model-writtenrun 8870 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Adam Scott reported a PostgreSQL 10.11 crash while planning numrange containment queries in #1, then supplied a data dump after finding that reproduction depended on session state in #5.
- Michael Paquier reduced the failure to a small analyzed numrange table in #7, and Tomas Vondra reproduced it on PostgreSQL 10 in #11; subsequent investigation established that compiler and statistics-memory changes affected reproducibility rather than introducing the underlying bug #12 #14.
- Michael identified an access beyond the histogram array and posted a tentative loop-index fix with a regression test in #14, while questioning whether always skipping the first examined bin was correct.
- Andrey Borodin found that the proposed test did not fail on his machine and suggested capping the starting index in #15, then supplied a revised patch that instead guarded the out-of-range get_position() call in #16.
- Tom Lane argued that values beyond the histogram maximum should not form an additional bin and posted explicit endpoint handling for both containment estimators in #17, also adding histogram-size and NaN checks and omitting the unreliable test.
- Michael reviewed the new code and challenged its comments and test coverage in #18; Tom explained why uninitialized memory made a reliable crash test impractical and declined to merge the two estimator functions as part of this bug fix in #19, which Michael accepted in #20.
- Tom committed 652686a334 from #17, adding a guard against negative subdiff results and refining comments, and announced the push in #21; the fix was backpatched to all supported branches without a regression test.
- Adam reported successful testing on his server after the commit in #22, and Michael closed the thread by confirming that the committed fix would reach him in PostgreSQL 10.12 and that the omitted test would be retained separately #23.