Re: Removing unneeded self joins

Hi, Michael!

On Fri, Jun 27, 2025 at 3:26 AM Michael Paquier <michael(at)paquier(dot)xyz> wrote:
> On Thu, Jun 26, 2025 at 08:54:55AM +0200, Andrei Lepikhov wrote:
> > Before diving into the pg_hint_plan code, I wonder why you don't have
> > similar issues with the remove_useless_joins. We intentionally designed SJE
> > coupled with the left-join removal feature to avoid such type of complaints:
> >
> > CREATE TABLE test (x integer PRIMARY KEY);
> > EXPLAIN (COSTS OFF)
> > SELECT t1.* FROM test t1 LEFT JOIN test t2 ON (t1.x=t2.x);
> >
> > /*
> > QUERY PLAN
> > ---------------------
> > Seq Scan on test t1
> > (1 row)
> > */
> >
> > It seems that this join removal is also beyond the pg_hint_plan control ...
>
> Yeah, remove_useless_joins() has been around since the 9.0 ages as far
> as I know, and it's not bothered the module much in the tests because
> we have unlikely relied on it.
>
> For now the trick I am going to rely on is just disable
> enable_self_join_elimination in the test paths where I rely on the
> same relation and self joins to keep the plans of the regression tests
> stable. This stuff could create separate relations, but then with
> back-patching in mind that's going to be just extra conflict noise,
> which is always annoying when dealing with plan outputs. We've done
> that in the past with max_parallel_workers_per_gather or jit, to keep
> the expected plans stable.
>
> The point regarding the search join hook may stand, though. Perhaps
> somebody should check if we're still OK with this change in the
> context of the self-join work. I tend to think that we are and I
> agree that removing the joins when calling the hook can show benefits,
> but it may be surprising and users tend to be very noisy with plan
> stability, particularly if some of the FROM aliases get silenced by
> the backend without the module knowing about that. At least there is
> the trick with SET enable_self_join_elimination available as a last
> resort method.

One thing I could additionally propose it to add hook, which gets
called before self-join elimination. If pg_hint_plan will use this
hook, it could prevent self-join elimination or track this fact in its
data structures.

On the other hand we didn't do similar thing for
remove_useless_joins(). Not sure if it's justified now.

------
Regards,
Alexander Korotkov
Supabase