Re: compute_query_id and pg_stat_statements

On Tue, May 11, 2021 at 03:04:13PM +0900, Michael Paquier wrote:
> On Tue, Apr 27, 2021 at 02:25:04PM +0800, Julien Rouhaud wrote:
> > On Mon, Apr 26, 2021 at 11:37:45AM -0700, Andres Freund wrote:
> >> On 2021-04-26 14:21:00 -0400, Tom Lane wrote:
> >>> That's sounding like a pretty sane design, actually.  Not sure about
> >>> the shared-library-name-with-fixed-function-name detail, but certainly
> >>> it seems to be useful to separate "I need a query-id" from the details
> >>> of the ID calculation.
> >>> 
> >>> Rather than a GUC per se for the ID provider, maybe we could have a
> >>> function hook that defaults to pointing at the in-core computation,
> >>> and then a module wanting to override that just gets into the hook.
> >> 
> >> I have a preference to determining the provider via GUC instead of a
> >> hook because it is both easier to introspect and easier to configure.
> 
> So, this thread has died two weeks ago, and it is still an open item.
> Could it be possible to move to a resolution by beta1?  The consensus
> I can get from the thread is that we should have a tri-value state to
> track an extra "auto" for the query ID computation, as proposed by
> Alvaro here:
> https://www.postgresql.org/message-id/20210426174331.GA19401@alvherre.pgsql
> 
> Unfortunately, nothing has happened to be able to do something like
> that.

My understanding was that there wasn't a consensus on how to fix the problem.

Anyway, PFA a patch that implement a [off | on | auto] compute_query_id, and
provides a new queryIdWanted() function to let third party plugins inform us
that they want a query id if possible.

As it was noted somewhere in that thread, that's a hack on top on the GUC
machinery, so compute_query_id will display "on" rather than "auto" (or "auto
and enabled" or whatever) since GUC isn't designed to handle that behavior.

For the record I also tested the patch using pg_qualstats(), which can be
loaded interactively and also benefits from a query identifier.  It works as
expected, as in "query idenfitier are enabled but only for the backend that
loaded pg_qualstats".