cleanup won't be any better about that.
I'm confused. Are you saying that pg_class will never get bloated, so
we don't need a way to debloat it? I realize that with HOT bloat is
much less of a problem than it used to be, but surely it's not
altogether impossible...
...Robert
