model-writtenrun 9245 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Rui Hai Jiang reported that a premature break prevented TopoSort() from marking additional members of a lock group and submitted its removal in #1.
- Robert Haas requested a reproducing test in #5, and Rui agreed to investigate a parallel-process scenario in #6.
- After the test remained unavailable, Tom Lane argued that the obvious fix should land before the upcoming minor releases without waiting for a reproducer in #9.
- Tom then supplied an advisory-lock isolation test in #13, disabling a parallel-mode guard to force workers into the problematic wait queues and confirming that the existing code hung while removing the break fixed it.
- Tom committed Rui's fix as 3420851a2c, adding an assertion and explanatory comments and backpatching through PostgreSQL 9.6.
- Robert proposed alternative ways for workers to acquire locks in #14 and #16, while Tom explained the LOCK TABLE restrictions and difficulties of constructing a reliable catalog-lock test in #15.
- Tom proposed removing the redundant advisory-lock runtime guard and relying on parallel-safety markings in #20; Robert examined the resulting same-group lock behavior in #21.
- Tom rejected parallel-safe labeling because worker-held advisory locks disappear before the documented transaction or session boundary, settling on parallel-restricted labeling in #22.
- Robert asked whether removing the guard could expose crashes in #23, Tom explained why the lock-manager behavior was already reachable in #26, and Robert accepted that reasoning in the thread's final message, #27.
- The agreed follow-up work landed on August 1: 4886da8327 changed advisory-lock catalog markings to parallel restricted, and da9456d22a polished the test from #13, installed it separately, and removed the runtime guard, completing all three changes.