model-writtenrun 7317 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
May - Jul 2021 · 7 entries
- Masahiko Sawada proposed skipping one remote transaction by XID and exposing replication errors in logs and statistics in #1, explaining in #5 that the goal was an easier alternative to manually advancing a replication origin.
- Amit Kapila emphasized streamed and prepared transactions and durable reset behavior in #2 and #8, while Peter Eisentraut framed the feature as a repair tool and cautioned against overengineering in #27.
- The first prototype in #33 stored the skip XID in pg_subscription, coupled origin advancement with clearing that catalog field, and bundled error-context and statistics patches.
- Amit identified the danger of skipping relation metadata in streamed transactions in #34, and Sawada clarified the change to skip only data modifications in #36.
- Alexey's requests for error text, counts, and database identity in #41 and #46, together with Amit's tablesync discussion in #53 and #57, expanded the statistics design beyond the initial minimal view.
- Following Amit's syntax feedback in #39, the series adopted SET(skip_xid), added statistics-reset support, and split RESET into a prerequisite across #62, #69, and #77.
- Hou identified an unused commit-helper argument in #68; the cleanup already present in #62 was extracted and landed as 16bd4becee, with the push confirmed in #72.
Aug 2021 · 4 entries
- Amit's callback review in #86 and #91 led to consolidation of tuple-conversion and apply error contexts in #93; Tang's missing-context report in #109 and Hou's callback-placement suggestion in #114 then produced the corrected outer-loop handling in #118.
- Sawada's combined Stream Stop terminology patch in #112 was split into the backpatchable documentation correction 0ac1aee0d7 and code rename 4cd7a18968, as Amit confirmed in #120.
- Timestamp review in #134 led to wording that also covered prepare and rollback in #136; Amit's final polish in #146 landed unchanged as abc0910e2e.
- Sawada resent the remaining statistics, RESET, and skipping work in #150, temporarily borrowing a fileset fix from another thread, and Mark Dilger supplied a testing/helper delta in #156.
Sep 2021 · 2 entries
- Discussion of asynchronous statistics and test duplication in #175 and #186 favored standalone view tests and deferred the shared wait helper; the flat worker-statistics design discussed in #183, #185, and #188 replaced nested hashes in #192.
- Peter reopened syntax and wraparound safety in #193, and later discussion in #218, #219, and #220 settled that transient skip targets should not be preserved by pg_dump.
Oct 2021 · 2 entries
- Sawada renamed the view to pg_stat_subscription_workers in #224 for future transaction metrics, while Amit's extensible SKIP(options) proposal in #228 displaced the SET/RESET approach and the statistics patch proceeded alone in #249.
- Amit requested database-local statistics in #253, leading to the storage rewrite in #264; review then fixed Vignesh's long-message failure from #281 and Tang's hidden tablesync-history problem from #292.
Nov 2021 · 2 entries
- After column naming and test revisions, Amit deduplicated purge entries and removed a flaky reset check in #321; the final attachment #324 landed with a catalog-version bump as 8d74fc96db.
- The sidewinder failure exposed unrelated tablesync errors and a test race, prompting Amit to move discussion from thread #225823 here in #326; after the initial fix #1099482 there, Sawada's local revisions #328 and #335 produced the single-query polling fix that landed as 41e66fee05, also removing redundant streaming configuration.
Dec 2021 · 3 entries
- The remaining feature settled on whole-transaction skipping and excluded initial table synchronization in discussion including #338, #340, and #343, then returned with dedicated SKIP(xid=...) syntax in #354.
- Streaming changes remained spooled until application, as explained in #359 and #369, while Vignesh's concurrent-target concern in #361 led toward locking and comparing the catalog value when clearing it in #377.
- Peter requested explicit NONE instead of a magic reset value in #381, and discussion in #385 and #386 retained persistent catalog storage so operator intent survived restarts; these refinements appeared in #390.
Jan 2022 · 3 entries
- Amit noticed a missing conflict-documentation update in #392; Sawada extracted #395, which landed unchanged as 85c61ba892, while automatic target selection from asynchronous statistics remained deferred in #393 and #396.
- The January revisions refined cache handling, privileges, psql display, and tests; Peter supplied the documentation fixup #439, Sawada incorporated it in #440, and Amit added catalog grants and privilege-test corrections in #451.
- David Johnston challenged arbitrary future targets and stale statistics in #459 and #484, and Peter highlighted XID wraparound in #472 and #477, prompting reconsideration of target identity and error storage.
Feb - Apr 2022 · 8 entries
- Sawada replaced XIDs with LSNs in #500, adding warnings and clearing an unmatched target after a successful nonempty transaction; Peter defended retaining a higher-level SKIP command over direct origin manipulation in #502.
- With the release deadline approaching, Sawada narrowed the plan in #503 to logged origin/LSN information plus SKIP, deferred catalog error-detail storage and recording skipped data, and bundled external error-context prerequisites for testing.
- Once those prerequisites landed elsewhere, #508 retained the SKIP patch and integrated disable_on_error behavior; reports from Shi Yu in #510 and Takamichi Osumi in #511 drove streaming and origin-position fixes in #513.
- Sawada restored explicit privilege checking and consolidated tests in #525, Amit polished the patch in #528, and Euler's review in #534 led to #536, which landed with a catalog-version bump as 208c5d65bb.
- Noah Misch reported failures on all five AIX buildfarm members in #539; Sawada added diagnostics in #541, Amit tried the unsuccessful DatumGetLSN conversion in #543, and expanded diagnostics followed in #546.
- Noah diagnosed the mismatch between eight-byte C alignment and four-byte tuple alignment on AIX in #547, and the discussion chose relocating subskiplsn rather than introducing a new alignment mechanism, producing #551.
- Noah requested a regression safeguard in #556 and consistent catalog ordering in #563; the completed fix #566 landed as 79b716cfb7 with a catalog-version bump and parallel-safe test helper.
- Noah then observed that the new test only detected the problem on affected platforms and posted #572 to check the preceding field's end offset; after Sawada's review in #573 and #575, it landed unchanged as c1da0acbb0.
Jun 2022 · 2 entries
- June discussion revisited catalog-order constraints and alternative alignment strategies in #576, #578, and #579, with ABI and upgrade concerns raised in #581 and #583 and Tom Lane supporting the existing reorder-and-check approach in #585 and #587.
- By the final exchange, the SKIP feature and follow-up fixes had landed, while #589 and #590 left a possible requirement that NAMEDATALEN be divisible by eight on affected platforms as an unresolved future improvement.