model-writtenrun 8863 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tomas Vondra reported that logical-replication walsenders repeatedly exposed their original transaction timestamp and proposed refreshing it before starting each decoding transaction in #1.
- Andres Freund questioned the cost and meaning of those timestamp updates in #2, while Kyotaro Horiguchi discussed commit timestamps and existing StartTransaction() behavior in #3; Tomas explained that stale timestamps caused false long-running-transaction alerts in #4.
- The discussion shifted toward suppressing the timestamp: Tomas proposed reporting none in #5, Alvaro Herrera agreed in #6, and Andres explicitly recommended NULL in #7; Craig Ringer supported suppression and argued that any commit timestamp belonged in pg_stat_replication in #8.
- Alvaro replaced the original refresh approach with an am_walsender early return in SetCurrentStatementStartTimestamp() in #10, then committed it with an expanded comment as b175bd59fa and announced backpatching through PostgreSQL 10 in #12.
- Tom Lane reported buildfarm failures in #13 and explained in #14 that changing globally shared timestamp state risked side effects, recommending suppression at the statistics-output layer instead.
- Alvaro posted an untested follow-up in #15 that reverted the internal timestamp guard and filtered walsender transaction timestamps in pg_stat_get_activity(); Tom caught its incorrect || operator and requested && in #16.
- Alvaro incorporated that correction and an explanatory comment into f5d28710c7, preserving internal timestamps while returning NULL for walsender transaction start times.
- In #17, Alvaro confirmed the follow-up had landed and explained that force_parallel_mode=regress caused subscription-related queries to launch workers that inherited a zero timestamp without inheriting am_walsender, triggering an assertion.
- The timestamp fix was committed, but concerns about replication code executing SQL and spawning parallel workers remained open: Alvaro raised the restriction mismatch in #17, and Tom questioned the layering and possible security implications in the final message #18.