model-writtenrun 116 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Sami Imseih proposed dedicated DSA/dshash storage for variable-numbered statistics kinds in #1, motivated by a future pg_stat_statements integration that would benefit from isolated scans, locks, and memory limits.
- Bertrand Drouvot reviewed the initial patch in #2, identifying allocation-failure hazards, unnecessary scans during per-kind resets, a redundant DSA helper, and a test-module preload regression; he also suggested making dedicated hashes and dedicated DSAs independent options.
- After the July proposal and August review, Sami returned with a rebased revision in #4 that addressed several findings while retaining coupled hash/DSA allocation and whole-file statistics loading or discard.
- Michael Paquier proposed initializing per-kind mappings at attachment time in #6; Sami adopted that direction in #7, removed redundant shared pointers, expanded restart and reset coverage, and exposed own_hash through pg_stat_kind_info.
- Michael's review in #8 led to direct hash and DSA array lookups, named test configurations, batched counter updates, and separation of the view change into its own patch in #9.
- Michael supplied a cleanup patch alongside the full series in #10, and Sami folded it into #11, added handling for unregistered kinds during transactional drops, and removed the redundant own-hash test helper.
- Michael committed the storage feature from #11 as 44539786b8, replacing the runtime kind-range check with an assertion and polishing comments and tests, then committed the view portion as 6296e78c3d with a catalog bump and checks for both hash configurations; he explained the final adjustments in #12.
- Zsolt Parragi reported a startup crash and supplied a fix with a reproducer in #14: reusing a variable-numbered kind ID for fixed-numbered statistics could make persisted hash entries encounter NULL storage mappings.
- Sami revised Zsolt's fix in #15, checking fixed_amount specifically for hash entries and adding a separate registration check rejecting serialized-name callbacks for fixed-numbered kinds.
- Michael questioned the cost and narrow coverage of the reproducer in #16, Zsolt agreed in #17, and Sami removed the test in #18; a broader framework for testing malformed statistics files remained a possible future improvement.
- The two code fixes from #18 landed unchanged in reverse submission order: the startup-crash check as 0d110901f0, credited to Sami and Zsolt, followed by callback validation as 3c6665bf02.
- Tom Lane reported intermittent counter failures and diagnosed a missing flush in #19; Sami reproduced the shutdown race and supplied pg_stat_force_next_flush() after the batched updates in #20, which Michael accepted in #21.
- Alexander Lakhin reported broader shutdown-delay testing in #22, including custom-statistics failures and unrelated test problems; the attached counter fix landed unchanged as 8afc886831.
- By the final exchange, the feature and both follow-up fix rounds had landed; Michael redirected the separate dropped-entry assertion issue in #23, and Alexander planned to investigate a possible standby-restart failure in another thread in #24.