model-writtenrun 7312 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - Sep 2019 · 7 entries
- Paul Guo opened the thread with a standby crash-recovery failure after CREATE DATABASE, DROP DATABASE, and DROP TABLESPACE, proposing recursive reconstruction of missing directories in #1.
- Asim preferred tracking invalid references rather than reconstructing objects in #2, while Kyotaro proposed ignoring directory-creation failures only before consistency in #4 and supplied tablespace-aware backup support and a regression test in #6.
- Paul identified the additional missing-source case caused by moving a database between tablespaces in #12, then moved the checks into dbase_redo() and added WAL-description changes in #13.
- Kyotaro warned that missing directories must remain fatal after consistency and that forced restartpoints could slow recovery in #14 and #16; Paul consequently added remembered missing directories and consistency-time validation in #18.
- After compatibility updates, Anastasia demonstrated that #22 could silently tolerate lost tablespace storage during ordinary standby operation, establishing the need to distinguish pre-consistency recovery in #23.
- Asim took over while Paul was busy and posted a restartpoint-based alternative in #29, but returned to tracking missing objects after the discussion favored that approach.
- Asim found duplicate missing-directory entries could cause a false PANIC and supplied expanded tests, including Alexandra's archive-recovery scenario, in #30; his replacement in #31 used a hash keyed by tablespace and database IDs.
Dec 2019 - Jan 2020 · 2 entries
- Anastasia rebased and tested the hash-based series in #32, then Alvaro refactored its copying support into callbacks in #33; Paul repaired the resulting test failures in #36.
- Fujii identified a crash window around tablespace deletion and recommended advancing minimum recovery before removing directories in #37, also requesting separation of the WAL-description enhancement from the backpatchable fix.
Jul 2020 · 1 entry
- After several months without progress, Paul incorporated Fujii's feedback and split out the WAL-description patch in #40, then restricted missing-directory bookkeeping to pre-consistency recovery and updated test APIs in #44.
Jul 2021 - Apr 2022 · 10 entries
- Robert requested clearer explanations of the test infrastructure in #47, and Tom rejected the WAL-description change and unused polling alterations in #50; Kyotaro removed those pieces while rebasing the series in #52.
- Michael raised Windows symlink and temporary-directory concerns in #53, leading Kyotaro to replace the callback approach with explicit tablespace backup management in #58.
- Julien reported Perl failures in #60, and Kyotaro corrected hash-return handling, tar-backup handling, and Windows junction paths in #63 and #64.
- Michael proposed using existing in-place tablespaces instead of extending the Perl framework in #68; Kyotaro adopted that simplification, removed the second test as outside the bug's scope, and combined the fix with standalone coverage in #70.
- Robert questioned why recovery should skip operations rather than recreate their prerequisites in #71, and Kyotaro confirmed that directory recreation also passed the existing test in #72, but the tracking implementation remained the immediate candidate.
- Alvaro refined diagnostics, restricted the recovery flush, and added a warning assertion in #74, which landed unchanged as 49d9cfc68b; he explained the limited backpatch range and unavailable older-branch automated tests in #76.
- Robert then showed that the new skipped-copy mechanism conflicted with other redo paths that already recreated directories in #82, weakening the intended consistency checks.
- Alvaro decided to abandon that mixed strategy in #86, and bf902c1393 reverted the first landing so the fix could be rebuilt around directory reconstruction.
- Robert favored reconstruction as the less risky backpatchable route in #91, and Kyotaro settled on creating temporary directories under pg_tblspc and verifying their disappearance at consistency in #95.
- Kyotaro implemented that replacement in #96, covering both FILE_COPY and WAL_LOG database creation and initially using ignore_invalid_pages to permit recovery despite detected problems; #98 corrected a Windows build warning.
Jul - Aug 2022 · 10 entries
- After a three-month gap, Alvaro rebased the replacement and adjusted its interaction with in-place tablespace tests in #99, then proposed link-check refactoring, an alternative override, and earlier consistency checking in #100.
- Kyotaro reviewed those fixups in #101, and Alvaro consolidated them in #105, replacing ignore_invalid_pages with allow_in_place_tablespaces and tightening directory reconstruction.
- While simplifying link checks, Alvaro exposed incorrect Windows junction classification in #103 and #104; Thomas linked the existing fix from thread #45630 in #109, and Alvaro endorsed it in #111; 9d3444dcce took the src/port/dirent.c fix from that other thread's #503669, expanding its comment.
- Because older branches lacked allow_in_place_tablespaces, Alvaro considered a separate recovery override in #108, but Thomas proposed backpatching the existing developer feature in #110 and Kyotaro supported that choice in #114.
- Alvaro assembled the in-place tablespace feature and its necessary backup and location fixes into 7bdbbb8734, explaining the bundle and omission of unrelated tests in #119.
- Alvaro posted the final reconstruction patch with comment and test corrections in #120, and it landed unchanged as 9e4f914b5e, with stable-branch backports.
- Matthias reported intermittent failures in the new test in #121, Tom confirmed buildfarm failures in #122, and Thomas traced them to prematurely removed WAL in #123.
- Kyotaro supplied the replication-slot delta in #124, which became 59be1c942a with Alvaro's added slot drop after connection detection, described in #125.
- Tom reported continued failures in #126, and Thomas identified a race in dropping the slot before the standby was safely using the WAL in #130, recommending proper standby slot configuration.
- At the last message, Alvaro committed to repairing reliability and investigating the test's substantial runtime in #134; the subsequent 6c1c9f88ad configured and retained the slot and adjusted synchronization, while test-performance improvement remained unresolved.