Re: Why does autovacuum clean fewer rows than I expect?

Hi Tom,

Thanks for the quick reply. 8.4.9. No db crashes, in fact not even a
restart in May until the 28th. I don't believe any HOT updates were
occurring in that table: all updates modify one or more indexed fields
(update status from pending to done, or update to store a stack trace and
bump the time at which to retry the row into future). We don't delete from
the table, just insert and update.

During the time interval in question a bug was leading to an unusually high
level of activity, both inserts and updates. I'm thinking this is what
was responsible for a huge amount of table and index bloat, but haven't yet
connected all the dots. Since I don't see autovacuum failing, or failing
to run, it's a bit of a mystery.

I guess the next step is to attempt to reproduce the bug on a test system
and see if we get the bloat. I'll let you know what turns up.

On Thu, Jul 3, 2014 at 5:35 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Nick Cabatoff <nick(dot)cabatoff(at)gmail(dot)com> writes:
> > I'm having trouble making sense of the these two autovacuum log entries.
> > I'm running PostgreSQL 8.4.
>
> 8.4.what?
>
> It'd probably be a good idea to monitor the counts in pg_stat_all_tables
> for the troublesome table(s). I suspect there are a whole lot of HOT
> updates going on in this table, which'd account for autoanalyze activity
> happening even though autovacuum finds few dead tuples to remove. That
> still doesn't directly explain physical table bloat, but it might be
> a piece of the puzzle.
>
> Also, have you had any database crashes in between these autovacuums?
> I'm wondering if the WAL-replay issue fixed in commits 6f2aead1f and
> siblings could have produced the bloat.
>
> regards, tom lane
>