model-writtenrun 9097 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Антон Власов reported PostgreSQL 12 crashing during concurrent updates in #1; after ruling out JIT in #4, he isolated a BEFORE UPDATE trigger that merely returned NEW and supplied a two-session reproducer in #9 and #10.
- Andres Freund diagnosed ExecBRUpdateTriggers copying a tuple slot into itself after EvalPlanQual, corrupting the slot, and proposed a caller-side guard, an assertion, clearer variable names, and better tests in #11.
- Tom Lane proposed making ExecCopySlot silently tolerate self-copying in #12, but Andres argued that this could conceal callers incorrectly expecting independent slots in #13, and Tom accepted the assertion approach in #14.
- Andres posted the crash fix, assertion, and variable cleanup in #15, then retained that implementation and added extensive isolation tests in #16; foreign-table coverage remained outside the submission because it required separate contrib tests.
- The two patches in #16 landed with identical diffs as d986d4e87f for the crash fix and c884119950 for the tests; Andres announced the fix for PostgreSQL 12.1 and test backpatches through 9.6 in #17.
- Buildfarm failures revealed nondeterministic ordering in the three-session tests, prompting Andres to propose disabling or rewriting them in #18; Peter Geoghegan explained that speculative insertions lacked heap-style lock arbitration and questioned the proposed fairness requirement in #19.
- Andres disabled the two UPSERT permutations in 6e61d75f52 and explained in #20 that speculative insertion was not actually involved: he suspected an unordered transaction-lock wait during uniqueness checking, a broader issue outside the trigger crash fix.
- A second follow-up, 3a68105154, disabled two more three-session UPDATE permutations; after Tom reported continuing buildfarm failures in #21, Andres confirmed this additional change and said he had seen no related failures since in #22.
- At the thread's close in #22, the crash fix and remaining tests were committed, four unstable permutations were disabled, and the broader question of deterministic ordering among tuple-lock waiters remained unresolved.