Re: Query result differences between PostgreSQL 17 vs 16

On Sat, Mar 1, 2025 at 7:16 AM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> So in the case at hand, we generate a clone of the "rc3.id IS NOT
> NULL" clause that doesn't have a nullingrel bit for the left join to
> rc3, but nonetheless is expected to be applied after that join is
> performed. There is also a clone that does have a nullingrel bit, but
> that doesn't save us in this example because it will only be applied
> in join orders other than the one that gets chosen.

Exactly!

> It'd certainly be nice to make that cleaner, but I'm not sure how,
> and we wouldn't risk back-patching the nontrivial changes that would
> be needed. So I think rejecting has_clone/is_clone clauses is the
> only path forward for now.

Yeah, a better fix might be to find a way to perform this NullTest
deduction during eval_const_expressions, as you mentioned in the other
thread. However, that would also require some nontrivial changes.

> Also, I found a test case proving that restriction_is_always_false
> needs it too.

Thanks. I've included this test case in v3 patch.

Thanks
Richard