model-writtenrun 7318 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
May 2021 · 2 entries
- Michael Paquier reopened the pg_upgrade TAP conversion in #1, aiming to replace test.sh and duplicated MSVC machinery while preserving cross-version testing through installation and source paths.
- Andrew Dunstan recommended historical regression dumps in #2, and Michael added an alternative olddump mode plus source-independent library-reference rewriting in #6, temporarily setting aside unfinished MSVC support.
Sep - Oct 2021 · 2 entries
- After several months, Rachel Heaton reported a failing test caused by the removed node-construction API in #7, and Michael refreshed the patch and broadened library-reference updates in #8.
- Buildfarm compatibility delayed the conversion: Tom Lane raised concerns about removing test.sh in #11 and #13, while Andrew clarified in #15 and #18 that ordinary buildfarm runs invoke make check and therefore do not require that script.
Dec 2021 - Jun 2022 · 17 entries
- Andres Freund emphasized retaining compatibility for existing buildfarm commands in #23, and Michael's revised implementation in #25 preserved upgradecheck as a TAP launcher, restored MSVC support, removed fragile directory changes, and reused upgrade_adapt.sql.
- Following January refreshes, Julien Rouhaud exposed rejection of a valid Windows socket path in #28, prompting Michael's temporary PGHOST and PGHOSTADDR workaround in #29.
- The reusable config_data() helper present in #29 landed as ba15f16107 with clearer documentation, but its latest matching submission was #498555 in the GUC flags thread thread #45167, which receives attribution for that commit.
- Andres's review in #31 called out lost test coverage and unnecessary output-path plumbing, and his tmp_check proposal in #34 led Thomas Munro to post the cleanup described in #35, which landed as 7e6124ca7d.
- Andres proposed using is_unixsock_path() for Windows and abstract socket paths in #30, and Michael separated that production fix from the revised TAP conversion in #39, also restoring small-WAL and group-access coverage and adding failure-log reporting.
- Michael concluded in #39 that running historical pg_regress binaries would create a fragile compatibility burden, so #40 removed that cross-version mode and retained loading old dumps instead.
- The unchanged server.c component of #40 landed as dc57366c58, leaving the rebased TAP conversion in #41; Michael explained that the socket-path fix was initially limited to the development branch.
- Andres corrected regression-option precedence and reviewed comments in #42 and #43, producing #45; Tom subsequently supported proceeding after the buildfarm compatibility discussion in #52.
- The TAP conversion from #45 landed as 322becb608 with clearer dump-generation documentation and small invocation fixes, replacing test.sh and most duplicated MSVC upgrade-test code.
- Noah Misch immediately reported that Solaris diff lacked the -q option and that failure logs were missing in #54; Michael reported switching dump comparison to File::Compare and improving diagnostics in #56 and #57.
- A Windows buildfarm failure with TAP disabled led Michael to revive Andres's compatibility-no-op design in #59, moving pg_upgrade testing into bincheck while keeping upgradecheck callable.
- Justin Pryzby requested corresponding CI cleanup in #60, and d2a2ce4184 landed #59 with that cleanup, removal of an unused helper, and revised messages; Michael noted remaining duplicate execution in older buildfarm clients in #61.
- Noah's detailed review in #62 found unchecked command failures, redundant connection options, incompatible EXTRA_REGRESS_OPTS parsing, and missing database-name quoting coverage.
- Michael addressed those findings in #64, which landed as 7dd3ee5084 with a small adjustment to the database-name backslashes.
- Noah questioned the changed database names in #66 and #68, prompting #69; his further review in #70 identified regression-name warnings, a missing command-line space, and failure to assert pg_regress success.
- Michael restored the historical regression-prefixed names and fixed both command-execution issues in #71, which Noah confirmed in #72 and which landed unchanged apart from whitespace as eaa5ebe046.
- The thread ends with diagnostic-output work still pending: Andres reported an almost 15 MiB failure log in #74, and Michael's final patch #76 proposed limiting printed logs and dump differences to 10,000 characters, with no landing shown.