model-writtenrun 710 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Bill Kim proposed making btree_gist’s float comparisons follow PostgreSQL’s NaN ordering, reporting passing tests and explaining that inconsistent comparisons caused missed rows, duplicate NaNs under exclusion constraints and an RLS-related wrong result #1 #3.
- After Daniel Gustafsson encouraged submission and Tom Lane pointed out the missing attachment, Bill supplied the comparison fix and regression tests #2 #4 #5.
- Tom initially found the fix reasonable but questioned backpatching because affected indexes would need rebuilding; Bill favored backpatching because of silent wrong answers and exclusion-constraint bypasses #6 #7.
- Tom’s deeper review found additional NaN and infinity problems in penalty and distance calculations, so he posted a substantially expanded revision with separate floating-point penalty logic, revised distance semantics and multicolumn tests that exercised previously uncovered penalty functions; he deferred changing generic gistpenalty behavior and leaned toward backpatching with a reindex notice #8.
- Bill confirmed that the revision applied cleanly and passed btree_gist regression tests, but his authorship excludes him from the suggested reviewer list #9.
- The revised patch landed unchanged apart from base-context and whitespace differences as 255bce4488, with backpatching through PostgreSQL 14 and advice to rebuild indexes containing NaNs; Tom’s final message confirmed it was pushed, leaving no pending patch in the thread #10.