Re: Expose lock group leader pid in pg_stat_activity

On Tue, Jan 28, 2020 at 02:26:34PM +0100, Julien Rouhaud wrote:
>On Tue, Jan 28, 2020 at 2:09 PM Tomas Vondra
><tomas(dot)vondra(at)2ndquadrant(dot)com> wrote:
>>
>> I agree a separate "leader_id" column is easier to work with, as it does
>> not require unnesting and so on.
>>
>> As for the consistency, I agree we probably can't make this perfect, as
>> we're fetching and processing the PGPROC records one by one. Fixing that
>> would require acquiring a much stronger lock on PGPROC, and perhaps some
>> other locks. That's pre-existing behavior, of course, it's just not very
>> obvious as we don't have any dependencies between the rows, I think.
>> Adding the leader_id will change, that, of course. But I think it's
>> still mostly OK, even with the possible inconsistency.
>
>There were already some dependencies between the rows since parallel
>queries were added, as you could see eg. a parallel worker while no
>query is currently active. This patch will make those corner cases
>more obvious.

Yeah, sure. I mean explicit dependencies, e.g. a column referencing
values from another row, like leader_id does.

>Should I document the possible inconsistencies?

I think it's worth mentioning that as a comment in the code, say before
the pg_stat_get_activity function. IMO we don't need to document all
possible inconsistencies, a generic explanation is enough.

Not sure about the user docs. Does it currently say anything about this
topic - consistency with stat catalogs?

regards

--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services