model-writtenrun 7493 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2017 - Mar 2018 · 5 entries
- Tomas Vondra proposed byte-based ReorderBuffer memory limits and streaming of large in-progress transactions in #1, combining immediate WAL metadata, output-plugin callbacks and subscriber-side disk spooling to reduce commit-time transfer lag.
- Erik Rijkers' testing exposed equal-LSN and multi-node failures, and Tomas corrected MULTI_INSERT assertions, unstreamed-subtransaction aborts and temporary-file collisions in #9.
- Peter Eisentraut questioned inheriting maintenance_work_mem and requested clearer configuration and observability in #12 and #16; Tomas added spill/stream statistics in #14, while the largest-transaction eviction strategy remained.
- Following Andres' proposal and Masahiko Sawada's review, Tomas embedded top-level XID association in the next WAL record in #37, reducing worst-case assignment overhead while retaining separate invalidation logging for standby compatibility.
- Peter supplied a catalog-declaration fix and judged the feature unready for PostgreSQL 11 in #47, citing unresolved memory-limit details, the two-phase-decoding dependency and missing performance evidence.
Dec 2018 - Feb 2019 · 1 entry
- After a long gap, Tomas revived the series with TRUNCATE support in #51; Alexey Kondratov found that tests had not enabled streaming and supplied worker, historic-MVCC and schema fixes in #53, which were integrated through #58.
Aug - Nov 2019 · 4 entries
- Alexey proposed avoiding subscriber disk spooling with a background-worker pool in #61, but rollback-heavy testing exposed a savepoint bug and severe wasted-work costs in #68.
- Amit isolated concurrent-abort handling as the actual two-phase dependency in #71; Tomas introduced the independent 64MB logical_decoding_work_mem setting in #85, and Dilip detached the unnecessary two-phase code in #94 and #96.
- Dilip split memory limiting and spill statistics for independent review in #101, while Tomas and Alexey deferred worker-pool and extra-buffering work to control scope in #110 and #111.
- Vignesh restored serialization coverage with #123, and Amit polished #137 into the memory-limit commit cec2edfa78; the statistics component last reposted in #142 became 9290ad198b with a catalog-version bump.
Dec 2019 - Mar 2020 · 8 entries
- A broken transaction-wide schema flag led Tomas to per-relation streamed-transaction tracking in #167, and Dilip unified commit and streaming replay while correcting invalidation handling in #169.
- Alvaro removed the already-landed transaction-flags cleanup from the remaining series in #192, while incomplete TOAST and speculative inserts became the next major correctness problem.
- Dilip first retained incomplete TOAST chunks in a local list in #198, then replaced that approach with WAL flags in #217 after Amit highlighted repeated work and unsafe catalog dependence in #202.
- Alvaro reviewed a separate subscription work_mem option in #200, but the option was dropped in #225 because its use case remained unclear and streaming did not require it, as discussed in #223.
- Dilip changed invalidation logging from individual messages to command-end batches in #227, reducing overhead without removing the commit records needed by existing standby processing.
- Tomas and Masahiko began exploring removal of redundant assignment WAL with the experimental standby fix in #229.
- The feature was held for PostgreSQL 14 in #231 because concurrent aborts, partial tuples, invalidations and subscriber file handling still needed substantial review.
- Tomas extended the assignment experiment with a standby cache in #244, but unlogged-subtransaction holes and standby accounting concerns led to retaining the existing assignment records, as explained in #245 and #253.
Apr 2020 · 1 entry
- Kuntal Ghosh's reviews in #252 and #263 improved error cleanup and subscriber handling; Erik's cascading tests prompted #261, and Mahendra's invalid-allocation diagnosis in #282 was incorporated in #287.
May 2020 · 2 entries
- Amit and Robert's discussion in #310 led to complete-prefix streaming in #317, allowing completed changes to leave ReorderBuffer while an incomplete TOAST or speculative-insert tail remained.
- Subscriber file leaks and size limits identified in #339 drove Dilip's BufFile prototype in #360, followed by SharedFileSet-managed process-exit cleanup in #384 and #389.
Jun 2020 · 3 entries
- June invalidation review removed duplicate decoding and collected invalidations per transaction in #356 and #377; further testing found missing invalidations after the last command-counter increment, fixed by commit-time logging in #402.
- The assignment prerequisite from #398 landed as 0bead9af48, preserving the old standby assignment records and adding only a WAL-format bump to the posted component.
- The callback component of #398 became 45fdc9738b, with the later review in #452 making test_decoding respect XID and timestamp options.
Jul 2020 · 4 entries
- Amit required consistent startup and snapshot semantics in #399 and #431, while Ajin supplied forced-streaming test patches in #409 and #417; Dilip merged the backend streaming, incomplete-change, abort and SQL-API pieces in #439.
- Ajin adapted streaming counters to the separate replication-slot statistics framework in #446 and #461, creating a side branch that was kept out of the main feature landing.
- The command-end invalidation patch from #402 landed as c55040ccd0, after fast-forward catalog tracking and WAL-structure reuse were corrected in the final review.
- Ajin reported improved throughput under bulk-insert load in #459, and the backend implementation from #464 landed as 7259736a6e after the additional TOAST test and cleanup in #466; the final approach spills incomplete changes and streams them once complete.
Aug - Sep 2020 · 8 entries
- Thomas Munro then diagnosed missing Windows imports in #470, producing the discussion-derived CheckXidAlive and bsysscan declaration fix b48cac3b10.
- Review of BufFile truncation corrected buffer state and dirty-data handling through #489; Dilip's single-fileset experiment in #492 was rejected because per-transaction filesets better preserve tablespace changes and distribution, and the prerequisite from #495 landed as 808e13b282.
- Amit found unsafe list mutation and stale SharedFileSet references in #501; the standalone unregister fix last reposted in #506 became 4ab77697f6, while subscriber review continued.
- After schema-tracking fixes, pg_dump and psql integration, and test separation in #512, the subscriber/protocol/pgoutput patch #514 landed as 464824323e with a catalog-version bump.
- Nathan Bossart's inline declaration correction in #516, extended by Amit in #517, became ac15b499f7.
- The remaining subtransaction, rollback and DDL tests from #518 landed unchanged as 58b5ae9d62, and Amit declared the main series complete while explicitly moving pending statistics work to its separate discussion in #520.
- Tomas' post-landing cleanup #523 became ddd5f6d260, removing an unused protocol declaration and correcting a typo.
- Tom Lane reported an invalidation-callback failure in #527; Dilip reproduced it and tested the first fix in #534, Tom proposed separating initialization from validation in #535, and the revised #538 landed unchanged as 69bd60672a.
Nov - Dec 2020 · 2 entries
- Noah Misch traced the sungazer temporary-file failure reported in #540 to integer XIDs being hashed as strings, supplying #547; a1b8aa1e4e retained the HASH_BLOBS fix and test while improving the background-psql helper.
- Tom broadened that diagnosis into explicit hash-key semantics in #549, then rolled HASH_STRINGS, mandatory HASH_ELEM, documentation and caller cleanup into #555; Noah's wording review in #556 preceded b3817f5f77.
Apr - May 2021 · 2 entries
- A compiler warning reopened concurrent-abort handling in #558, revealing that callbacks could raise the same rollback error for unrelated reasons; Amit's guarded fix #560, reviewed by Dilip, landed unchanged as 2ce353fc19.
- The final exchange #562 and #563 confirms that last fix landed, leaving the main streaming feature and these follow-up fixes complete, with the previously separated statistics work outside this thread's completed series.