On Fri, Feb 16, 2024 at 8:55 PM Andres Freund <andres@anarazel.de> wrote:
>
> Hi,
>
> On 2024-01-12 17:16:53 +0100, Magnus Hagander wrote:
> > On Thu, Jan 11, 2024 at 5:55 PM Bertrand Drouvot
> > <bertranddrouvot.pg@gmail.com> wrote:
> > > On Thu, Jan 11, 2024 at 02:24:58PM +0100, Magnus Hagander wrote:
> > > > On Wed, Jan 10, 2024 at 3:12 PM Bertrand Drouvot
> > > > <bertranddrouvot.pg@gmail.com> wrote:
> > > > >
> > > > > If we go the 2 fields way, then what about auth_identity and auth_method then?
> > > >
> > > >
> > > > Here is an updated patch based on this idea.
> > >
> > > Thanks!
> > >
> > > + <row>
> > > + <entry role="catalog_table_entry"><para role="column_definition">
> > > + <structfield>auth_method</structfield> <type>text</type>
> > > + </para>
> > > + <para>
> > > + The authentication method used for authenticating the connection, or
> > > + NULL for background processes.
> > > + </para></entry>
> > >
> > > I'm wondering if it would make sense to populate it for parallel workers too.
> > > I think it's doable thanks to d951052, but I'm not sure it's worth it (one could
> > > join based on the leader_pid though). OTOH that would be consistent with
> > > how the SYSTEM_USER behaves with parallel workers (it's populated).
> >
> > I guess one could conceptually argue that "authentication happens int
> > he leader". But we do populate it with the other user records, and
> > it'd be weird if this one was excluded.
> >
> > The tricky thing is that pgstat_bestart() is called long before we
> > deserialize the data. But from what I can tell it should be safe to
> > change it per the attached? That should be AFAICT an extremely short
> > window of time longer before we report it, not enough to matter.
>
> I don't like that one bit. The whole subsystem initialization dance already is
> quite complicated, particularly for pgstat, we shouldn't make it more
> complicated. Especially not when the initialization is moved quite a bit away
> from all the other calls.
>
> Besides just that, I also don't think delaying visibility of the worker in
> pg_stat_activity until parallel worker initialization has completed is good,
> that's not all cheap work.
>
>
> Maybe I am missing something, but why aren't we just getting the value from
> the leader's entry, instead of copying it?
The answer to that would be "because I didn't think of it" :)
Were you thinking of something like the attached?
--
Magnus Hagander
Me: https://www.hagander.net/
Work: https://www.redpill-linpro.com/