model-writtenrun 7138 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar - May 2022 · 8 entries
- Aleksander Alekseev opened the thread by extracting XID formatting and 64-bit SLRU page numbering from the larger 64-bit XID project, hoping these preparatory changes could reach PostgreSQL 15 independently #1.
- Peter Eisentraut rejected embedded formatting macros in translatable strings, and Tom Lane recommended literal long-long formats with explicit casts; Maxim Orlov implemented that approach, with further corrections from Kyotaro Horiguchi #3 #6 #8 #9 #15.
- Peter explained that translation files should be maintained separately and questioned premature formatting changes, so the authors dropped that attachment and moved SLRU work to the front of the series #18 #21.
- An experiment with unsigned SLRU page numbers followed Andres Freund's suggestion, but Peter's concerns about signed APIs and sentinel values led the authors back to int64; filename expansion was deferred and SlruFileName became an inline function #16 #23 #26 #27.
- Epochless output exposed test and pg_upgrade problems, prompting Maxim's test correction and Pavel Borisov's temporary removal of control-data output changes before formatting and epoch removal were split into separate patches #30 #31 #33.
- Kyotaro found additional width mismatches and unnecessary widening, leading to corrections in SLRU callbacks and FileTag while keeping existing segment filenames #32 #33 #34.
- Peter resisted declaring the SLRU refactoring ready while reviews still found bugs and tests did not exercise the change, so Aleksander started a separate SLRU test-coverage effort #36 #38 #39.
- The series then spent months being rebased by Maxim and Aleksander; Simon Riggs considered it ready, while the authors retained this smaller thread to separate it from the larger XID implementation #41 #45 #48 #50 #52.
Sep 2022 - Mar 2023 · 9 entries
- Maxim accidentally posted the full eight-patch XID implementation here, but the mistake was acknowledged and Aleksander resumed the smaller three-patch branch #54 #55 #56 #57.
- After the separate test_slru infrastructure became available, the series adapted its callbacks and SQL interfaces; Andres's UBSAN report then exposed int32 cutoff pointers being read as int64, which Aleksander fixed before removing unnecessary casts at Maxim's suggestion #63 #66 #67 #69 #70.
- Following Robert Haas's advice reported by Maxim, the effort pivoted away from XID formatting toward SLRU internals, 64-bit caller addressing, and pg_upgrade conversion #71.
- Andrey Borodin questioned whether the new series really eliminated wraparound and why migration required its complicated conversion algorithm, while Japin Li supplied a cleanup that Maxim incorporated with additional comments #72 #73 #77.
- Jacob Champion identified that the supposedly internal-only patch also changed WAL records, leading Aleksander to revise the commit explanations and clarify that this was major-version work rather than a backportable refactoring #78 #80 #81.
- Aleksander then removed the attempted epoch-derived addressing, merged the series into one indexing patch, and replaced the page-copying migration algorithm with simpler file copying or linking under padded names #82 #83.
- Heikki Linnakangas challenged incomplete upgrade handling and the lack of actual wraparound elimination, then proposed preserving existing filenames until each SLRU individually adopted wider addressing #85 #89.
- That plan produced opt-in long filenames without pg_upgrade changes, followed by a notification-queue conversion that removed wraparound bookkeeping and retained the old default space limit through max_notify_queue_pages #92 #94.
- Maxim added FullTransactionId-based two-phase filenames as another bounded step, avoiding a much larger PGPROC interface conversion, and Aleksander explicitly approved that patch #95 #96.
Jul 2023 · 1 entry
- Heikki's renewed review found that twelve hexadecimal characters could not represent the required SLRU segment range, so Aleksander increased long filenames to fifteen characters and documented the choice #97 #98.
Sep - Dec 2023 · 10 entries
- After another review gap, Alexander Korotkov revived the series, requesting range assertions, a safer notification-limit comparison, and tests using genuinely large page numbers; the contributors also agreed to preserve attribution to the earlier authors #99 #101 #102 #103 #546105.
- Aleksander and Maxim independently submitted revisions, and Aleksander temporarily returned to his passing branch after Maxim's high-page tests failed; Maxim subsequently traced the failure to strtol truncation on 32-bit systems and restored the tests with strtoi64 parsing #108 #110 #111 #112.
- Heikki proposed normalizing all short filename lengths, but Pavel and Korotkov favored preserving existing formats for this change; they settled on replacing the unresolved comment with an explanation of the legacy lengths #116 #117 #118 #121.
- Korotkov committed the SLRU indexing patch from the final submission with only formatting differences as 4ed8f0913b, preserving existing filenames for SLRUs that had not opted into long names #121.
- The notification conversion and two-phase filename patches from the same submission landed unchanged as 2cdf131c46 and 5a1dfde833 #121.
- Maxim's high-page-number regression tests also landed unchanged from the final submission as a60b8a58f4, exercising SLRU pages above the 32-bit range #121.
- Alexander Lakhin had reported a declaration warning and a spelling error before the landing; after Tom Lane highlighted widespread buildfarm warnings, Pavel supplied the declaration fix, which landed unchanged as b589f211e0 #114 #122 #123 #124.
- Korotkov separately corrected Lakhin's reported spelling error directly from the discussion in ae2ccf66a2, without a corresponding attached fix patch #114.
- The focused commitfest entry was closed as complete, while the broader 64-bit XID implementation remained separate #126 #127 #128.
- Thomas Munro subsequently reported an epoch assertion failure in two-phase recovery; Korotkov found that recovery checked prepared transactions before advancing nextXid and posted a repair whose procarray.c portion landed as c64086b79d, omitting an incidental variable rename #129 #130.
Jun - Aug 2024 · 6 entries
- In June, Noah Misch found surviving int conversions and incorrect filename-range comments, reopening the work as a PostgreSQL 17 issue; Aleksander supplied corrections and expanded them after Noah identified another async.c case #132 #133 #134 #135 #136.
- Michael Paquier found further missed notification variables and requested clearer filename examples, leading Aleksander to expand both the code fixes and documentation #137 #138 #139 #140.
- Michael committed the documentation with editorial simplifications as ffb0603929 and the page-number fixes with three additional predicate.c conversions as 3937cadfd4 #140 #141.
- Noah then found missed segment-number types in slru.c and multixact.c; Michael posted a further patch and Noah approved it, but c9e2457390 included only its logging changes #143 #144 #145 #147.
- In the final messages, Noah pointed out the missing slru.c hunk and Michael acknowledged missing multixact.c changes as well, attributing the omission to staging during a rebase and promising to repair it #148 #149.
- The four omitted type-conversion hunks first proposed in #144 ultimately landed as bd06cc338d, completing the repair and backpatching it to PostgreSQL 17; its final source is Alexander Korotkov's later complete attachment #1113188 in thread #238152, credited to Michael Paquier.