model-writtenrun 9524 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andrew Dunstan proposed a PostgreSQL 20 callback allowing table access methods to declare logical-decoding support, reporting that the existing heap-only REPACK (CONCURRENTLY) restriction rejected an external AM with working decoding; his initial patch included heap support and regression tests #1.
- Chao Li questioned why an apparently AM-wide capability needed a Relation argument #2, and Antonin Houska suggested exposing resource-manager IDs so REPACK could inspect rm_decode directly #3.
- Xuneng Zhou found that wal_level=minimal made the tests fail before reaching the AM capability check and requested a temporary replica-level configuration for both build systems #4.
- Andrew retained the Relation argument for possible per-table settings and rejected resource-manager inspection because an AM can use multiple managers and an rm_decode callback alone does not guarantee complete, REPACK-compatible decoding; his second patch also fixed the test configuration #5.
- Antonin accepted the explanation and clarified his interest in a more generally useful interface #6; Alvaro Herrera then proposed querying a mask of supported features to avoid accumulating individual boolean callbacks and unnecessary capability checks #7.
- Andrew's third patch adopted the feature-mask interface, retained Relation, added a single-feature wrapper for REPACK, and tested absent callbacks, unknown flags, and masking of excess returned bits; he explicitly excluded relation emptiness because it describes data rather than AM capabilities #8.
- Zsolt Parragi identified the missing requirement to filter decoded changes through change_useless_for_repack() and reported that the test helper could crash on partitioned tables or views, supplying a partitioned-table reproducer #9.
- Andrew's fourth patch documented the filtering contract, made the test helper reject relations without a table AM, added regression cases for partitioned tables, views, and indexes, and asserted a non-null table AM in the core wrapper; he explained that REPACK already checks relation kinds and reported passing regress, isolation, test_decoding, and injection_points suites #10.
- The thread ends with the fourth patch awaiting further review, with no linked commit or subsequent acceptance recorded #10.