model-writtenrun 8610 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct 2019 - Apr 2020 ยท 26 entries
- Amit Langote proposed automatically publishing a partitioned table's existing and future partitions in #1, initially maintaining explicit partition membership through DDL while requiring matching leaf tables on subscribers.
- Rafia Sabih's synchronization test exposed an attempt to synchronize the partitioned parent as a regular table in #3, prompting the tablesync fix in #4; automatic creation of subscriber partitions was left for a separate project.
- Petr Jelinek raised attachment/detachment semantics and the need to distinguish leaf publication from root publication in #5, and Amit demonstrated the existing detach behavior in #8.
- Peter Eisentraut argued in #9 that publication membership should store only the explicitly added parent and discover ancestors during replication, leading Amit to replace the DDL-maintained design in #11 and supply missing dump support in #13.
- After Peter questioned strict publisher/subscriber relation-kind matching in #14, Amit expanded the implementation in #16 to publish changes under the root identity and route subscriber inserts, while subscriber UPDATE and DELETE support remained incomplete.
- Peter argued in #17 and #20 that leaf publication remained valuable for partition-by-partition initial synchronization and that root publication should be explicit, which Amit accepted in #22.
- The four-patch series in #23 restored leaf publication as the default and separated publication membership, the publish_using_root_schema option, worker refactoring, and root publication with subscriber routing.
- Peter's cross-version and synchronization review in #24 and Rafia's code and wording review in #27 led to revised validation, version-aware queries, and a rebased series in #25 and #28.
- Peter's API review in #30 led Amit to move partition expansion entirely onto the publisher in #33, eliminating specialized subscriber queries and unnecessary executor API changes.
- Amit abandoned duplicate parent/partition membership restrictions and separated subscriber routing from root publication in #34; Peter and Rafia's subsequent reviews led to three partition-expansion modes and attach/detach tests in #38.
- After the January review round, Peter committed the membership component of #38 as 17b9e7f9fe in March, converting the partition-mode constants to an enum and polishing comments as reported in #39.
- Peter extracted synchronization preparation in #41, replacing COPY (SELECT *) with an explicit column list to exclude generated columns; Amit reviewed it in #42 and #43, and the redundant publishability check was dropped before c314c147c0.
- Amit resent the remaining features in #43, and Peter's review in #47 prompted helper renaming, relocation, corrected index-opening order, and more tests in #48.
- Peter committed the worker helper extraction from #48 as 97ee604d9b, while warning in #49 that subscriber UPDATE routing still lacked meaningful coverage.
- Testing uncovered faulty UPDATE routing, so Amit corrected it, documented attribute-map indexing, and introduced a separate tuple-lookup refactor in #51, then clarified comments and tests in #52.
- The tuple-lookup component of #52 landed as d8653f4687, with only small comment and statement-placement edits.
- Peter identified further untested routing branches and partition-map initialization concerns in #53, Petr Jelinek examined the memory-context issue in #54, and Amit supplied partition-map and UPDATE fixes with expanded tests in #56.
- The publication option became publish_via_partition_root following the naming discussion in #53 and #54, while misleading pgoutput coverage results prompted investigation by Amit, Petr, and Tom Lane across #56, #57, and #447062 would be unsupported by the supplied messages.
- Peter committed the subscriber-routing component of #56 as f1ac27bfda, revising documentation and tests, and Amit resent the remaining ancestor-publication patch in #67.
- Amit improved ancestor-publication tests and catalog documentation in #68, then incorporated Peter's pg_dump and psql fixups from #69 and simplified ancestor selection into get_rel_sync_entry() in #70.
- Amit repaired TRUNCATE handling in #71, and Peter committed ancestor publication as 83fd4532a7, omitting the extra root replica-identity check and announcing completion of the feature series in #72.
- Amit reported a buildfarm failure in #73, and Peter traced it to RELCACHE_FORCE_RELEASE while identifying immediate test-node shutdown as the coverage-reporting problem in #74.
- Amit's test delta in #75 restored an omitted routing UPDATE, exercised publisher row conversion, and stopped test nodes cleanly so coverage data would be saved.
- Peter diagnosed dangling tuple-descriptor references and posted the inline copying fix in #78, while Amit's alternative in #79 combined the same lifetime correction with broader pgoutput cleanup.
- Peter committed only the focused descriptor-lifetime fix as 12fb189bfe and explained the narrower scope in #80.
- At the thread's end, the main features and relcache fix had landed, while Amit's rebased pgoutput cleanup and test-coverage improvements remained pending in #81.