On Sun, Apr 13, 2014 at 9:14 PM, Andres Freund <andres@2ndquadrant.com> wrote:
> On 2014-04-12 17:40:34 -0400, Robert Haas wrote:
>> On Fri, Apr 11, 2014 at 10:28 AM, Andres Freund <andres@2ndquadrant.com> wrote:
>> > VACUUM sometimes waits synchronously for a cleanup lock on a heap
>> > page. Sometimes for a long time. Without reporting it externally.
>> > Rather confusing ;).
>> >
>> > Since we only take cleanup locks around vacuum, how about we report at
>> > least in pgstat that we're waiting? At the moment, there's really no way
>> > to know if that's what's happening.
>>
>> That seems like a pretty good idea to me.
>
> What I am not sure about is how... It's trivial to set
> pg_stat_activity.waiting = true, but without a corresponding description
> what the backend is waiting for it's not exactly obvious what's
> happening. I think that's better than nothing, but maybe somebody has a
> glorious better idea.
>
> Overwriting parts of the query/activity sounds like it'd be somewhat
> expensive ugly.
>
>> I think we've avoided doing
>> this for LWLocks for fear that there might be too much overhead, but
>> it's hard for me to imagine a workload where you're waiting for
>> cleanup locks often enough for the overhead to matter.
>
> Hm. I am not sure I see the cost as a very compelling thing here. Sure,
> we can't list the acquired lwlocks and such, but it should be cheap
> enough to export what lwlock we're waiting for if we're going to
> sleep. I think it'd be worthwile making that visible somehow.
> But that's a separate issue...
How about having a view like pg_lwlocks similar to pg_locks which
can be used to capture and provide the useful information?
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com