model-writtenrun 2537 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Bertrand Drouvot proposed removing duplicate WAL write/fsync counts and timings from pg_stat_wal, retaining wal_bytes because it measures something different, and removing PendingWalStats and track_wal_io_timing to simplify per-backend statistics work in #1.
- Michael Paquier supported the removal but requested shorter documentation and merging the timing-helper cleanup into the removal patch in #2; Bertrand incorporated those changes in #3.
- Michael extracted the WAL-configuration documentation from #3 into 2f70871c2b and committed its monitoring documentation separately as 4538bd3f1d, adjusting markup and paragraph placement.
- Michael corrected the newly introduced sync_time typo to fsync_time in f11674f8df, acknowledged the oversight, and posted the remaining code-removal patch in #4.
- Bertrand endorsed the remaining patch in #5, and Michael committed it with catalog and statistics-format version bumps as 2421e9a51d, announcing completion in #6.
- Andres Freund objected in #8 that timing WAL operations under highly contended locks can impair scalability and must remain independently controllable.
- Bertrand explained the earlier benchmark-based decision in #9, but Andres showed in #10 why slow clocks, higher concurrency, and workloads without a disk-flush bottleneck could invalidate those conclusions.
- Bertrand responded with a corrective patch in #11 that restored track_wal_io_timing and required both timing settings for WAL write/fsync measurements while leaving the duplicate pg_stat_wal fields removed.
- Michael reworked that fix in #12 so WAL and non-WAL timing could be enabled independently, included WAL reads and initialization, replaced an extra function argument with a zero-time sentinel, and adjusted a backend assertion; #13 resent the same content with corrected patch prefixes.
- Bertrand accepted the independent controls and sentinel approach in #14, while expressing a mild preference for the earlier treatment of WAL reads and supplying additional documentation restrictions.
- Michael committed #13 with those documentation additions and a comment clarification as 6c349d83b6, confirming in #15 that the independent controls were the chosen balance; finer per-object/per-operation controls remained only a possible future idea, with no submitted patch left pending.