Re: Properly pathify the union planner

On Wed, 27 Mar 2024 at 16:15, Richard Guo <guofenglinux(at)gmail(dot)com> wrote:
> if (root->parent_root != NULL &&
> root->parent_root->parse->setOperations != NULL &&
> IsA(root->parent_root->parse->setOperations, SetOperationStmt))
> qp_extra.setop =
> (SetOperationStmt *) root->parent_root->parse->setOperations;
> else
> qp_extra.setop = NULL;
>
> This check cannot tell if the subquery is for a set operation or a CTE,
> because its parent might have setOperations set in both cases. Hmm, is
> there any way to differentiate between the two?

As far as I see, there's nothing to go on... well unless you counted
canSetTag, which is false for the CTE (per analyzeCTE())... but that's
certainly not the fix.

I did wonder when first working on this patch if subquery_planner()
should grow an extra parameter, or maybe consolidate some existing
ones by passing some struct that provides the planner with a bit more
context about the query. A few of the existing parameters are likely
candidates for being in such a struct. e.g. hasRecursion and
tuple_fraction. A SetOperationStmt could go in there too.

The other CTE thread about the PathKey change you worked on highlights
that something like this could be useful. I posted in [1] about this.

David

[1] https://postgr.es/m/CAApHDvrF53ErmonnpO77eDiJm7PyReZ+nD=4FSsSOmaKx6+MuQ@mail.gmail.com