model-writtenrun 8811 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Amit Langote reported recursive relcache loading and session-dependent anonymous-record failures, proposing to reject anonymous and self-referencing composite partition keys in #1.
- Tom Lane argued for reusing CheckAttributeType() rather than narrow special cases in #2 and #4; after Amit raised diagnostic concerns in #5, Tom separated pseudo-type rejection from the relcache recursion problem and supplied the generalized type check in #6.
- Tom's independent patch in #7 made partition metadata load on demand, allowing references to the table's rowtype, while identifying stale-descriptor memory retention and the now-unused equalPartitionDescs() as follow-up work.
- Tom added partition-specific diagnostics and discovered that CheckAttributeType() also needed to recurse into range subtypes in #8; its two patches landed separately as fc7695891d, with an added self-inclusion test, and 39ebb943de, with only incidental comment wrapping.
- Amit examined descriptor lifetime handling and redundant conditions in #9, while Tom decided against backpatching the invasive relcache change because of its scope and potential extension ABI effects in #11.
- Drawing on accessor placement from Amit's earlier work discussed in #13, Tom revised the relcache patch in #14 to encapsulate builders, release stale descriptors when references disappear, and construct descriptors safely in temporary memory contexts.
- Amit analyzed how deferred rebuilding avoided descriptor accumulation during repeated invalidations in #17, then supplied a patch on top of the relcache revision to permit whole-row partition expressions in #18.
- Tom clarified in #19 that older branches would receive neither the relcache redesign nor a new prohibition on self-rowtype expressions, since reliably detecting every problematic expression was impractical.
- Tom's tests in #21 and #22 confirmed reduced memory accumulation and modest performance gains, while leaving broader cache-lifetime concerns and other performance problems for future work; the final relcache attachment landed unchanged apart from whitespace as 5b9312378e.
- The deliberately separated dead-code cleanup landed as 42f74f4936, and Amit's whole-row patch landed with minor cleanup as bb4114a4e2, confirmed by Tom in #23.
- Amit supplied comments identifying the required partitioning accessors in #25, and Tom committed them unchanged as fbe0232358, confirming completion in #26.
- Julien Rouhaud discovered that the range-recursion fix rejected ranges with collatable subtypes and supplied the correct subtype-collation lookup in #27; Tom expanded its tests as described in #28 and committed 74b35eb468 before the regression reached a release.
- The thread ended with all submitted work landed, but the older-branch limitation remained: Jobin Augustine reported a PostgreSQL 12 DROP TABLE failure in #30, and Julien demonstrated in #31 that accessing the table first populated its relcache entry and allowed the drop, without proposing another patch.