model-writtenrun 2176 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut proposed simplifying objectNamesToOids() through get_object_address(), gaining object locking while retaining special handling for relations, types, and parameters; he requested scrutiny of the changed isolation-test output and lock strength in #1.
- Bertrand Drouvot endorsed the refactoring and verified that locking moved the concurrent-drop error into relation lookup, producing a more appropriate user-facing error; his review also questioned the unexplained NULL-relation assertion in #2.
- Peter proposed moving that assertion into get_object_address() in #3, and Bertrand supplied the cleanup in #4; it landed as e468ec0fdd, allowing callers to pass NULL when no relation output is expected.
- The original locking patch from #1 landed as d31bbfb659 with a small adjustment to use the new NULL-output convention, and Peter confirmed both commits in #5.
- After the landing, Noah Misch explained that AccessShareLock still permits concurrent catalog updates and that eliminating those races would require self-exclusive locking across all relevant update paths, with autovacuum and lock-table costs; he also identified lost SearchSysCacheLocked1() test coverage in #6.
- Peter accepted documenting the remaining concurrency limitation in #7, while Noah outlined two ways to recover coverage, including replacing DROP TABLE with direct deletion from pg_class, in #8.
- Returning to the PostgreSQL 18 open item, Peter submitted a comment explaining the locking tradeoff and a test patch implementing direct catalog deletion in #9, while also discussing the alternative of allowing unlocked get_object_address() calls.
- Noah mildly preferred the unlocked alternative but accepted either direction in #10; Nathan Bossart similarly questioned the extra lock acquisitions while suggesting broader reconsideration could wait until PostgreSQL 19 in #11.
- Heikki Linnakangas endorsed the comment and restoring test coverage in #12, suggested considering a separate test for the artificial catalog deletion, and removed the issue from the release-blocking open-items list.
- The two patches from #9 landed separately with unchanged diffs as 118601a7e8 for the comment and baf45ba053 for test coverage, both crediting Noah as co-author; Peter confirmed completion in #13, leaving AccessShareLock in place.