Tore Halset <halset@pvv.ntnu.no> writes:
> On Oct 15, 2012, at 2:27 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Tore Halset <halset@pvv.ntnu.no> writes:
>>> On this box I drop a 80GB database each night followed by a restore of a similar sized database. It is a restore of
ourproduction database to a development server. This box is running 9.2rc1 (sorry).
>>> du and df reported quite different numbers and lsof show that autovacuum is holding lots of deleted files. After
killingthe autovacuum daemon, some disk space was restored and the du and df numbers was more equal.
>> Hm. I've been able to reproduce some leakage of file descriptors in the
>> autovac launcher, but it required (a) fairly small shared_buffers and
>> (b) very heavy update activity on large tables. So I'm not sure that it
>> would explain the consistent leakage you seem to be seeing. Can you
>> tell us more about your usage pattern on the development server?
> A cron job dropdb one of the databases and createdb it and then pg_restore. Roughly 80GB dump.
So far, my guess is that this is fixed by commits a1f064fc2 + d7598aeea.
> Out production PostgreSQL running a 9.1 variant does not have this problem. It does not have the nightly
dropdb/createdb/pg_restore,but otherwise similar usage patterins. It is also configured to use more memory.
9.1 has the same bug (and the same patches), but I think that the aspect
you're running into is specific to DROP DATABASE scenarios.
regards, tom lane