model-writtenrun 5229 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun - Jul 2021 · 4 entries
- Tomas Vondra opened the thread with #1, building on Cary Huang's earlier work to keep sequences usable after failover to a logical replica by adding decoding callbacks, test_decoding support and built-in replication.
- The initial design in #1 emitted ordinary sequence changes immediately but queued changes to newly created or rewritten relfilenodes transactionally, using a relfilenode-to-transaction hash to preserve rollback semantics.
- Peter Eisentraut argued in #6 for automatic handling of owned sequences, less duplicated table/sequence code and subscription tests; Tomas accepted much of that direction in #7 while retaining the need to publish independent sequences.
- Tomas added TAP coverage in #9, exposing that replication waits could miss sequence changes from aborted transactions, while Robert Haas suggested treating sequence changes as non-transactional operations whose emission could depend on transaction completion in #10.
Sep 2021 - Apr 2022 · 25 entries
- At Peter's request in #11, Tomas split infrastructure, test_decoding and built-in replication into separate components in #12, also adding streaming callbacks so the foundational work could proceed independently of publication syntax.
- Hannu Krosing proposed logging sequence state once at commit in #13; Tomas implemented and refined that alternative in #15 and #16, finding ordering, dropped-sequence and cached-value problems.
- Peter and Petr Jelinek favored the original approach in #17 and #18, and Andres Freund emphasized the difficulty of tracking sequence use without WAL in #19; Tomas returned to direct decoding in #22, incorporating Peter's documentation, test and tuple-access fixups from #21.
- The next revision, #24, removed a redundant sequence-creation argument and restricted extra XID assignment to logical WAL, while Amit Kapila examined synchronous replication and prepared transactions in #25 and #31.
- Álvaro Herrera reviewed publication grammar and completion details in #29, while Tomas separated the broader sequence durability problem into another discussion as explained in #33.
- Tomas temporarily added individual-fetch WAL logging and corresponding test changes in #34, then explicitly returned to the earlier approach in #35 because the durability problem predated sequence decoding and could be fixed independently.
- Peter wanted sequence output enabled by default, while Petr wanted consumers to retain an opt-out in #36 and #37; #38 and #40 implemented that compromise, incorporated Peter's fixups from #39, and settled on include-sequences with portable integer formatting.
- The infrastructure component of #40 landed as 0da92dc530 with light documentation and code polishing, and Tomas posted the two remaining components in #41.
- The test_decoding component of #41 landed as 80901b3291 after removal of two redundant comments, leaving built-in replication and its test workaround in #42.
- The revisions through #44 removed nonportable debugging output and made TAP waits reliable after rollback; Peter's detailed review in #45 then drove the shared relation handling, pg_dump support, psql displays and publication tests in #48.
- Publication syntax remained contested: Amit supported combined table/sequence publication syntax in #50, Euler preferred aggregated psql output in #54, and Peter argued for an all-objects interpretation of FOR ALL TABLES in #55.
- Amit reported assertion and unexpected-output failures in #56, then identified a missing transaction-registration call in #57; Tomas added it without an attached patch in d5ed9da41d and announced the fix in #61.
- Tomas reproduced checkpoint-dependent extra sequence output in #62 and disabled sequence decoding in unrelated tests with a180c2b34d, announcing the fix and remaining sequence-test instability in #63.
- Peter supplied further fixups in #67 and #72; Tomas folded them into the series, reduced subscription-side duplication, documented the protocol and replaced the schema-publication boolean with a character type in the revisions through #74.
- The design in #74 retained separate table and sequence schema-publication entries and allowed a combined all-tables/all-sequences publication, rather than adopting automatic publication of every owned sequence.
- Amit and Petr questioned unnecessary relfilenode creation in #78 and #79, leading to in-place non-transactional updates in #82; Petr also clarified in #81 that replication overwrites subscriber state rather than reconciling distributed sequence allocation.
- The built-in replication patch from #82 landed as 75b1521dae with light documentation, error-message and protocol-tag edits, as Tomas described in #84.
- Masahiko Sawada supplied missing CREATE PUBLICATION completion in #87, and that attachment landed unchanged as 2d2232933b.
- Amit's multi-sequence publication crash report in #86 prompted Tomas's unattached parser and regression-test fix, 002c9dd97a, correcting sequence handling in preprocess_pubobj_list.
- The cleanup commit 05843b1aa4 combined the unused-parameter removal from vignesh C's #92 with Amit's privilege, lookup and documentation corrections from #90; Tomas explicitly declined the attached completion-rule consolidation in #94.
- Amit's synchronization and privilege concerns led Tomas to post the OID-and-relation-map fix in #93, but #95 left apply-worker synchronization, skip-LSN handling, privilege checks and missing-relation diagnostics for further work.
- Amit demonstrated in #98 that sequence changes could be lost before snapshot consistency while corresponding table changes were replicated; Tomas's isolation tests in #102 also exposed a transactional assertion failure and analogous logical-message behavior.
- Tomas explored global and per-transaction buffering in #107, finding difficult interactions with snapshots, aborted transactions, streaming and memory accounting, while Masahiko identified replay-after-crash concerns in #108.
- Amit distinguished the logical-message assertion fix from the broader sequence redesign in #110 and challenged cross-transaction snapshot assumptions in #111; Tomas then revived commit-time sequence logging in #112, leaving locking costs and dropped-sequence behavior unresolved.
- Tomas reverted the sequence feature and its follow-up changes with 2c7ea57e56, explaining in #113 that the necessary redesign was too extensive after the final commitfest and attaching a complete commit-time-logging replacement.
May 2022 · 1 entry
- Zheng Li proposed an independent non-transactional-message LSN-tracking solution in #115, reporting successful concurrency tests and preliminary sequence experiments, but the supplied thread contains no landing of that proposal.
Aug 2022 · 4 entries
- Months later, Noah Misch supplied a deterministic standby-promotion PANIC reproducer in #116, initially associating it with the revert; Tomas's testing in #117 and Noah's confirmation in #118 instead implicated recovery prefetching.
- Thomas Munro proposed error-path cache invalidation in #120, and Kyotaro Horiguchi explained the nonblocking retry problem in #121 and #122; Thomas subsequently changed to invalidation before reads and added error-reset and retry-control patches in #127.
- Tomas questioned the cache-performance implications in #128, prompting the explanation-only revision #129; Thomas later combined those three patches, adjusted the error-exit invalidation and committed 932b016300.
- At the last message, #129, the recovery fix was still posted for review, while sequence replication remained reverted and the replacement proposals in #113 and #115 remained uncommitted in the supplied evidence; the subsequent recovery commit did not restore sequence replication.