On Sat, Dec 13, 2025 at 7:02 PM Tender Wang <tndrwang@gmail.com> wrote: > Tom Lane <tgl@sss.pgh.pa.us> 于2025年12月13日周六 00:54写道： >> PG Bug reporting form <noreply@postgresql.org> writes: >> > After upgrading to Postgres 18 I've come across an error I wasn't getting >> > beforehand. Here's a minimal way to reproduce the issue, that used to work >> > well in Postgres 12 and 17 at least.
>> Thank you for this well-crafted bug report! Bisecting shows that >> it broke at >> >> f5050f795aea67dfc40bbc429c8934e9439e22e7 is the first bad commit >> commit f5050f795aea67dfc40bbc429c8934e9439e22e7 (HEAD) >> Author: Richard Guo <rguo@postgresql.org> >> Date: Tue Sep 10 12:36:48 2024 +0900 >> >> Mark expressions nullable by grouping sets
> When there is SRF in the query, the grouping_target changes in grouping_planner when entering the following function: > split_pathtarget_at_srfs(root, grouping_target, scanjoin_target, > &grouping_targets, > &grouping_targets_contain_srfs);
Yeah, the issue happens here. In split_pathtarget_at_srfs(), if we find a subexpression that matches an expression already computed in the previous plan level, we should treat it as a Var and should not split it further. setrefs.c will later replace the expression with a Var referencing the subplan output.
However, when processing the grouping target for grouping sets, this function can fail to recognize that an expression is already computed in the scan/join phase. This happens because the comparison crosses the grouping boundary: expressions in the grouping target may carry the grouping nulling bit, while the corresponding expressions in the scan/join target do not.
The nulling bits in the grouping target should be ignored when processing across the grouping boundary.
This mismatch leads this function to incorrectly assume that the expression (e.g., unnest(markets)) needs to be re-evaluated from its arguments, which are often unavailable in the subplan.
To fix, I think we should ignore the grouping nulling bit when checking if an expression from the grouping target is available in the pre-grouping input target. This is actually what we do in setrefs.c.
Hence, attached patch.
The patch works for me. All regression tests pass with this patch.
(I'm currently on vacation and will take a closer look when I return. Anyone who wants to move this forward before then is welcome.)
