model-writtenrun 7492 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar 2021 · 23 entries
- Mark Dilger opened this thread with #1, moving work from thread #42262 into a series containing a multi-database ParallelSlots refactor, the pg_amcheck command-line application, and optional PostgresNode corruption-test infrastructure.
- Robert Haas's review in #2 exposed confusing database-name versus pattern semantics, schema-exclusion problems, and inadequate diagnostics; Mark's #3 revised those behaviors and added relation/page progress reporting.
- Robert and Peter Geoghegan argued that corrupted-database errors should remain visible rather than be suppressed using potentially unreliable relation sizes, and their reviews led to portable block arguments, clearer result handling, and pattern arrays in #11, followed by cleanup in #13.
- Sampling was deferred in the discussion around #9 and #10, while Peter Eisentraut requested placing the application in src/bin in #14 and the streamed-backup validation idea in #20 was left outside this project's scope in #21.
- The ParallelSlots portion of #13 landed unchanged as f71519e545, providing lazy connections and reusable slots across databases.
- Robert's database-selection design in #19 and documentation patch in #23 were incorporated into #24, which retained only the application after ParallelSlots landed and dropped the optional corruption framework from the submission.
- Mark incorporated Robert's further documentation edits from #25 and moved the utility into src/bin in #26; Robert committed it with small diagnostic and build-integration edits as 9706092839.
- Peter's compiler-warning report in #29 prompted Robert's direct declaration and assert-only-variable fixes in d60e61de4f.
- Buildfarm failures exposed endian-sensitive corruption setup and alignment-dependent TOAST chunk counts, which Mark repaired in #35 and Robert committed unchanged as 6611256127.
- Mark's #37 supplied separate fixes for authentication-message variation and apparent IPC::Run wildcard expansion; the wildcard workaround landed as f371a4cdba.
- The authentication-test correction from #37 was combined with the portable option ordering in #39 as 24189277f6, and Mark's remaining one-line correction in #41 landed as 9e294d0f34.
- Unsupported Perl q packing led to #54; Tom Lane's word-swapping analysis in #61 shaped the final test data in #62, committed unchanged as 945d2cb7d0.
- Tom's report of pg_statistic failures in #67 started another investigation, and Mark expanded #69 through #73 with autovacuum disabling, cleaner missing-TOAST diagnostics, checks before intentional corruption, and TOAST relation validation.
- Noah Misch reported 196 successful runs with autovacuum disabled in #74, but Mark reproduced the failure with manual ANALYZE in #75 and supplied a focused pg_statistic reproducer in #78.
- The TOAST value identifiers added inline in #82 let Tom identify dead initdb-created values in #83, and Robert traced the false corruption reports to incorrect visibility decisions in #86.
- Robert first committed the missing-TOAST diagnostic correction from #73 as 4078ce65a0, while the visibility and additional-validation questions remained under discussion.
- Mark's #93 proposed visibility repairs with regression coverage and broader TOAST checks, initially adding four SQL result columns to carry corruption context.
- After Tom insisted in #94 that intentionally corrupted tables must not be vacuumed, Robert committed the 003_check.pl autovacuum setting and pre-corruption check from #73 as 87d90ac61f, retaining autovacuum in the logical opclass-damage test as explained in #95.
- That distinction prompted Peter Geoghegan's analysis in #97 and #99, leading directly to 5b861baa55, which lets nbtree VACUUM continue when a corrupt or inconsistent opclass prevents locating a parent downlink.
- Mark substantially reworked visibility in #101 around HeapTupleSatisfiesVacuumHorizon and replaced the extra SQL result columns with relevant fields in diagnostic text, while also exploring compression validation.
- Robert distinguished tuple readability from TOAST-pruning safety in #102 and #103, identified unsafe database access under a heap buffer lock, and supplied an alternative visibility sketch in #104.
- Mark incorporated those ideas in #105, collecting TOAST checks per heap page and executing them after releasing its buffer lock, then split the implementation into four patches in #107 at Robert's request.
- Review led Mark to restore lost validation and remove unnecessary reporting-function renaming in #109; Robert then consolidated its visibility portion in #115, correcting transaction-status ordering and handling corrupt xmax values without unnecessarily refusing to inspect the tuple.
Apr - May 2021 · 9 entries
- After small corrections in #117 and #119, the jointly authored visibility implementation landed unchanged as 3b6c1259f9.
- Robert's review in #122 separated remaining bug fixes from broader improvements near feature freeze, and Mark's #123 divided the work into duplicate-check cleanup, deferred TOAST scans, diagnostics, and additional validation.
- The first two patches of #123 landed as 4573f6a9af and ec7ffb8096, with Robert changing the latter to free the collected TOAST allocations deeply.
- Mark carried the remainder through #125 and #133, and the messaging/alignment patch landed unchanged as 9acaf1a621.
- Robert's objections in #126 to cascading errors and competing chunk-number concepts drove increasingly elaborate missing/extraneous-chunk reporting through #133, #135, and #138, but its size became a concern near feature freeze.
- Robert supplied a much smaller alternative in #139, using actual and expected chunk_seq values instead of ordinal chunk counts, and Mark's testing in #141 and #143 exposed false positives and last-chunk calculation problems.
- Mark independently reconstructed a reporting-only fix in #149, explicitly reserving broader validation and refactoring for PostgreSQL 15; Robert's counterproposal in #150 retained his smaller implementation while incorporating findings from their comparison.
- Mark demonstrated another false positive in #151 and caught a missing-final-chunk boundary error in #154, leading to the arithmetic corrections in #152 and #155.
- The final reporting patch from #155, with wording adjusted through #156 and #157, landed as 50529e5b4e, confirmed by the thread's final message #161; the application and these corrections were committed, while the broader TOAST improvements remained deferred.