Re: compute_query_id and pg_stat_statements

On Wed, May 12, 2021 at 08:36:18PM -0400, Bruce Momjian wrote:
> On Wed, May 12, 2021 at 05:51:49PM +0800, Julien Rouhaud wrote:
> > On Wed, May 12, 2021 at 11:42:12AM +0200, Pavel Stehule wrote:
> > > this check just can check if there is "any" query-id provider. In this
> > > context is not important if it is buildin or external
> > 
> > Yes, the idea is that if you execute "SELECT * FROM pg_stat_statements" or
> > similar, then if the executing query itself doesn't have a queryid then
> > it's very likely that you didn't configure compute_query_id or an alternative
> > query_id implementation properly.  And loudly complaining seems like the right
> > thing to do.
> 
> I understand the desire to make pg_stat_statements require minimal
> configuration, but frankly, if the server-side variable query id API is
> confusing, I think we have done more harm than good.
> 
> The problem with compute_query_id=auto is that there is no way to know
> if the query id is actually enabled, unless you guess from the installed
> extensions, or we add another variable to report that, and maybe another
> variable to control the provier, unless we require turning
> compute_query_id=off if you are using custom query id computation.  What
> if it is auto, and pg_stat_statments is installed, and you want to use a
> custom query id computation --- what happens?  As you can see, this is
> all becoming very complicated.

Well, as implemented you can get the value of compute_query_id, and if it's
still "auto" then it's not enabled as calling queryIdWanted() would turn it to
on.  I agree that it's not ideal but you have a way to know.  We could document
that auto means that it's set to auto and no one asked to automatically enabled
it.

Or you can just do e.g.

SELECT query_id FROM pg_stat_activity WHERE pid = pg_backend_pid();

and see if you have a query_id or not.

If you want to use third-party modules, they you have to explicitly disable
compute_query_id.  If you don't, every query execution will raise an error as
we documented that third-party modules should error out if they see that a
query_id is already generated.  Such module could also explicitly check that
compute_query_id is off and also raise an error if that's not the case.

> I think we might be just as well to go with compute_query_id=on/off, and
> just complain loudly from CREATE EXTENSION, or in the server logs on
> server start via shared_preload_libraries, or when querying
> pg_stat_statements system view.  We simply say to change
> compute_query_id=on or to provide a custom query id implementation.

I'm not opposed to that, but it was already suggested and apparently people
didn't like that approach.