Nov 2025

4 commits
Clear 'xid' in dummy async notify entries written to fill up pages
12 Nov 2025Heikki Linnakangasmaster19beta1REL_19_STABLE19beta1REL_18_STABLE18.2REL_17_STABLE17.8REL_16_STABLE16.12REL_15_STABLE15.16REL_14_STABLE14.21
LLM analysis, 100% surewritten from the discussion, no matching patch1 of 1 reviewer saw the final version

The single added assignment qe.xid = InvalidTransactionId in asyncQueueAddEntries exactly matches Álvaro's inline proposal in #98. Heikki confirms in #99 that it was committed separately because he forgot to include it in the main fix.

  • Heikki Linnakangassaw the final versioncommittercommitted the patch
Fix remaining race condition with CLOG truncation and LISTEN/NOTIFY
12 Nov 2025Heikki Linnakangasmaster19beta1REL_19_STABLE19beta1REL_18_STABLE18.2REL_17_STABLE17.8REL_16_STABLE16.12REL_15_STABLE15.16REL_14_STABLE14.21
LLM analysis, 89% suremodified before commitsignificant changesfrom patchset #961 of 2 reviewers saw the final version

Combines the SLRU-locking refactor from #89 with the latest contributing attachment #96, preserving its empty-listenChannels guard with revised comments. Relative to that small attachment alone, the commit also contains the substantial page-processing and notification-buffering refactor.

  • Joel Jacobsonsaw an earlier versiontrailerReported benchmark results for both versions in #77 and #82, before the first-LISTEN behavior added in #96.
  • Arseniy Mukhinsaw the final versiontrailerIdentified error-path notification loss in #79, raised first-LISTEN concerns in #92, and accepted the final guard in #97.
Fix bug where we truncated CLOG that was still needed by LISTEN/NOTIFY
12 Nov 2025Heikki Linnakangasmaster19beta1REL_19_STABLE19beta1REL_18_STABLE18.2REL_17_STABLE17.8REL_16_STABLE16.12REL_15_STABLE15.16REL_14_STABLE14.21Also in: BUG #16961: Could not access status of transaction (not this thread)Also in: BUG #18804: LISTEN on channel fails with "could not access status of transaction"
LLM analysis, 99% surepart of a posted patchsetlight changesfrom patchset #892 of 7 reviewers saw the final version

The freezing patch in #89 supplies the unchanged AsyncNotifyFreezeXids implementation and vac_truncate_clog hook; the commit omits its async-notify isolation-test changes and lightly simplifies the retained xid_wraparound TAP test. This November submission establishes the source here, contrary to the from_discussion attribution in thread #70682, whose August patch concerns tail advancement; thread #72973 has no competing patchset.

  • Daniil Davydovsaw an earlier versionLLM analysis, 99% sureIdentified a two-listener failure in shared minXid tracking and reviewed locking assertions and naming in #4, before the substantial redesign.
  • Yura Sokolovsaw an earlier versionLLM analysis, 96% sureProposed replacing completed queue-entry XIDs with FrozenTransactionId while reviewing the scanning approach in #16, a change later adopted in the freezing implementation.
  • Arseniy Mukhinsaw an earlier versionLLM analysis, 99% sureIdentified unsafe SLRU-slot reuse in #30 and stale-tail and unsignalled-listener problems in #61, before freezing replaced the horizon-retention approach.
  • Masahiko Sawadasaw an earlier versionLLM analysis, 100% sureRejected fallible SLRU work during abort in #50 and provided detailed locking, iteration, and test feedback in #59 before the freezing redesign.
  • Álvaro Herrerasaw the final versionLLM analysis, 100% sureRequested an explanation of freezing in #81 and tested the final approach sufficiently to identify unnecessary dummy-entry XID lookups in #98.
  • Tom Lanesaw an earlier versionLLM analysis, 95% sureReviewed the freezing implementation attached to #66, which has no indexed patchset, and identified the stale-local-copy race and lock-contention concern in #70.
  • Heikki Linnakangassaw the final versioncommittercommitted the patch
Escalate ERRORs during async notify processing to FATAL
12 Nov 2025Heikki Linnakangasmaster19beta1REL_19_STABLE19beta1REL_18_STABLE18.2REL_17_STABLE17.8REL_16_STABLE16.12REL_15_STABLE15.16REL_14_STABLE14.21
LLM analysis, 100% suremodified before commitlight changesfrom patchset #892 of 3 reviewers saw the final version

Matches the first patch in #89: ExitOnAnyError replaces the PG_TRY/PG_FINALLY recovery and obsolete volatile qualifiers disappear. Differences are confined to explanatory comments and commit-message wording.

  • Matheus Alcantarasaw an earlier versiontrailerDiscussed notification loss and recoverable errors in #85 and #88 before the FATAL-handling implementation appeared in #89.
  • Álvaro Herrerasaw the final versiontrailerExplained in #90 why reconnecting applications must reestablish LISTEN and reread current state, supporting the final FATAL behavior.
  • Arseniy Mukhinsaw the final versiontrailerFound the error-handling regression in #79 and reviewed removal of recovery handling and its first-LISTEN consequences in #92.

Credits across 4 commits

Author
Heikki LinnakangasLLM2
Alvaro Herrera1
Joel JacobsonLLM1
Reviewer
Arseniy Mukhin2
Heikki Linnakangascommitter2
Alvaro Herrera1
Joel Jacobson1
Matheus Alcantara1
Committer
Heikki Linnakangas4

N LLM means N of the credits come from LLM analysis, not from commit trailers.