model-writtenrun 9786 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2016 - May 2017 · 6 entries
- Stas Kelvich began with PREPARE-time logical decoding, carrying transaction GIDs through WAL and initially reporting prepared-transaction events through the existing commit callback #1.
- Simon Riggs questioned GID overhead, while Stas and Robert Haas explained why external coordinators and failover need preserved transaction names; measurements and discussion led toward logging GIDs only where logical decoding needs them #3 #6 #20 #34 #42.
- Craig Ringer explained that decoding before the commit decision enables synchronous distributed transactions, then identified premature catalog visibility in the snapshot builder and called for separate two-phase callbacks #16 #39 #40.
- Concurrent rollback exposed the risk of catalog tuples disappearing during decoding; Robert rejected unbounded locking, Petr Jelinek explored a two-pass catalog scan, and the initial revision instead adopted plugin filtering to defer unsafe transactions #48 #51 #52 #58 #64.
- Stas implemented separate prepare, commit-prepared, and abort-prepared callbacks with filtering and origin metadata, then corrected premature reorder-buffer cleanup; Craig expanded tests that exposed filtering inconsistencies #66 #68 #75.
- Unresolved locking, rollback, restart, and testing questions led Craig to recommend deferral beyond PostgreSQL 10, while Stas supplied an experimental prepare/decode/commit regression harness and addressed Masahiko Sawada's build feedback #81 #82 #83 #84.
Sep 2017 - Apr 2018 · 15 entries
- After the long pause, Yura Sokolov refreshed the patch and repaired duplicate WAL metadata, subtransaction snapshot cleanup, filtering, prepared output, and GID alignment #92 #93.
- Nikhil Sontakke added built-in pgoutput protocol and subscriber support, although the attachment retained the earlier decoder structure and omitted several Yura fixes despite claiming his latest base #95.
- Petr argued against making rollback wait indefinitely, and Nikhil tried retaining recently aborted catalog tuples while introducing per-change abort checks and an abort callback #100 #103.
- Peter Eisentraut supplied a compiler-warning fix and reported an assertion failure in the vacuum changes; Nikhil added concurrent-rollback TAP tests and continued refining the implementation #108 #109.
- Nikhil made prepare filtering stable across restarts, freed reorder-buffer state at PREPARE, restored commit-time invalidations, and added crash tests, while moving the vacuum proposal out for separate discussion and adding callback documentation #111 #114.
- The next approach introduced a flags cleanup and PGPROC decoding groups with LogicalLockTransaction/LogicalUnlockTransaction to protect catalog access against concurrent rollback #115 #116.
- Stas found a savepoint-related assertion, and Andres Freund demanded independently explained components and objected to exposing abort decisions and locking complexity to plugins; Nikhil split the series and restored test_decoding coverage with a savepoint test #117 #119 #122 #126.
- Simon Riggs extracted an independently committable WAL GID and replication-origin component, which Nikhil incorporated and Simon then announced as pushed #128 #131 #132.
- Tomas Vondra reviewed locking invariants, protocol behavior, callback policy, and memory costs, then supplied a review delta; Andres rejected requiring every output plugin to manage explicit locking #127 #133 #134.
- Following Petr's exploration of automatic protection, Nikhil moved locking around apply_change in core, folded Tomas's review into the remaining series, and added ERROR-path cleanup #137 #140 #147 #151.
- Tomas found cleanup assertions and missing opt-in behavior, while Stas identified protocol incompatibility with PostgreSQL 10; Nikhil removed built-in pgoutput support pending protocol negotiation and per-subscription design, retaining test_decoding and external-plugin support #152 #153 #161.
- Tomas's next review exposed decoding-group leader races, leading to XID rechecks, cleanup fixes, and another full revision #162 #163 #166 #167.
- Andrew Dunstan proposed committing the reduced series, but Andres objected to the late architectural changes and continuing bugs, and Andrew withdrew the proposal without a landing #168 #170 #171.
- The built-in subscriber component last attached here in #151 contributed to a8fd13cab0, with substantial changes to protocol messages, generated subscriber GIDs, and deferred two-phase enablement until table synchronization completed.
- Combining streaming with two-phase subscriptions subsequently landed as 63cf61cdeb from #484680 in thread #39143, rather than from an attachment in this local thread.
Jul 2018 · 4 entries
- Nikhil resumed after a three-month gap with a rebase and restored omitted locking calls, but renewed reviews questioned PGPROC lifetime and synchronization #176 #177 #178.
- Robert Haas identified unsafe reads, interruptible abort waits, and cleanup races, while Tom Lane rejected blocking abort because it could create unbreakable deadlocks; Nikhil accepted that the design needed reconsideration #179 #180 #183 #190.
- Andres proposed checking the decoded transaction's status after catalog scans and catching a special error to stop decoding without stopping the WAL stream, replacing the entire abort-blocking design #193 #195 #197 #199.
- Nikhil implemented that replacement with CheckXidAlive and catalog-scan checks, then addressed Andres's review with better status checks, heap-access guards, plugin capability options, and deterministic check-xid testing #205 #206 #209 #210.
Oct 2018 - Feb 2019 · 7 entries
- The discussion split across threads, and the January refresh adapted the remaining four-patch series to current TAP infrastructure #212 #213 #214.
- Arseny Sher and Petr Jelinek found unresolved snapshot-consistency, premature-forgetting, empty-transaction, callback, and abort-semantics problems, while Alvaro Herrera identified specific macro and comment defects in the independent flags cleanup #216 #217 #218 #219.
- The local thread ended waiting on the author, with no replacement attachment after the January review and a proposed return-with-feedback disposition after months of inactivity #220 #221 #222.
- Alvaro later committed the flags component of the final local series as a7b6ab5db1, applying small macro, type, comment, and rebase adjustments to #214.
- The output-plugin API eventually landed as 0aa8a01d04 from the newer submission #465274 in thread #39143, superseding the callback patches attached here.
- Prepare-time ReorderBuffer decoding landed as a271a1b50e from #465673 in thread #39143, followed by the snapshot-consistency isolation test 5a3574d7b3 from #465912 there.
- Replication-protocol and pg_recvlogical two-phase enablement landed as cda03cfed6 from the newer work in #480700 in thread #44319.