model-writtenrun 9591 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2018 - Mar 2019 ยท 24 entries
- Michael Banck proposed offline checksum enabling and disabling in #1, combining his work with Michael Paquier's implementation as a useful maintenance tool and a fallback if online enabling missed PostgreSQL 12.
- Robert Haas, Fabien Coelho, and Magnus Hagander favored separate action switches, while Coelho and Hagander favored verification by default; Paquier accepted that direction in #24 after summarizing the discussion in #22.
- Coelho's first code review in #6 found unsafe control-file unlinking, unnecessary write access during verification, missing documentation, and concurrent-start risks; Banck addressed the first implementation issues in #27.
- Changing initdb to enable checksums by default attracted support but remained separate work, with Tomas Vondra raising corruption-recovery concerns in #12 and Stephen Frost arguing for earlier detection in #25.
- After Andres Freund returned the undocumented patch with feedback in #37, Banck resumed with documentation and review fixes in #38, then extracted the shared control-file writer and migrated pg_rewind in #40.
- Paquier took over preparation for commit in #43, separating the common helper, tool rename, checksum feature, and new --no-sync option, while expanding tests and removing the proposed PID-file check.
- The shared helper landed as ce6afc6823 from #43, with a small clarification of callers' locking responsibility.
- Sergei Kornilov's testing in #46 exposed missing cross-version protection; Paquier added a control-file version check in #47, which landed unchanged as c9ae7f704c and was backpatched to PostgreSQL 11.
- The rename component of #47 landed unchanged as 6dd263cfaa, and Paquier resent only enable/disable and --no-sync in #55.
- Coelho demonstrated concurrent-operation hazards and incorrect durability ordering in #58, then supplied the synchronization-order delta in #63; Banck supplied a control-file timestamp guard in #67, incorporating Hagander's observation that a server could start and stop during the operation.
- Kornilov also raised replication and block-size compatibility in #65, prompting Banck's separate BLCKSZ-check patch in #69.
- Hagander argued that warnings alone would not prevent accidental startup and proposed temporarily renaming pg_control in #84; Banck implemented the idea in #87, and Paquier folded it into #90 with signal cleanup and data-before-control synchronization.
- The replication debate considered changing the system identifier, but Banck identified WAL and standby-rebuild consequences in #88, while Paquier's analysis and testing in #89 distinguished ordinary WAL replay from dangerous physical page copying by tools such as pg_rewind.
- Coelho started a separate pg_resetwal cleanup in #102 and expanded it to backend control-file updates in #104, consolidating duplicated writers and moving wait-event tracking into the shared helper.
- Paquier committed that refactor as 8b938d36f7, correcting path handling, backend errors, and no-sync behavior and simplifying the API to a boolean, as detailed in #107.
- Paquier adapted the checksum patch to the new helper in #109, improved signal cleanup with durable_rename, and drafted replication warnings; Coelho's testing in #110 still identified races around the temporary rename.
- Freund opposed the rename mechanism in #114 and #120 because an interrupted operation could leave ordinary users unable to start the cluster or rerun the tool without manual recovery, and argued that a general maintenance lock belonged in separate work.
- Paquier consequently returned to the simpler implementation in #122, following his explicit proposal in #121 to retain the clean-shutdown check without PID-file manipulation, system-identifier changes, or control-file renaming.
- Coelho's renewed review in #126 and #131 emphasized that data must become durable before pg_control advertises enabled checksums; Paquier restored that ordering and adjusted --no-sync in #133.
- Banck supplied clearer interruption and concurrent-start warnings in #136, while Coelho's final testing in #138 found an untranslated write-error message and suggested more accurate replication terminology.
- The main enable/disable component of #133 landed as ed308d7837 with those light documentation and translation edits, followed by its unchanged --no-sync component as e0090c8690.
- Coelho separately proposed applying data-before-control synchronization to pg_rewind in #139, but Paquier requested a new thread for that work in #145, so this thread does not establish its eventual disposition.
- Coelho then reordered the documentation to put corruption risks first and explain problems before remedies in #146; after wording discussion in #147, Paquier committed the lightly edited result as a7cc52370b.
- The thread ends with confirmation of that documentation commit in #149: the checksum feature, no-sync option, rename, and shared-helper changes had landed, while a general concurrency guard remained outside the implementation and Coelho's preference for successful no-op enable/disable requests remained unadopted.