Обсуждение: Re: [GENERAL] Postgres stats collector showing high disk I/O

Поиск
Список
Период
Сортировка

Re: [GENERAL] Postgres stats collector showing high disk I/O

От
Alvaro Herrera
Дата:
Excerpts from Justin Pasher's message of jue may 20 16:10:53 -0400 2010:

> Whenever I clear out the stats for all of the databases, the file
> shrinks down to <1MB. However, it only takes about a day for it to get
> back up to ~18MB and then the stats collector process start the heavy
> disk writing again. I do know there are some tables in the database that
> are filled and emptied quite a bit (they are used as temporary "queue"
> tables). The code will VACUUM FULL ANALYZE after the table is emptied to
> get the physical size back down and update the (empty) stats. A plain
> ANALYZE is also run right after the table is filled but before it starts
> processing, so the planner will have good stats on the contents of the
> table. Would this lead to pg_stat file bloat like I'm seeing? Would a
> CLUSTER then ANALYZE instead of a VACUUM FULL ANALYZE make any
> difference? The VACUUM FULL code was setup quite a while back before the
> coders knew about CLUSTER.

I wonder if we should make pgstats write one file per database (plus one
for shared objects), instead of keeping everything in a single file.
That would reduce the need for reading and writing so much.

--

Re: [GENERAL] Postgres stats collector showing high disk I/O

От
Tom Lane
Дата:
Alvaro Herrera <alvherre@alvh.no-ip.org> writes:
> Excerpts from Justin Pasher's message of jue may 20 16:10:53 -0400 2010:
>> Whenever I clear out the stats for all of the databases, the file
>> shrinks down to <1MB. However, it only takes about a day for it to get
>> back up to ~18MB and then the stats collector process start the heavy
>> disk writing again.

> I wonder if we should make pgstats write one file per database (plus one
> for shared objects), instead of keeping everything in a single file.
> That would reduce the need for reading and writing so much.

Well, the real problem here is the OP is running 8.1, so he isn't
getting the benefit of any of the subsequent changes to reduce the
stats file I/O.  I suspect the bloat also comes from a
subsequently-fixed bug, but not totally sure about that yet.

            regards, tom lane