On Fri, Mar 27, 2026 at 3:19 AM Tender Wang <tndrwang@gmail.com> wrote:
> Andrei Lepikhov <lepihov@gmail.com> 于2026年3月27日周五 03:59写道：
> >
> > On 20/3/26 15:02, Alexander Korotkov wrote:
> > > OK. I've pushed this. Let's go back to
> > > restrict_infos_logically_equal(). I'm still not convinced that we
> > > need to check if required_relids is singleton. Why we can ignore
> > > outer_relids for singleton, but can't do if, for instance, two
> > > relations involved?
> >
> > Let's continue. In the attachment, the Tender's proposal that I changed
> > a little and added some tests.
> >
> > As you can see in the tests, the SINGLETON limitation keeps duplicates
> > of clauses like 'a.x + b.y = c'.
> > This example shows the main flaw of this approach. Introducing the
> > restrict_infos_logically_equal(), we do a little more job than just the
> > equal() routine could because of the context where we call this function
> > and on which clauses.
> > But skipping all other RestrictInfo fields except required_relids seems
> > excessive. - see the example with security_level difference - ignoring
> > its value, we potentially might remove the clause with enforced security
> > level in favour of an unsecured one.
>
> Yes, it seems too strict to require all fields to be equal, but
> skipping some fields is unsafe.
>
>
> > That's more, further some new optimisations might introduce more fields
> > into RestrictInfo that should be checked to correctly decide on the
> > equality, and we may forget to arrange this specific place.
> >
>
> Agree.
>
> > So, formally it works, and making the following replacement, we close
> > the singleton issue:
> >
> > - if (bms_membership(a->required_relids) == BMS_SINGLETON &&
> > - a->security_level == b->security_level)
> > + if (bms_equal(a->required_relids, b->required_relids) &&
> > + a->security_level == b->security_level &&
> > + a->is_pushed_down == b->is_pushed_down)
> >
>
> The singleton issue does not seem to be the correct way; I don't dive
> deeply to cover all cases.
>
> > but I'm unsure, in general, that this approach is conservative enough.
> > Maybe we shouldn’t change this logic and invent one more optimisation
> > ‘deduplication’ stage later, before the selectivity estimation stage.
I have another approach about to deduplication of RestrictInfo's. The
field, which differs in this case, is outer_relids. AFAICS,
outer_relids and incompatible_relids serves as the restriction on what
we can do with RestrictInfo. So, what we can do is to ignore both
outer_relids and incompatible_relids during comparison, but compose a
union of their values for remaining RestrictInfo. That means that
remaining RestrictInfo will ancest all the restrictions, and that
should be safe.
What do you think?
------
Regards,
Alexander Korotkov
Supabase
