On Jul 8, 2009, at 3:42 PM, Tom Lane wrote:
> Joe Uhl <joeuhl@gmail.com> writes:
>> On Jul 8, 2009, at 3:00 PM, Tom Lane wrote:
>>> Hmm. In any case that shouldn't have led to a lock left hanging.
>>> Assuming that it was a regular and not autovacuum, do you know what
>>> the exact command would have been? (In particular, FULL, ANALYZE,
>>> etc options)
>
>> They were VACUUM VERBOSE ANALYZE. Specifically run with "/usr/bin/
>> vacuumdb -v --analyze $DB_NAME" in the cron job each night.
>
> We had bugs long ago (8.0 or before) where SIGTERM'ing a backend could
> result in lock entries not getting cleaned up, but that's all fixed
> now
> AFAIK. I tried some simple experiments with SIGTERM'ing a VACUUM in
> 8.3.x and couldn't reproduce a problem. So there must have been some
> other contributing factor here, but without any idea what it was, it's
> hard to investigate further :-(
>
> regards, tom lane
No worries, appreciate the response. I should have collected more
information at the initial state before diving in with the mass kill.
I'll restart tonight, and get a zabbix script monitoring the vacuums.
If they start piling up again i'll be able to gather more useful
information.