model-writtenrun 7361 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Justin Pryzby proposed changing a table's access method through a tuple-level rewrite, with explicit detoasting and tests using heap2, after experimenting with zedstore #1.
- Michael Paquier questioned the extra TABLE keyword, and Justin adopted SET ACCESS METHOD while adding documentation, completion, and a separate proposal for newly created partitions to inherit their parent's access method #2 #4.
- Michael supported inheritance affecting future partitions without changing existing ones, reviewed dependencies and comments, and recommended combining the partition extension with its get_rel_relam() helper #5.
- Jeff Davis argued that access methods must already handle toasted input for operations such as INSERT SELECT, and Michael agreed that tablecmds.c should not duplicate that responsibility #11 #13 #14.
- A dedicated dummy table access method was considered for broader testing but was not required for this feature, and Jeff took over the main patch with explicit detoasting removed and the partition extension left aside #15 #17 #18.
- Michael's next review led Jeff to add catalog-value checks, invalid-command tests, explicit OID initialization, and rejection of partitioned tables; optional CLUSTER and VACUUM FULL support remained outside the proposal #19 #20 #23.
- Zhihong Yu questioned the duplicate-command error wording, Michael identified redundant checks, and Justin favored the SET TABLESPACE wording and checking before the no-op return #21 #22 #24.
- After Vignesh reported compilation problems, Justin rebased and cleaned up the main patch, restored the partition extension with its helper merged in, and then corrected expected output following Vignesh's test failure report #25 #26 #28 #29.
- Michael asked that partitioned-table support move to a separate discussion, polished the main patch's tests and comments, and committed #30 unchanged as b0483263dd #27 #31.
- Returning from vacation, Jeff identified a missing object-access hook, inconsistent handling of repeated no-op subcommands, and a completion-comment typo; Michael fixed the typo directly in 15c6ede045 #32 #33 #35.
- Michael found that existing file-swap hooks concerned internal objects rather than the altered table and supplied a small follow-up patch to invoke the table-level hook after a rewrite #35.
- Jeff asked about invoking hooks for no-op commands but accepted the simpler fix; Michael subsequently demonstrated the extra state needed for no-op coverage and chose not to include it, while Jeff also declined to press for stricter repeated-subcommand handling #36 #38 #39.
- Michael committed the original hook patch #35 unchanged as 7b565843a9, leaving no-op behavior unchanged and avoiding backpatching because hook consumers could observe a behavior change #40.
- The hook addition produced two extra sepgsql regression-log entries; after rhinoceros confirmed the differences, Michael updated expected output in 1e3445237b, closing the immediate follow-up work at the thread's final message #40 #41.
- Partitioned-table support remained deferred when this 2021 discussion ended; it subsequently landed as 374c7a2290 from the March 2024 patchset #557369 in thread #45970, which superseded the precursor #29 with revised default and dependency semantics, pg_dump support, and broader tests.