model-writtenrun 7386 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jan - Mar 2021 · 2 entries
- Dilip Kumar reported a standby becoming stuck on an old timeline after another standby was promoted: the new history file reached the archive before its WAL, leaving recoveryTargetTLI and expectedTLEs inconsistent #1.
- Robert Haas identified the violated expectedTLEs invariant, and Dilip proposed initializing it from recoveryTargetTLI in an inline one-line fix, while investigating why the older code used receiveTLI #2 #3 #4.
May - Jun 2021 · 23 entries
- After the investigation resumed in May, Dilip attached a different fix that rebuilt expectedTLEs after checkpoint reading, while Kyotaro Horiguchi supplied diagnostic instrumentation and initially could not reproduce the failure #5 #6.
- Dilip clarified that the failing standby had no local WAL, allowing Kyotaro to reproduce the problem and replace post-checkpoint rebuilding with conditional timeline selection inside WaitForWALToBecomeAvailable #9 #10.
- Robert argued that expectedTLEs should always follow its documented recoveryTargetTLI invariant, leading Kyotaro to simplify the fix to that substitution while retaining delayed history loading #14 #16.
- Further investigation separated the historical need to wait for streamed history files from the incorrect choice of timeline, and Kyotaro supplied diagnostic checkpoint suspension and then a revised patch with an additional compatibility check #27 #31 #32.
- Robert constructed a reproducible scenario using an archive command that supplied history files but withheld WAL, and Dilip converted that shell reproducer into a new standalone TAP test #37 #39 #42.
- Robert's review moved the archive helper into a checked-in file and exposed a test race: the cascading standby could start before the history file was archived, so Dilip added an archiving wait and renamed the test #46 #47 #48 #54.
- Robert and Kyotaro settled on a separate test using a backup without WAL and a deliberately selective archive command, avoiding dependence on WAL-generation timing or manual removal of an existing standby's WAL #58 #60 #66.
- A related report from Tatsuro Yamada showed duplicate archiving failures with archive_mode=always, prompting Kyotaro to develop a separate file-comparison proposal and then add archive-status tests #59 #69 #71.
- Robert cleaned up the recovery test's restoration commands and helper, and Dilip fixed another race by waiting for the promoted server's insert LSN rather than its old replay LSN #70 #72.
- Kyotaro's Linux and Windows testing found that the helper needed an explicit Perl interpreter and that the posted patch still omitted the actual recovery fix; Robert restored that fix and continued addressing Windows path escaping #73 #75 #76.
- The duplicate-archiving proposal was moved to a separately launched discussion, leaving its proof-of-concept patches uncommitted in the supplied record #83.
- Robert explained that StartupXLOG already performs the timeline compatibility check Kyotaro proposed adding, and Kyotaro accepted leaving the redundant check out of the final recovery fix #84 #87.
- Backport testing exposed older SQL function names, asynchronous promotion, and missing PostgresNode.pm helpers; Robert combined promotion polling and portability cleanup with a prerequisite backport for 9.6 #86 #92 #93 #94.
- Dilip found that older branches also needed an explicit latest-timeline setting and backup_options support, then supplied working branch variants and a 9.6 workaround that removes backup WAL because that version lacks -Xnone #95 #96 #97.
- Robert committed the backup_options portion of the final bundle as 99a0a2ada8, extending the prerequisite backport through v13 #97 #98.
- Robert separately committed the 9.6 lsn and wait_for_catchup prerequisite from the final bundle as 75212a854f, then landed the recovery fix and standalone test with small adaptations as caba8f0d43 #97 #99.
- Tom Lane reported an immediate buildfarm failure, and Kyotaro diagnosed missing WAL retention and supplied branch-specific one-line fixes #100 #101.
- Those retention patches landed unchanged as 2208d71a00 for wal_keep_segments and 4dcb1d087a for wal_keep_size #101 #102.
- Further buildfarm investigation separated intermittent conchuela failures from jacana's Windows archive-command problem, and Andrew Dunstan reported pushing a portability fix tested on a Windows replica #104 #108 #110 #113.
- Tom questioned an accompanying coalesce change in the archiving wait query, and Andrew later reported reverting that incidental change while retaining the portability work #114 #122.
- Robert's code review requested explanations for Perl path handling and PERL_BADLANG, and Andrew confirmed that perl2host accepts a complete filename and posted a simpler inline patch with explanatory comments #121 #125 #126 #127.
- Andrew's inline cleanup landed as 54a5ed2201, correcting the inline typo and refining the locale comment #127.
- By the thread's end the recovery fix, prerequisites, WAL-retention fixes and portability refinement had landed, but Kyotaro's final message reported reproducing jacana's error after obtaining a working MSYS TAP setup, leaving that test-environment investigation without a final resolution in the supplied discussion #131.