model-writtenrun 6255 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug 2018 · 3 entries
- Nikhil Sontakke opened this thread segment with prepare-time logical decoding, concurrent-abort safeguards, and deterministic tests that coordinate through an XID instead of sleeps in #1.
- Petr Jelinek requested clearer restrictions on plugin catalog access in #2, Andres Freund requested broader heap-scan checks in #3, and Nikhil incorporated both in #4.
- Arseny Sher identified a fundamental missing-data case when a consistent snapshot is reached between PREPARE and COMMIT PREPARED, then developed its restart and tablesync implications in #5 and #9.
Sep - Oct 2020 · 5 entries
- After a two-year gap, Ajin Cherian revived the work in #10, rebasing the remaining feature and dropping concurrent-abort infrastructure already supplied by large-transaction streaming.
- Amit Kapila's review restored attention to concurrent-abort correctness and pgoutput support, leading to additional tests in #12 and publisher/subscriber implementation in #14.
- Reviews by Amit, Dilip Kumar, and Peter Smith drove prepared-state retention, invalidation handling, shared GID lookup locks, explicit enablement, callback cleanup, and dedicated tests across #19, #33, #54, and #55.
- At Amit's request, Ajin separated API, backend, and subscriber layers in #57, and Peter became an implementation contributor by adding subscriber streaming support in #58.
- The team eliminated redundant streaming finish callbacks, shared spool processing, and independently revised backend and subscriber branches before reuniting them in #67, #69, #83, #89, and #90.
Nov - Dec 2020 · 10 entries
- Masahiko Sawada's timestamp testing and Peter's cascading tests helped refine the implementation, while Ajin added streaming concurrent-abort coverage and Amit consolidated cleanup through ReorderBufferResetTXN across #97, #99, #105, and #125.
- The snapshot-boundary problem resurfaced in #132; discussion of restart-safe duplicate detection led to the begin-prepare approach proposed in #143 and implemented in #149.
- Peter introduced a subscription two_phase option in #145, then fixed missing publisher support reported by Sawada and kept temporary tablesync slots outside two-phase mode in #148.
- Amit corrected the delayed-PREPARE LSN, retained original prepare metadata, prevented new snapshots from being attached to prepared transactions, and separated wire operations in the substantial revision #150.
- The team chose to document catalog-lock deadlock limitations rather than introduce automatic fallback, as discussed in #151 and accepted in #159.
- Sawada's duplicate-GID concerns in #152 led to origin-LSN and timestamp checks in #161, followed by rollback prepare metadata and separate replication-origin progress handling in #165.
- After further review and Osumi's successful Valgrind testing in #170, the output API component of #168 landed with cosmetic edits as 0aa8a01d04, announced in #173.
- Amit removed that committed component, centralized filtering, and merged documentation in #174; its ReorderBuffer component then landed with only cosmetic changes as a271a1b50e.
- Ajin's snapshot-boundary isolation test evolved from #158 through a three-session version to #179, which landed with explanatory wordsmithing as 5a3574d7b3.
- Ajin also identified two pre-existing comment errors in #166, and Amit fixed them directly from that report in e02e840ff7.
Jan - Feb 2021 · 4 entries
- The new isolation test exposed a buildfarm failure caused by autovacuum-generated empty transactions; Amit reported the diagnosis in #182 and announced its repair in #183.
- Peter resumed the remaining subscriber work in #184, incorporating a separately developed tablesync prerequisite and later dropping that prerequisite after its landing in #190.
- Intermittently failing development-only concurrent-abort tests were removed in #195, while Markus Wanner's report that ReorderBufferFinishPrepared unnecessarily created transaction state in #196 produced ade89ba5f4.
- A missing-data race during initial synchronization prompted prepare spooling in #199, followed by reviews that improved file persistence, restart recovery, replay, interrupts, and subscription-drop cleanup through #248.
Mar 2021 · 6 entries
- Because that spooling design did not support streaming prepares safely, Ajin temporarily prohibited combining streaming and two-phase subscriptions and prohibited toggling two-phase mode in #213.
- Independent spool refactoring and tablesync early-exit work were removed or moved elsewhere in #218, and committed origin-progress work was removed from the remaining series in #224.
- Amit showed that prepare spooling could deadlock synchronous replication and lacked sufficient retained metadata in #253, leading Ajin to replace it with delayed activation, a two_phase_at boundary, and disabled/pending/enabled states in #255.
- Amit supplied design-comment patches in #262 and #264, Osumi supplied additional tests in #269, and Peter integrated those contributions and further review changes in #272.
- Generated subscriber GIDs in #275 addressed multi-subscription conflicts, while feedback patches through #289 fixed pending-state refresh behavior, empty subscriptions, and table-state lookup efficiency.
- Vignesh found another generated-GID collision after subscription replacement in #292, and Peter corrected it by incorporating the subscription OID in #293.
Apr - Jun 2021 · 6 entries
- Amit requested another development cycle for testing, streaming compatibility, and empty-prepare handling in #295; Peter restored streaming support in #299 and split it from core subscriber support in #301.
- Protocol version 3 and detailed wire documentation arrived in #305, followed by Vignesh's protocol and test review and repeated compatibility updates through #319.
- Ajin added empty prepared-transaction suppression in #321, then replaced inappropriate global prepared state and incorporated crash-regression tests in #329.
- Tang's copy_data=false crash report in #330 exposed transaction-boundary ownership in UpdateTwoPhaseState, which Ajin corrected in #332 and explained in #334.
- Further review strengthened copy_data=false coverage and replaced protocol assertions with proper errors across #345, #348, and #358.
- Ajin separated slot-creation and pg_recvlogical support in #354, Amit combined and polished those components in #363, and their landing was announced in #366, leaving the PostgreSQL 15 remainder in #367.
Jul - Aug 2021 · 7 entries
- Ajin added empty-stream suppression in #368; Tang then found that TRUNCATE could transmit schema before stream-start in #369, prompting the correction #371.
- Amit prepared the core subscriber feature for commit in #373 and #378, with its landing announced in #380; Peter then retained streaming support in #381, moved empty-transaction work elsewhere, and omitted the empty-stream optimization.
- Tom Lane reported potential GID buffer overruns through Coverity in #383; Peter's bounded-copy fix, revised to use destination sizeof in #388, landed unchanged as 01c3adcdd8.
- Amit revisited duplicated protocol and worker logic in #393, so Peter extracted helpers and streamlined tests in #396; the helper component landed with local interface changes as 91f9861242.
- Amit relayed a buildfarm failure and diagnosed a test waiting for only one of two subscriptions in #403; Ajin's fix #404 passed review and landed unchanged as eaf5321c35.
- Final streaming reviews refined protocol wording, redundant return values, and cascade assertions, culminating in #417; successful final testing preceded its unchanged landing as 63cf61cdeb.
- Amit announced the last series commit in #421 and closed the CommitFest entry in #422, completing the core and streaming two-phase work while the empty-transaction optimization continued separately.
Jan 2022 · 1 entry
- The final exchange revisited whether replication slots need a two_phase property: Sawada questioned it in #423, and Amit explained in #424 that delayed activation until initial synchronization completes prevents data loss; no further patch followed.