model-writtenrun 6875 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul - Aug 2026 · 24 entries
- Noah Misch opened the thread with a reviewed, AI-generated reproducer showing that an in-flight worker could overwrite REFRESH SEQUENCES with stale values while marking sequences READY, and supplied additional findings for investigation #1.
- Vignesh first proposed skipping already-initializing sequences with warnings and attached an implementation plus an injection-point test #2 #5.
- Hayato suggested coordination resembling slot synchronization, while Shveta preferred checking for a running worker; Vignesh then implemented a single command-level error rather than potentially many per-sequence warnings #3 #4 #7.
- Noah warned that correctness must not depend on users noticing warnings, and Amit chose rejection after considering future automatic synchronization; Shveta also identified restart-loop inconvenience and suggested an alternative DATASYNC state that was not adopted #8 #9 #11 #12.
- Vignesh expanded the series to address read-only defaults and older publishers, then moved publisher validation to subscription connection changes and narrowed the read-only override for backpatching after review #15 #19 #22.
- The concurrency component of #22 landed as f38afa4abb, with Hayato's requested translator comment and the existing-test synchronization wait.
- Vignesh redirected the concurrent-publisher-drop finding to thread #248581 in #17; the implementation from #1148044 there landed as 55f5186848 with light comment and test refinements, and its completion was reported here in #72.
- Tom Lane reproduced intermittent failures under valgrind and argued that ordinary users should not have to know whether a synchronization worker was running; Vignesh's test repairs removed a redundant refresh and waited for worker exit, but Tom called for a backend fix instead #28 #29 #30 #32 #33.
- Amit replaced rejection with stopping the worker and resetting sequence states, initially using an exclusive pg_subscription_rel lock, and announced that the original rejection would be reverted #35.
- The rejection was undone by fd27c7e64f; Amit then simplified the replacement to rely on the subscription object's existing exclusive lock and added an assertion documenting that requirement #36.
- Vignesh and Shveta tested the replacement's interlocks, but the injection-point test failed intermittently on Amit's Windows system, so it was left out with Vignesh's agreement #37 #38 #39 #40.
- The stop-and-reset implementation from #36 landed as a8c9d2be28 with explanatory-comment edits, and Hayato also confirmed its locking and interrupt behavior #41.
- For read-only defaults, Noah accepted changing the GUC after Amit explained the complexity of overriding every transaction; Vignesh submitted separate branch variants in #42, and 4cd02d5522 landed the fix and test #24 #26 #27.
- Although the read-only commit message describes a broader development-branch override, the supplied diff and checkout for 4cd02d5522 actually place the override in SequenceSyncWorkerMain(), matching the narrow variant described in #42.
- The remaining publisher-version patch was resent, but Hayato demonstrated that ALTER SERVER could bypass connection-change validation, causing repeated worker failures against PostgreSQL 18 #47 #49.
- Vignesh restored version checks in refresh and worker execution; Shveta accepted worker-side errors instead of easily missed warnings, and Hayato verified the revised behavior #50 #51 #53 #54 #55.
- Amit standardized the publisher-version error codes and comments in Vignesh's patch, Vignesh tested CONNECTION and SERVER scenarios on both branches, and #56 landed unchanged as 57e7c52c0f #57.
- Vignesh next bundled five fixes for disabled subscriptions, irreversible slot-drop ordering, psql completion, monitoring timestamps, and catalog documentation; Hayato redirected worker shutdown toward rereading subscription state rather than launcher-side termination #58 #59.
- The revised bundle in #60 was approved by Hayato and committed unchanged as d2768065a9 #61 #62.
- Vignesh reproduced shared-lock-table exhaustion while collecting many sequences and proposed immediately releasing weaker locks; Hayato verified the fix, and Amit expanded the explanation of why early release is safe #63 #64 #65.
- The lock implementation from #63 landed as e46c2a6571 with the expanded comment and without the reproducer, which Vignesh had supplied only for local verification #67.
- Vignesh and Amit left four findings unchanged: uncommon SET ROLE failures, intentional sharing and prioritization of synchronization-worker capacity, a narrow concurrent subscriber-drop race, and existing documentation distinguishing snapshot synchronization from incremental replication #68 #70.
- Peter Smith spotted a typo in the landed lock comment; Vignesh fixed it together with misleading table-oriented sequence-origin warnings, and #71 landed unchanged as c7d8049682 #69.
- Vignesh summarized the accepted fixes as complete, with the mixed-snapshot sequence-definition issue tracked separately and four findings rejected; Amit subsequently closed the open item #72 #73 #75.
Sep 2026 · 4 entries
- After the closure, Nikolay Samokhvalov demonstrated that publication refresh could still remove a sequence already queued by the worker, causing an internal error and potentially disabling the whole subscription; he supplied a worker-stop patch and deterministic test #76.
- Andrey reproduced and tested Nikolay's fix, but Shveta questioned stopping a worker shared by many sequences and supplied another reproducer; Vignesh replaced termination with checking membership immediately before updating each sequence, preserving unrelated batch progress #77 #78 #79.
- Vignesh omitted the follow-up automated test because of its recurring setup cost for a rare case, then incorporated Zhijie's suggestions to use a direct cache-existence check and delay relation locking #79 #80 #81.
- The thread's last message reports Shveta's successful verification of the final patch #82; that version subsequently landed unchanged apart from whitespace as 883da6d2c2, leaving no submitted fix awaiting commitment in the supplied record.