model-writtenrun 5641 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct 2021 - Jan 2022 · 17 entries
- Andres Freund proposed in-tree Cirrus CI for testing development branches on FreeBSD, Linux, macOS, and Windows, choosing Cirrus for existing cfbot integration, custom virtual machines, platform coverage, and publicly accessible results in #1.
- Thomas Munro and Tom Lane supported maintained CI configurations alongside possible future providers, while Tom requested usage documentation and separated the production implications of Windows crash handling from CI configuration in #4, #5, and #15; a plain-text README was chosen in #17.
- Daniel Gustafsson requested broader OpenSSL and SSL-test coverage in #6, and the resulting Windows testing exposed broken slurp_file offset handling in #19.
- Andres replaced the faulty Windows seek and cleanup operations in #20, Andrew Dunstan reviewed the fix in #22, and 2f74db1236 landed with only an additional unused-import removal.
- The revised CI submission added optional-feature coverage and documentation in #32, and Melanie Plageman supplied a tested YAML anchors-and-aliases fixup in #33 that was incorporated into #36.
- Andres expanded testing to documentation builds, MinGW cross-compilation, dtrace configurations, and Windows Unix sockets in #36, while explicitly leaving the Windows crash-handling patch unready for integration.
- Thomas Munro reported successful use and suggested shared image infrastructure and reusable helpers in #41, while Peter Eisentraut challenged parallelism settings, image provisioning, duplicated configuration, and manual Windows test registration in #47.
- Andres retained a non-LLVM FreeBSD configuration because of image and startup costs in #42, then moved support into src/tools/ci, extracted core-backtrace handling, and switched to shared prebuilt containers in #55.
- The final CI submission #58 omitted the unready Windows crash patch and refined compiler caching and Linux UUID configuration; Justin Pryzby's proofreading in #59 supplied the light edits preceding 93d9734946.
- Justin discovered that Windows CI missed failures because vcregress skipped NO_INSTALLCHECK modules in #63; Andres rejected duplicated module configuration in #64, leading to the temporary-instance approach attached in #66.
- Andres restored Windows compiler summaries and prevented line wrapping through the inline proposal #67, committed as 8a1ce5e54f, and separately took the initdb --no-sync patch from #66 as dfc0cb3940.
- Justin expanded the follow-up series with macOS system information, shared configure options, Linux log capture, and initdb options in #69; Andres questioned unnecessary Linux script wrapping and suggested checksum coverage through existing regression runs in #70.
- Andrew Dunstan requested a Windows equivalent of check-world in #71, while Tom Lane urged a design compatible with Meson in #79 and Andres described discoverable tests, concurrent execution, and separate logs in #80.
- Peter Eisentraut requested discovery of TAP tests anywhere in the tree in #82, and Andrew's concerns about optional security-sensitive tests led to Andres's proposal to move PG_TEST_EXTRA checks into the tests themselves in #83, #84, and #85.
- Tom Lane traced slow and failing recovery tests to unnecessary fsync calls in #88, Andres confirmed the diagnosis in #89, and a7f4171071 removed the fsync override.
- The same investigation exposed a postmaster left running after pg_ctl failure in #91; Tom's attached Cluster.pm fix #92 landed as 6c4a8903b9 with the END-block comment clarification requested by Andres in #93.
- Andres also found pg_basebackup synchronization despite --no-sync in #89, supplied the rename fixes in #95, and committed them unchanged apart from rebasing as 9c86d9337e.
Feb - Mar 2022 · 12 entries
- Justin's subsequent branch update #97 supplied a ccache-variable correction committed as 79b79dd57a, while the macOS sysinfo patch previously attached in #69 landed as 0a156058aa.
- Andres used Justin's artifact examples and glob feedback to propose one consolidated log-artifact group in #98, implemented by 16eb8231d1.
- Justin caught a missing wildcard introduced by the artifact consolidation in #106, and Andres immediately corrected it in f79c94f39d, confirming the fix in #107.
- Andres proposed an independent documentation task and challenged comparison against master for every branch in #105; Justin then attached package-install hints, Windows warning checks, and changed-document artifacts in #108, expanding the series with coverage and parallel TAP experiments in #118.
- Robert Haas and Tom Lane questioned fixed build defaults and requested configurability in #110 and #111, while Andres defended a common minimum configuration and explained personal-repository overrides in #112 and #114.
- Parallel TAP experiments exposed test-directory and log-placement problems, leading to Andres's Utils.pm proposal in #119, Justin's inline implementation in #121, and a decision to move that infrastructure work to another thread in #122.
- Justin resent the remaining CI enhancements separately in #127; Andres's review in #128 and #129 required separate documentation-task setup and shell helpers, rejected routine Windows binary uploads because of their size and limited use, and left vcregress work outside his planned commits.
- Justin initially restricted artifact comparisons to cfbot commits in #134, then restored a missing helper, added an HTML index, and introduced configurable BASE_COMMIT defaulting to HEAD~1 in #135; comparison-base selection and reuse of unchanged documentation remained unsettled.
- Andres accepted Unix -Og optimization but objected to MSVC optimization because it degraded backtraces in #142, Thomas Munro measured a substantial FreeBSD speedup in #144, and Justin removed the Windows optimization and refined package setup in #149.
- The package-install hints and Unix optimization from #149 landed as 225fb558cd and 4a288a37f9, with only noninteractive apt-get examples added by the committer, as confirmed in #150.
- Windows warning detection still awaited committer time and minor polishing in #152, and Thomas reported new LLVM-header warnings caused by -Og in CXXFLAGS in #153, with no resolving patch shown in this thread.
- Andres enabled zstd for FreeBSD, Linux, and macOS in #155, committed unchanged as ff50baec65; Windows remained excluded because its image lacked zstd.
Sep 2022 · 1 entry
- The thread ends with the original CI integration and selected follow-ups landed, documentation and coverage artifacts still unresolved here, and Thomas's final update #156 noting that GNU make had fixed the output-locking issue but package availability would lag the planned Meson/Ninja transition.