"Andrus" <kobruleht2@hot.ee> writes:
>> I looked back at the previous thread you mentioned (bug #7914) and was
>> reminded that we never did understand what was going on in that report.
>> I'm not sure if you are seeing the same thing though. That user reported
>> that he was able to see pg_dump's memory consumption bloating well beyond
>> what it ought to be (I suppose he was watching the process in whatever
>> Windows' equivalent of ps or top is). Do you see that?
> This occurs in customer computer which I din't observe.
> I added --inserts parameter to pg_dump and ssl=false in postgresql.conf
> file.
> After that backup works OK.
Now that we've identified the actual problem [1], I'm guessing the reason
why --inserts appears to make it go away is that then pg_dump uses a
SELECT which requires more server-side effort than COPY. So that slows
down the server just a bit, and if the phase of the moon is favorable the
timing no longer results in this buffer-bloat behavior. But I wouldn't
call that a reliable fix. You'd be better off applying the patch.
regards, tom lane
[1] http://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=2f557167b19af79ffecb8faedf8b7bce4d48f3e1