model-writtenrun 9434 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Rajkumar Raghuwanshi reported incorrect statistical aggregate results with partitionwise aggregation in #1, and Jeevan Chalke reproduced the failure with parallel partial aggregation and traced its visibility to variable-length function-call storage in #2.
- Kyotaro Horiguchi identified a strictness check reading a nonexistent combine-function argument and supplied a proof-of-concept decrement in #4, then moved the correction into aggregate-state initialization in #5.
- Robert Haas and Tomas Vondra treated the failure as a release open item in #7 and #8, despite evidence that the underlying argument-count error predated the development release.
- Greg Stark's question about Valgrind coverage in #9 led Kyotaro to demonstrate an allocation-boundary blind spot with a one-byte allocation adjustment in #11, followed by allocator-wide sentinel protection in #12 at the cost of increased memory use; neither diagnostic patch was included in the aggregate fix.
- Andres Freund challenged the initialization approach in #13, proposing either correct arguments from ExecInitAgg or separate count initialization in the combine branch, and concluded that PostgreSQL 11 needed a fix while older releases did not use the faulty count for combining.
- David Rowley supplied #16 with an explicit combine-input count, argument-name cleanup and a regr_count regression test, explaining that the earlier numInputs-based correction only happened to produce the right value.
- Andres questioned the argument-type plumbing and preferred a minimal fix before beta1 in #17; David clarified that the regression query exercised partial and final aggregation in #18.
- Andres committed the minimal portion of #16 as 2657283256, hardcoding the combine-input count in the helper and retaining the regression test with expanded EXPLAIN output; the change was backpatched to PostgreSQL 11 and credited David, Kyotaro and Andres as authors.
- After confirming the commit, Andres proposed moving combine-specific setup out of build_pertrans_for_aggref in #19, and David posted the deferred cleanup in #20, supplying combine-state argument types explicitly and removing the separate combine-expression builder.
- Kyotaro requested clearer argument naming and explanations of final-function and ordered-set behavior in #21, David explained those distinctions in #22, and Robert and Andres continued discussing transition-versus-combine terminology in #23 and #24.
- When Andres revisited the cleanup in July in #25, David said it remained low priority and outlined broader planner-side aggregate work for separate threads in #26; their final exchange in #27 and #28 left sorting strategy and planner responsibility open, with the correctness fix landed but the cleanup still pending.