model-writtenrun 8789 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Hsu, John opened the thread with a reproducible logical-decoding failure under a subtransaction-heavy pgbench workload, showing several replication slots failing after decoding restarts #1.
- Arseny Sher diagnosed a false error when the first transaction-assignment record precedes restart_lsn and proposed removing the check; his initial series also addressed a separate race where a newly created slot restores a snapshot and becomes consistent before it can decode complete transactions #2.
- Andres Freund questioned whether removing the check would conceal another bug, requested isolation-test coverage, and suggested ignoring serialized snapshots during initial snapshot construction #3; Arseny explained why the reported failure could occur during valid decoding and why relaxing the check would require additional information #4.
- After renewed reports and questions, Arseny explained that slots sharing an unfortunate restart_lsn could fail together, while slots at different restart positions need not fail simultaneously #10.
- Amit Kapila explored skipping assignment records instead of removing the check, while Arseny distinguished existing-slot decoding, new-slot snapshot restoration, and assignment processing during snapshot construction #15 #16 #17.
- Arseny's revised series adopted Andres's simpler prohibition on serialized snapshots during slot creation and corrected the explanation of why transaction IDs alone cannot identify partially observed transactions #19; Amit agreed that changing persisted data or substantially redesigning decoding was unsuitable for backpatching and prioritized the check removal #20.
- Arseny supplied a preliminary isolation test #21, and Amit requested its inclusion despite initially being unable to reproduce the error #22; the complete patch integrated that test, and subsequent revisions added a second get_changes call to handle intervening running-transaction records reliably #24 #25 #26.
- Dilip Kumar found that hint-bit WAL could introduce a top-level transaction record and defeat the test #27, but confirmed that his nondefault wal_log_hints setting explained the discrepancy after Amit questioned it, so the proposed extra savepoint was not incorporated #28 #29.
- Amit polished comments and the commit message, supplied variants for newer and older branches, and tested through PostgreSQL 9.5 #30; Dilip and Arseny subsequently confirmed successful back-branch testing #31 #33.
- Amit committed the final check-removal patch and isolation test as e3ff789acf, backpatched through PostgreSQL 9.5, and announced the push in #34; the available diff matches the preceding executable changes, with the comment polishing described in #30.
- Review then returned to the uncommitted snapshot-creation patch: Amit proposed an explicit flag controlling snapshot restoration, while Arseny defended using an invalid start_decoding_at to represent an unknown safe streaming point #35 #36 #37 #38.
- The thread ended with the subtransaction-check fix landed but the snapshot fix unresolved: Amit raised replication-slot copying concerns #39, and Arseny argued that copying needed a separate correction and clarified that restart_lsn governs WAL reading whereas start_decoding_at governs transaction streaming #40.