model-writtenrun 1360 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jan - Mar 2025 · 4 entries
- Corey Huinker opened #1 to complete statistics transfer across dumps and upgrades, proposing extended-statistics set/restore/clear functions, working pg_ndistinct and pg_dependencies input functions, and reusable attribute-statistics helpers.
- Tomas Vondra proposed structured JSON arrays and validation against the statistics object in #2, then clarified in #4 that attribute-number validity mattered more than enforcing statistical consistency; Corey incorporated duplicate checks and moved MCV construction into mcv.c in #7.
- Jian He found an empty-input crash and integer/special-value problems in #9 and #10; #13 initially retained the historical formats, treated ndistinct counts as integers, and validated both positive column numbers and negative expression numbers.
- The revision #14 replaced the separate setter with a warning-tolerant restore interface, using text object names and returning false for lookup failures to match the relation/attribute facilities.
May - Jun 2025 · 1 entry
- After reconsidering the format at pgconf.dev, Corey argued in #16 that adding input functions was the last convenient opportunity to change output; #17 implemented structured JSON arrays and pg_dump translation from older server formats.
Aug 2025 - Jan 2026 · 34 entries
- After the summer and autumn rebases through #20, Michael Paquier requested independently reviewable input/output changes and initially questioned old-server dump support in #21; Tomas defended compatibility in #26, and Michael accepted the translation approach in #27.
- Corey split the series in #28, and Michael extracted its attribute-argument documentation as d6c132d83b, rewording the comments as reported in #29.
- Michael proposed separate datatype files in #31, Corey implemented the moves in #34, and Michael reordered them ahead of the functional changes in #35; the unchanged move patches resent in #38 landed as a552312343 and 2ddc8d9e9b.
- Man Zeng identified inconsistent output-routine comments inline in #40, leading directly to 040a39ed25, acknowledged in #42.
- Jian He corrected the documentation examples and reviewed the output formats in #36, #37, and #44; the output patches from #47 landed with light documentation and presentation edits as 1f927cce44 and e76defbcf0.
- The input-function review uncovered crashes, soft-error handling mistakes, duplicate keys and excessive dimensions in #46, #49, #50, and #58; #60 enforced ordered attribute numbers and duplicate-key rejection while the tests became substantially more comprehensive.
- Corey demonstrated in #57 that dependencies may omit statistically insignificant combinations, and Michael accepted relaxed completeness rules in #69; Chao Li and Yuefei Shi also reviewed state handling, comments and memory ownership in #63 and #64.
- Michael posted the final input-function pair in #69 and applied it after comment, style and encoding-safe test polishing as 44eba8f06e and e1405aa5e3, reporting roughly ninety-percent coverage in #72.
- Corey resumed the three uncommitted pieces in #81; Michael’s detailed helper-contract review in #84 led to relocation into stat_utils.c and clearer documentation in #85, while the tests stayed with their existing stats_import fixtures.
- Tom Lane reported old-GCC warnings and unsuitable error details in #86; Michael posted a combined fix in #87, and the five direct ErrorSaveContext checks landed separately as 7bc88c3d6f after Tom’s agreement in #88.
- Michael independently applied Corey’s compact replace()-based formatting idea from #51 and #85 to the existing extended-statistics regression output in 2f04110225, rather than taking an attached patch for those files.
- Corey supplied the small follow-on #90 against Michael’s remaining error-message patch #89; Michael extended its unreachable-state elog conversions and committed the combined result as f68597ee77, explaining the additions in #91.
- The stale rebase #92 was immediately corrected by #93, which included allocation and expected-output updates; Michael then committed its shared-helper component with additional naming, header and documentation cleanup as 213a1b8952 and posted the remainder in #94.
- Tender Wang’s typo report, represented by #95 and acknowledged in #96, produced the one-word stat_utils.c correction eee19a30d6.
- Following the January rebase #97, Michael extracted the reusable ndistinct/dependencies validation and free routines, renamed and documented them, and committed 32e27bd320, as reported in #98.
- Michael found unsafe namespace and lock ordering in #97 and redesigned clear in #98 to lock the relation first, take both relation and statistics names, and check MAINTAIN; Chao found incorrect inherited decoding in #99, Corey fixed it in #103, and Michael added lock/test polish before d756fa1019.
- Further review exposed that a user-supplied relation/statistics mismatch could raise an inappropriate hard error; 395b73c045 changed this to a warning with regression coverage, and #106 carried the same handling into the pending restore implementation.
- The restore/dump remainder in #106 and #107 adopted the new relation-first API; Chao’s review in #110 identified cleanup, diagnostics and old-server query problems, which Michael gathered into the appended fix-up patch in #112.
- Michael extracted his new statext_mcv_free helper unchanged from #112 and committed it separately as 7ebb64c557, announcing the extraction in #115.
- After Michael demonstrated a malformed-MCV crash and missing negative-input coverage in #112, Corey added about thirty tests and explicit success flags in #118; #120 then added multirange tests and exposed the need to separate MCV type parsing from attribute-statistics operator lookup.
- Michael again requested smaller independently testable statistics-kind patches in #122; Corey clarified that expressions could be separate and produced the nine-part split #125, using direct type lookup for MCV input rather than extending the shared helper.
- The independent attribute-statistics multirange test from #125 landed with light placement/comment changes as d4504d6f60, including a stable-branch backport, as reported in #126.
- Michael combined the ndistinct-only restore component and selected documentation from #125, removed unused argument scaffolding and polished diagnostics/tests, producing 0e80f3f88d.
- Michael immediately found missing tuple null-flag initialization while reviewing the remaining work and corrected it in 114e84c532, explaining the unexpected NOT NULL failure in #126.
- The separate MAINTAIN/lock test from #125 landed as d9abd9e105 after Michael corrected its mixed-up role names and expected output, as documented in #126.
- The dependencies restore component of #125 followed as 302879bd68, with argument declarations, documentation and test/diagnostic polishing, completing backend support for the two simpler kinds as reported in #127.
- Michael narrowed the pg_dump patch from #125 to ndistinct and dependencies, used the shared format constants and updated upgrade-test comments, then committed c32fb29e97; #127 noted possible cross-version formatting differences to watch in the buildfarm.
- Corey posted the remaining MCV/expression work and a missing dependencies-output check in #128; that complete test landed with SQL formatting and comment edits as f9562b95c6, confirmed in #130.
- Michael discovered that the earlier split had left expression counting in the pending MCV patch, wrongly rejecting negative attribute numbers for ndistinct/dependencies; he extracted it from #128, added focused tests, and committed e3094679b9 following the report in #132.
- Michael demonstrated an overread from an undersized MCV nulls array in #133 and proposed removing that redundant parameter entirely in #134; Corey’s #138 made that change but accidentally omitted several edits from Michael’s intervening version, a misunderstanding acknowledged in #140.
- Michael reconciled #138 with his missing validation, comments and test changes, split out multirange coverage, and committed MCV restore plus pg_dump support as efbebb4e85, explaining the final edits in #139.
- Two further extractions from #138 landed as the reformatted multirange test 0b7beec42a and the reduced cloning test fc365e4fcc, leaving out unimplemented expression-statistics restoration and checking that supported statistics survive copying.
- Michael noticed two malformed diagnostics during the final review in #139 and corrected them, with expected-output updates, in 740a1494f4.
- At the end, clear and restore/dump support for ndistinct, dependencies and MCV had landed, while expression-statistics restoration from #138 remained uncommitted because of its opaque positional text-array interface and limited review time; #140 discussed the security/type constraints, and #141 favored a named-key input, preferably JSON, on the existing restore function and requested a new thread for that remaining work.