Re: BUG #11444: autovacuum stuck for 5 days and waits on a lock

On Wed, Sep 17, 2014 at 6:12 PM, Andres Freund <andres(at)2ndquadrant(dot)com> wrote:

>> There are also UPDATE statements constantly running on this table, in the
>> overlapping manner, so at a single moment there is at least one update
>> running on it. We are investigating why is it done this way, but can it be a
>> reason behind this strange vacuum behavior?
>
> Which quite possibly is caused by this.
>
>
> I've recently commented on -hackers that this is a very hard to debug
> behaviour and should be made more visible, but IIRC we didn't come to a
> conclusion...

Thank you, we've verified that the index in question is scanned as a
result of the sub-SELECT statement producing the data for the
resulting UPDATE. We are looking into simplifying it to avoid the
long-running scans, hopefully, we'll pull this before the transaction
wraparound will come for this cluster :-)

--
Regards,
Alexey Klyukin