Re: Fix HAVING-to-WHERE pushdown with nondeterministic collations

On Wed, Apr 1, 2026 at 11:19 AM wenhui qiu <qiuwenhuifx(at)gmail(dot)com> wrote:
> > + vars = pull_var_clause(node, PVC_RECURSE_PLACEHOLDERS);
> > +
> > + foreach_node(Var, var, vars)
> > + {
> > + if (var->varno == *group_rtindex &&
> > + OidIsValid(var->varcollid) &&
> > + var->varcollid != inputcollid &&
> > + !get_collation_isdeterministic(var->varcollid))
> > + {
> > + list_free(vars);
> > + return true;
> > + }
> > + }
> > +
> > + list_free(vars);

> This might be overthinking, but I wonder if calling pull_var_clause() at each walker step could introduce some overhead due to repeated subtree scans

That's a good point, but I doubt that it'd be an issue in practice.
HAVING clauses are typically very small expressions. Even in unusual
queries, the clause size is bounded by what a human writes, which is
negligible compared to the work the planner does elsewhere.

Maybe we can avoid this by calling pull_var_clause once at the top of
each clause and reusing that var list at every node. But that can
introduce false positives. The pre-pulled list contains all GROUP
Vars from the entire clause, but a given operator node only acts on
the vars in its own subtree.

- Richard