model-writtenrun 5670 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar - Apr 2022 ยท 15 entries
- Mark Dilger opened the thread by extracting Joshua Brindle's string-based object access hooks from a dependent GUC-privileges patch and adding regression coverage for hook ordering and permission checks; he also reported an apparently separate crash-restart test failure #1.
- Joshua reviewed and tested both patches, accepting the design despite reservations about using subId for privilege information, while Andrew Dunstan accepted using InvalidOid where no catalog existed rather than introducing another family of hook functions #2 #5.
- Andrew's review led Mark to remove unused test-module packaging, expand its documentation, and distinguish Oid-based and name-based entrypoints in function comments across the next two submissions #4 #6 #8 #9.
- Andrew committed the hook constituent unchanged as d11e84ea46 and the test constituent with redundant loading and message-setting commands removed as 90efa2f556, completing the original series from #9 and allowing the dependent GUC work to proceed #13.
- Julien Rouhaud reported a buildfarm failure, Mark supplied preload-related startup errors, and Tom Lane reproduced duplicate audit notices with force_parallel_mode and startup failures with EXEC_BACKEND #14 #15 #17 #22.
- Mark initially proposed deleting ExecutorCheckPerms instrumentation, but Tom suggested retaining it and suppressing parallel-worker notices instead; the replacement patch also removed the preload restriction and switched the test to explicit LOAD #20 #21 #23.
- After initially questioning the loading change, Tom agreed that the module should support shared, per-session, and explicit loading, and Andrew committed the replacement patch unchanged as f0206d9952, avoiding a revert of the test module #27 #28.
- Tom then tested installcheck successfully and requested removal of its restriction plus comment and spelling fixes, which Andrew reported completing #29 #30.
- The newly enabled tests exposed encoding-dependent namespace-search notices: Tom traced them to cached search-path construction, and Justin Pryzby supplied corresponding Windows failures #31 #36 #37.
- Andres Freund proposed explicitly testing parallel execution rather than suppressing worker notices, but discussion identified worker-count and message-ordering complications, and no replacement parallel test was submitted in this thread #32 #33 #34 #35.
- Andrew temporarily restored the installcheck restriction in 5b29a9f772 to quiet locale-dependent buildfarm failures, while Tom warned that make check could fail too #38 #39 #40.
- Andrew then forced UTF8 and NO_LOCALE in 7c51b7f7cc; subsequent Windows Meson testing also needed these settings, which Tom emphasized were a workaround for misplaced namespace hooks #41 #42 #44 #45.
- Mark later reproduced the separate crash-restart failure and supplied logs; Tom identified an obsolete SIGQUIT warning expectation and an unhandled frontend error, while Mark examined additional libpq failure paths and endorsed less restrictive matching #46 #53 #55 #56 #57 #59.
- Tom committed the discussion-derived fix as 7a43a1fc52, updating the SIGQUIT warning and accepting could-not-send errors in four crash-test checks, with a backport to PostgreSQL 14 #60.
- The namespace-hook placement issue remained open: Michael Paquier warned about compatibility risks and cache-filtering complexity, then recommended postponing redesign near beta2 and considering suppression of namespace-search test output to restore broader installcheck coverage #61 #62.