model-writtenrun 4630 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Nathan Bossart proposed replacing logical-worker polling with scheduled deadlines in #1, adapting the walreceiver approach and adding a tablesync retry deadline; he noted that separate worker-notification fixes were needed when periodic wakeups disappeared.
- Hayato Kuroda reported successful basic testing but found that repeated SIGHUP signals could indefinitely postpone pings and timeouts in #2; Nathan acknowledged the behavior and discussed preserving existing deadlines in #3, but the submitted revisions retained the reset behavior.
- Tom Lane's review in #5 requested a unified wakeup array, tablesync accessors, timestamp abstractions, safer enum handling, and fresher clock readings; Nathan incorporated much of this in #6, added one-second retries for exhausted tablesync slots, and supplied a separate walreceiver cleanup patch.
- Thomas Munro suggested reusing TimestampDifferenceMilliseconds in #7; Nathan identified its overflow limitation in #8, and Thomas explained in #10 that the wait infrastructure requires an INT_MAX bound rather than LONG_MAX.
- Tom Lane committed the walreceiver cleanup portion of #6 as 24ff700f6a, adding explicit timestamp-infinity names and moving clock reads closer to their uses while leaving the logical-worker changes pending.
- Tom supplied a separate timestamp-helper patch in #12 that clamped wait calculations to [0, INT_MAX], replaced walreceiver's manual arithmetic, and corrected existing caller mistakes in basebackup_copy.c and postmaster.c.
- Nathan identified signed-overflow and underflow cases in #13 and #15; Tom addressed them in #16, Nathan approved that revision in #17, and it landed unchanged as 3a28d78089.
- Returning to the logical-worker patch in #18, Tom questioned retaining one-second tablesync polling and suggested explicit slot-release notifications or launcher involvement; Amit Kapila explained the transaction-boundary synchronization requirement and questioned the busy-loop workaround in #19.
- Nathan considered exit callbacks and the risk of stale apply-worker PIDs in #20, but noted that removing polling from other synchronization and copy paths would require more invasive changes and questioned whether the benefit justified the complexity.
- Amit encouraged pursuing simpler improvements in #21, and Nathan returned with the remaining logical-worker patch in #22, adopting the committed timestamp helpers while explicitly leaving the larger tablesync design questions unresolved.
- Amit's review in #23 distinguished clean tablesync exits from error cases, suggested possible worker-exit notifications, and questioned additional clock reads and clearing the sync-start deadline during configuration reload.
- After Daniel Gustafsson asked about those outstanding comments in #24, Nathan confirmed in #25 that he had not addressed them; the two supporting fixes had landed, but the logical-worker wakeup patch remained pending.