Re: Teaching planner to short-circuit empty UNION/EXCEPT/INTERSECT inputs

On Wed, 5 Nov 2025 at 17:00, Alexander Lakhin <exclusion(at)gmail(dot)com> wrote:
> But while playing around, I've just discovered another interesting error:
>
> SET enable_seqscan = 'off';
> SELECT * FROM t EXCEPT SELECT * FROM t
> UNION SELECT * FROM pt;
>
> leads to:
> ERROR: XX000: no relation entry for relid 0
> LOCATION: find_base_rel, relnode.c:541

:-(

This happens because of the code I added to try to give better
estimates to the number of groups when only a single child remains in
the UNION. Because the only surviving Path is for the EXCEPT set-op,
there are Vars with varno==0, which estimate_num_groups() does not
like. I'm considering restricting that code Path to where the
path->parent->reloptkind != RELOPT_UPPER_REL. It doesn't feel great
adding the special case, but likely having better row estimates is
worthwhile, when it's possible. I'll sit on that for a bit and see if
I can think of a better way.

(Yet another reason we should probably fix the varno==0 issue.)

Thanks for the report and reproducer, again!

David