model-writtenrun 6907 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - Jul 2026 ยท 34 entries
- Andres Freund opened the thread by announcing Cirrus CI's June shutdown and asking which replacement properties mattered, including self-hosting, speed, platform coverage, and CI available without posting patches to the mailing list #1.
- Jelte Fennema-Nio prioritized accessible, fast Linux/macOS/Windows testing, while Peter Eisentraut and Heikki Linnakangas emphasized longer-term self-hosting and local execution; Peter supported an established hosted service as a short-term bridge #3 #7 #9.
- David E. Wheeler suggested Woodpecker, and Thomas Munro outlined reusable QEMU images and a more independent cfbot implementation, but neither alternative became the immediate migration patch #11 #12.
- Robert Haas questioned repeated full CI runs for unchanged small patches, and Tom Lane distinguished valuable immediate submission tests from periodic retesting that could be reduced #13 #15.
- With the shutdown approaching, Jelte posted an AI-assisted GitHub Actions prototype that he had only cursorily reviewed; Nazir Bilal Yavuz posted his independently developed workflow with reusable provider helpers and offered to combine the implementations #17 #18.
- Following the developer unconference, Nazir merged the prototypes around Jelte's Linux containers, restored io_uring testing, and deferred BSD jobs, public logs, and provider-helper extraction, with eventual backpatching to PostgreSQL 15 proposed #19.
- Andres requested less duplication and safer cancellation, Zsolt Parragi found a branch filter that missed names containing slashes, and Peter tested dependencies and supplied corrections; Nazir responded with YAML anchors, a Linux matrix, cached MacPorts provisioning, and a less collision-prone workflow filename #21 #24 #25 #31.
- Jacob Champion's PATH and third-party-action concerns led Nazir to revise Windows setup and relocate MSYS2 to the faster drive; Peter's subsequent shell and spelling patches were folded in, while Andres's discovery that Windows crash logs were never generated became a deferred repair #37 #38 #43 #44 #46.
- Alvaro Herrera and Jakub Wartak reported slow runs, and Andres's measurements shifted the Linux diagnosis from primarily disk I/O toward the smaller number of physical CPU cores available on the new runners #34 #39 #42.
- Andres took over the main restructuring, replacing the Linux matrix with shared anchors and environment updates, moving containers to GitHub's registry, improving cancellation and caches, and attaching Cirrus removal plus temporary segment-size and reduced-test experiments #50.
- Jakub reported successful tests, Thomas connected the new workflow to cfbot, and Nazir supplied stable-branch cancellation fixes while identifying additional Cirrus cleanup; Andres then devised unique concurrency groups so separately pushed stable commits would neither cancel nor block each other #51 #52 #53 #56.
- The next consolidated series adopted pg-ci.yml, moved ASan to Linux Meson, rewrote the CI README, and completed Cirrus cleanup; the segment-size experiment was not ready to commit, and the reduced-test patch had only supported faster iteration #61 #64.
- Peter supplied missing coverage formerly provided by FreeBSD, while Andres repaired Windows dependency installation, error handling, and missing logs; they agreed to place the extra tests on faster Linux jobs rather than concentrating them on macOS #65 #69 #71 #73.
- Andres incorporated that coverage and split Windows Visual Studio tests across two runners, reducing their reported duration from roughly 33 to 23 minutes; pull-request triggers and reusable workflow calls were deferred until basic CI was restored #70 #75.
- Jacob continued pressing for explicit downstream consent, and joint testing confirmed that existing forks could automatically acquire enabled workflows; Andres added PG_CI_ENABLED gating and an explanatory notification, then refined its wording with Nazir and Peter #77 #78 #79 #80 #83 #85.
- Andres added partial MacPorts cache reuse and early cache saving, Nazir reviewed it, and Jacob agreed that unfinished CPAN caching should not delay the initial migration #87 #89 #91.
- The workflow and MacPorts changes from #87 landed together as 9c126063b1, with only a robocopy threading-option removal and README corrections, while the unchanged removal attachment landed separately as 68c8a365d4; Andres announced the push in #92.
- Nazir then found that simultaneous failures in the two Windows slices produced identically named artifacts and posted a small patch adding the slice number to artifact names; none of the supplied candidate commits represents that fix #94.
- Jakub continued investigating Windows performance with ReFS experiments, typeperf counters, and then RAM-disk and per-process measurements, but reported no useful RAM-disk speedup and left the slow-system-drive investigation unfinished #95 #97 #104.
- Andres identified restored Windows crash reporting and faster cold-cache CompilerWarnings builds as priorities before backpatching, alongside possible improvements to runner selection, concurrency, container granularity, and failure-log access #99 #100.
- The MSYS2 discussion returned when larger runners lacked the assumed drive layout; Jacob clarified that his earlier request was to pin the action rather than prohibit it, and Andres ultimately used a pinned setup-msys2 action in the follow-up series #101 #102 #106 #108.
- Andres's new nine-patch series included Nazir's Windows debugger setup, Meson-based compiler-warning builds, smaller and freshly installed MSYS2 environments, optimized MinGW builds, configurable runners, automatic concurrency, and job-specific containers; the optional pacboy simplification remained explicitly discretionary #108.
- Nazir reviewed that series and found it satisfactory except for a question about feature coverage lost when CompilerWarnings switched from Autoconf options to the common Meson feature set #111.
- Tatsuo Ishii separately reported sanitizer instrumentation corrupting LLVM JIT bitcode, relayed an inline fix credited to Matheus Alcantara, and reported successful testing; Thomas Munro agreed to investigate, but the supplied candidates do not establish its resolution #112 #113.
- Jacob supplied CPAN installation caching and an MSYS2 environment cleanup as explicit additions to the follow-up series, explaining that caching installed Perl modules avoided repeated CPAN rebuilds #114.
- Henson Choi reported that cfbot could mistake a non-patch archive for a submission and test plain master instead of the real patch, and Thomas acknowledged the need to improve attachment handling #115 #116.
- On July 17, the MinGW optimization attachment from #108 landed unchanged apart from rebasing as 47a20d872b, and Nazir's crash-log attachment landed with identical code and polished explanatory text as d374280e18; although the latter also matches the June 10 submission #1006122 in thread #139919, this thread's June 12 repost is the latest matching patchset.
- The last thread reply clarified that easier log access should expose the first failed test per job, using a separately uploaded artifact and a readable link; no implementation was attached #118.
- Automatic concurrency originally proposed in #108 subsequently landed on November 6 as 8736129a48, but its exact final attachment was posted in #1160722 in thread #254549, so that follow-up thread owns the landing.
- The smaller MSYS2 installation originally attached in #108 landed as f8cabd1f17, preserving tools through PATH inheritance and dropping unnecessary packages; its later matching repost in #1160722 assigns the landing to thread #254549.
- The fresh MSYS2 installation and pacboy changes from #108, together with Jacob's environment cleanup from #114, were incorporated into the final attachment in #1160722 in thread #254549 and landed as 510512876f, including the updated action and default temporary installation location.
- Configurable runners originally proposed in #108 landed as d073ddd518 from the exact final attachment in #1160722 in thread #254549, which included simpler fallback expressions, shared setup-runner configuration, architecture-specific MacPorts caches, and RUNNER_TEMP socket directories.
- Per-job container naming originally proposed in #108 landed as 29fffc1a66 from the exact final attachment in #1160722 in thread #254549, using a shared image-format template and omitting the earlier proposed shared container options for CompilerWarnings.
- The initial migration, MinGW optimization, and Windows crash reporting remain attributed to this thread, while the five November follow-up landings are attributed to the later matching series in thread #254549 and were also backpatched to PostgreSQL 19; Jacob's CPAN caching and the CompilerWarnings conversion have no separately verified landing among the supplied candidates, and the CompilerWarnings coverage question and first-failure log presentation remain unresolved in this thread #108 #111 #114 #118.