model-writtenrun 5671 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - Jun 2021 · 4 entries
- Thomas Munro proposed running the core regression suite against a primary with a streaming standby to exercise otherwise poorly covered WAL decoding and redo routines, initially skipping tablespaces and enabling WAL consistency checking in #1.
- Takayuki Tsunakawa proposed comparing primary and standby dumps and identified sequence-cache differences in #2, while Andres Freund and Tom Lane supported routine replay testing but argued against enabling expensive WAL consistency checking by default in #4, #6, and #7.
- Thomas Munro incorporated direct pg_regress invocation, opt-in consistency checking, parallel execution, and simpler assertions in #10; testing contrib and external access methods remained a possible extension rather than part of the initial implementation.
- Anastasia Lubennikova reviewed and rebased the patch in #11, then removed unnecessary authentication setup and added cleanup rules in #12, although unexplained CI failures remained.
Oct 2021 · 1 entry
- After several months, Thomas Munro expanded the work into relative tablespaces under pg_user_files, NEW LOCATION directory creation, regression-test conversion, and dump comparison in #14, replacing the earlier plan to skip tablespace tests.
Nov 2021 - Feb 2022 · 13 entries
- Windows junction problems persisted through the Perl API rebase in #15; after Andrew Dunstan explained symlink privileges, Thomas restricted relative tablespaces to a developer-only setting and removed NEW LOCATION syntax in #18.
- Andrew Dunstan explained the existing PG_REGRESS environment interface in #19, enabling Thomas Munro to make execution and output paths work on Windows and VPATH builds in #20.
- Andres Freund proposed creating directories directly under pg_tblspc in #22 and requested separate coverage of ordinary tablespaces in #25; Thomas implemented both in #26, abandoning pg_user_files and relative symlink handling.
- Andres Freund's review in #29 and Michael Paquier's comments in #30 and #31 led to simplified in-place checks, more tablespace movement tests, prepared-transaction coverage, pg_dumpall comparison, and EXTRA_REGRESS_OPTS handling in #32, while shared invocation infrastructure was deferred.
- The in-place backend support and regression-test conversion from #32 landed as 7170f2159f and d6d317dbf6, with added configuration documentation and removal of an obsolete cleanup rule.
- The same series supplied the unchanged standalone tablespace test in d1511fe1b0 and the replay test in 123828a7fa, whose final edits quoted shell paths and polished documentation.
- Immediate buildfarm failures prompted dccee0f2b7 to revert the replay test and 0c53a6658e to revert the standalone tablespace test; Thomas reported incompatible buildfarm invocation and Windows paths in #33 and #34.
- Andres Freund identified that buildfarm EXTRA_REGRESS_OPTS overrode the replay test's chosen port in #35, and Thomas relanded that component of #32 with reordered arguments as f47ed79cc8, confirming local buildfarm-client testing in #36.
- Noah Misch diagnosed the MSYS path translation problem in #37; Thomas relanded the tablespace component of #32 using perl2host and removal of internal dot components as cfe7bd17e4, explaining the corrections in #38.
- The broader replay workload exposed further failures: Andres Freund reported tablespaces remaining nonempty in #40, and Thomas linked that to separate open-file deletion work in #41, then reported standby dump cancellation and existing statistics-test instability in #43.
- Andrew Dunstan reported excessive Windows runtime in #44; Andres Freund traced it to repeated failed localhost connection attempts in #47 and committed the explicit-host patch from #51 unchanged as 7340aceed7.
- The accompanying Windows hang investigation was separated from the replay-test work when Andres Freund connected it to an existing socket-shutdown problem in #54, with Thomas confirming work on that problem elsewhere in #55.
- After another standby dump cancellation, Thomas rejected hot_standby_feedback because it altered test results and proposed an unlimited replay-conflict delay in #57; Tom Lane requested a finite ten-minute delay in #63, which landed in 4d7c3e3447.
Mar - Apr 2022 · 4 entries
- A vacuum failure whose details were missing from buildfarm output prompted Thomas's logging patch in #66; Andrew Dunstan accepted it in #67, and it landed unchanged as a096813b6f.
- The new diagnostics let Masahiko Sawada report a concrete vacuum truncation mismatch in #69; Thomas's temporary-table and DISABLE_PAGE_SKIPPING fix in #70 landed unchanged as 383f222119 and was backpatched through PostgreSQL 12.
- After that vacuum fix had landed, Peter Geoghegan proposed fault-injection testing and VACUUM FREEZE in #71 and #73, but discussion with Andres Freund and Thomas Munro concluded that other backends holding back OldestXmin could still require a temporary table in #74 and #76.
- Thomas reported further small-buffer nondeterminism in #77, and Tom Lane identified unexpected synchronized scans in #78, then disabled synchronize_seqscans directly in cbf4177f2c and confirmed the change in #79.
Aug 2022 · 1 entry
- By the end of the thread the main series and subsequent fixes had landed, but test scheduling remained open: Justin Pryzby proposed starting the long replay test earlier in #80, Thomas considered merging it into the first streaming test in #81, and Justin supplied parallel-run timing improvements in #82 without a further attached patch or landing.