model-writtenrun 8026 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2019 · 3 entries
- Julien Rouhaud proposed backend online checksum verification in #1, using a SQL set-returning function, private reads of on-disk pages, dirty-page skipping, an optimistic read followed by a locked retry, and vacuum-style throttling.
- Robert Haas discussed structured results versus immediate notices in #2, while Michael Paquier advocated logging and database-level checksum statistics in #4; Julien incorporated statistics and WARNING output in #9.
- Masahiko Sawada raised persistent corruption reporting and the risk of corrupted disk pages already cached in shared buffers in #6, and Julien clarified in #7 that clean cached pages were still read from disk while dirty pages were skipped.
Mar - Apr 2020 · 3 entries
- Michael and Masahiko's reviews in #10 and #11 led Julien to remove the all-relations helper, separate the SQL entry point, add dedicated throttling settings, and document the feature in #14, with test and GUC-name corrections in #15 and #19.
- Further review from Masahiko in #20 and #22 produced simpler cost settings, temporary-table safeguards, buffer-state checks, and clearer page-validation logic in #21 and #23.
- Julien adopted Masahiko's reusable block-checking interface in #24, then adjusted SQL overloads and permissions in #26 and removed redundant declarations in #28; Masahiko marked the patch ready for a committer in #29.
Jul 2020 · 1 entry
- After the spring review round, Justin Pryzby supplied a separate language-cleanup patch in #32, which Julien folded into the complete rebased submission #33.
Sep - Nov 2020 · 13 entries
- Michael resumed detailed review in September in #34, prompting backend-header separation, relocated TAP tests, and a throttling wait event in #39, followed by Makefile and Windows-build repairs in #42 and #43.
- Julien's benchmark in #44 favored optimistic reads, but Michael's measurements in #47 showed competing effects on checker speed and concurrent workload throughput, renewing the discussion of partition-lock contention and throttling.
- Julien and Michael agreed to simplify the implementation in #48 and #49, and #50 moved the core operation into bufmgr.c, removed optimistic retries and throttling, and split infrastructure from SQL support.
- The split series needed packaging repairs: #51 accidentally resent unchanged content, #52 supplied the promised documentation correction, and #53 restored the omitted SQL implementation file.
- Michael proposed sharing the standard page-verification machinery and dropping expected and stored checksum values from the result in #54 and #56, with Julien accepting the shared checking approach in #55.
- Michael's substantial rewrite #57 introduced pg_relation_check_pages(), returned physical paths and block numbers, allowed checks without data checksums, skipped unsupported and temporary relations, restricted execution to superusers by default, and separated ordinary SQL tests from corruption tests.
- Following Julien's feedback about warning behavior in #58, Michael's final attachment #60 used PageIsVerifiedExtended() without WARNING output, retained failure statistics, strengthened I/O-lock sequencing, and streamlined the recovery TAP tests.
- Michael split #60 into the lightly edited infrastructure commit c780a7a90a and SQL-interface commit f2b8839695, confirming both landings and an immediate buildfarm-driven correction in #62.
- Noriyoshi Shinoda reported a crash when checking multiple corrupted pages in #63, Julien identified the misplaced pfree(path) in #64, and Michael moved it outside the loop in 60a51c6b32.
- Andres Freund challenged the implementation in #66, #67, and #68, identifying I/O under buffer mapping locks, an inadequately aligned page buffer, questionable buffer-state handling, and undocumented assumptions about concurrent writes.
- The discussion in #69, #70, #73, and #74 explored buffer pins, BM_IO_IN_PROGRESS, and relation locking, but did not produce a replacement patch; Andres also identified a relation-existence race and concerns about the SQL function's guarantees.
- Robert Haas objected in #78 that CheckBuffer() could not expose page contents for other checkers, silently skipped dirty pages, and had confusing overlap with amcheck, while Andres called for a revert and outlined alternative buffer APIs in #79.
- Michael accepted that the problems required more than a small revision in #82 and reverted the feature and its follow-up fixes with e152506ade, confirmed in the final message #84; safe concurrency, dirty-page handling, reusable APIs, and placement in core or amcheck remained unresolved.