model-writtenrun 7785 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andreas Seltenreich reported a sqlsmith-triggered partition-pruning assertion failure in #1, and Tom Lane reduced it to a query without a window aggregate in #2.
- After initially questioning the assertion in #3, Tom traced the failure to ordered Append path construction omitting an intermediate partition from partitioned_rels in #4.
- Zhihong Yu proposed returning NIL instead of asserting in #5, but Tom rejected masking the upstream problem and suggested reconstructing partition ancestry instead in #6.
- Tom posted two patches in #7: discover the necessary ancestors through AppendRelInfo links at plan creation, bounded by the Append relation and its subpaths, then remove partitioned_rels and its bookkeeping; he reported that this fixed both reported cases and deferred adding a regression test.
- David Rowley identified singleton Append flattening as the precise source of the lost ancestor in #8, favored removing the fields over his local repair, and explained that the bug did not require backpatching.
- Amit Langote expressed support for the pruning patch in #9, and David supplied a revised version in #10 that used bitmapsets to avoid list insertion and duplicate-elimination costs.
- Tom questioned reliance on RT-index ordering and future support for partitionwise joins in #11 and #12, while David explained in #13 that pruning joins would already require substantial changes elsewhere.
- Tom accepted the bitmapset representation and committed #10 with naming, comment and minor defensive changes as fb2d645dd5, without backpatching or adding a regression test.
- Tom split the cleanup from #7 into 5076f88bc9, removing incidental dependencies, and f003a7522b, removing the obsolete fields and bookkeeping.
- Tom announced the completed commits and explained his preference for Bitmapset terminology in #14; David's closing thanks in #15 left the series fully landed with no outstanding patch.