model-writtenrun 3289 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov 2020 - Mar 2021 · 8 entries
- Heikki Linnakangas proposed generic ResourceOwner callbacks and a shared array-plus-hash representation in #1, aiming to make extension-defined resources convenient while keeping recently acquired resources cheap to track.
- Michael Paquier questioned array sizing and JIT ownership placement in #2, Julien Rouhaud corrected hash-capacity documentation in #5, and Kyotaro Horiguchi examined release-loop correctness in #8.
- Craig Ringer supplied SystemTap instrumentation on top of the initial patch in #6, but said ordinary profiling probably offered similar value; that experimental attachment was not incorporated into the eventual series.
- Heikki revised hash cleanup to handle duplicate references and possible table growth in #9, retaining conservative array rescanning because callbacks could release other resources.
- Hayato Kuroda's consistency review in #14 led to standardized wrappers, while Heikki's reservation audit produced a separate preparatory patch in #15; Hayato subsequently checked all enlargement call sites in #17.
- Benchmarks in #18 exposed slower hash-heavy cases, and Álvaro Herrera and Robert Haas challenged accepting that regression in #19 and #22.
- Heikki added a separate cheaper-hash patch in #26, Hayato reproduced the results in #27, and further analysis in #29 identified 32 recent entries as a promising array size.
- The series survived several base changes, including explicit plancache ownership in #32 and newly added HMAC resources in #35, while Zhihong Yu scrutinized release-loop termination in #36 and #38.
Sep - Nov 2021 · 1 entry
- Aleksander Alekseev supplied a rebased series and cleanup patch in #39; Heikki folded those changes into #41, and Aleksander tested and endorsed the next revision in #42 and #43.
Oct - Nov 2022 · 2 entries
- After a long gap, Heikki revived the work in #45, increasing array and initial hash sizes and adding an explicit array-full error; Aleksander then reviewed casts and reported successful Raspberry Pi testing in #46 and #47.
- Andres Freund objected that resources still needed a defined release order and raised future reservation and intrusive-list requirements in #48; Robert Haas supported preserving resource-type ordering in #50.
Feb - Mar 2023 · 3 entries
- Heikki answered with priority-sorted release in #52, forbidding Remember/Forget after release begins and making callbacks stop unregistering themselves, while adding test_resowner and a pgcrypto conversion.
- The stricter lifecycle required a temporary parent-owner workaround during subtransaction invalidation in #52; broader reservation and intrusive-list APIs were discussed but were not included in the submitted redesign.
- Aleksander's sanitizer and coverage review in #53 and missing-test report in #54 led to corrected test integration in #55, which also made PinBuffer callers responsible for both reservations and retained legacy callbacks for extension compatibility.
May - Jul 2023 · 2 entries
- Heikki fixed AbortBufferIO's violation of the new callback contract in #57, and Aleksander considered that series ready for the next development cycle in #58.
- Peter Eisentraut's API review in #59 and Andres's correctness and performance review in #60 led to const descriptors, descriptive warning callbacks, typed wrappers, reverse searches, and separate release/sort state in #63.
Oct - Nov 2023 · 8 entries
- Andres proposed tighter counter placement in #64, Peter identified LLVM adaptation problems in #65, and Heikki incorporated fixes into #67, leaving stronger assertion-build leak diagnostics for possible follow-up.
- After Alexander Lakhin identified typos in #68, Heikki landed the final series and confirmed it in #69: reservation movements became b70c2143bb, the refactor became b8bff07daa, faster hashing became 954e43564d, and pgcrypto conversion became cd694f60dc, with only light cleanup to the main refactor.
- Alexander immediately found an assertion failure in #70; 8f4a1ab471 aligned ResourceOwnerReleaseAll with ResourceOwnerSort's empty-hash test, as explained in #71, without an attached fix patch.
- Alexander's repeated database-creation and deletion test exposed commit-time DSA ownership failures in #72, prompting Heikki's separate test, ownership-fix, and transaction-cleanup series in #73.
- Andres supported explicit DSA ownership in #74, but Alexander found a stale-owner access under Valgrind in #75, leading Heikki to replace mapping_pinned with a nullable owner in #76.
- Thomas Munro endorsed the revised ownership semantics in #77; the series landed as a8b330ffb6, 325f54033e, and c21e6e2fd4, with formatting and a small test adjustment.
- Heikki confirmed the DSA fixes in #78, deferred renaming dsa_pin_mapping, and declined immediate backpatching because no live failure was known on stable branches.
- Andres independently encountered the empty-hash assertion while rebasing AIO in #79, then recognized in #80 that it was already fixed and suggested additional test coverage; his other cleanup suggestions received no further resolution in this thread.
Jan - Feb 2024 · 2 entries
- Anton A. Melnikov identified missing descriptor export markings in #81; Heikki fixed those and two walsummarizer declarations in c6b86eaa55, confirming the additional scope in #82.
- Alexander's disk-full test in #84 exposed an unsafe assumption that forgetting any resource creates room for another; d212957254 added the missing local-buffer reservation and documented the narrower guarantee, as explained in #85.
Jun 2024 · 3 entries
- Alexander found a logical-decoding invalidation failure in #86, and Heikki concluded in #87 that relcache invalidation must work without a live ResourceOwner rather than relying on the parent-owner workaround.
- Heikki posted the focused relcache fix and expanded regression coverage in #88, moving broader relcache refactoring to another thread; e6cd857726 landed that patch with SQL style edits and removed the earlier subtransaction workaround.
- After Robert Haas noticed the final push was missing in #90, Heikki confirmed it in #91, leaving the main refactor and all corrective series landed, with broader refactoring and optional cleanup outside this thread.