model-writtenrun 484 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Bertrand Drouvot proposed separating table and index statistics to reduce unused index fields and improve naming, introducing distinct statistics kinds, accessors, and additional toast/partition tests in #1.
- Andres Freund requested consistent table naming, caller-level separation of creation operations, and less accessor duplication in #3; Bertrand incorporated those changes in #4.
- Melanie Plageman supplied optional counter-prefix, layout, and flush-function cleanup patches in #5, but Bertrand deferred those changes to a separate naming effort in #6 and adopted her source-file split in #7.
- Andres pushed for macros that generate complete accessor functions and a shared-relation lookup simplification in #8 and #10; Bertrand pursued preparatory work separately and incorporated the resulting accessor infrastructure in #14, then simplified its macro interface in #15.
- Review of #17 exposed duplicated helpers, inappropriate index transaction handling, incorrect accessor routing, and missing index test coverage in #18; Bertrand acknowledged the problems in #19, and Nitin Jadhav supported retaining common implementation code in #20.
- The next submissions maintained applicability without resolving those reviews: #23 and #25 explicitly left comments outstanding while transaction-accessor preparatory work proceeded separately.
- The work missed the feature-freeze window, was returned with feedback after stalling in #30, and resumed with Bertrand's November rebase in #32.
- Andres renewed concerns about buffer-manager layering, duplicated code, and unnecessary index transaction fields in #33; Bertrand proposed implementing relfilenode statistics first in #34 and confirmed that ordering in #35, after which the inactive entry was returned with feedback in #36.
- Michael Paquier revived the work in August 2026 with #37, reversing the dependency order after investigating relfilenode statistics and separating the implementation into shared-memory and pending-data patches.
- Michael's rework in #37 retained one relcache pointer, used a statistics-kind-tagged union for pending data, kept common relation machinery, and removed index dependence on table-only subtransaction data.
- Bertrand found missing database tuple-counter accumulation and an outdated catalog description in the shared-memory patch in #38; Michael fixed both in #39, explained why the undocumented table accessors could return zero for indexes, and received confirmation in #40.
- Michael committed the shared-memory and SQL-interface portion of #39 as 72a6dad1c9, with version bumps and a comment cleanup, and resent the pending-data remainder in #41.
- The first landing exposed the available index transaction counters through dedicated functions; Michael raised the absence of an index transaction statistics view in #41, and Bertrand recommended handling such a view separately in #42.
- Reviewing the remainder, Bertrand proposed smaller index pending allocations, kind assertions to protect union members, and retaining the table-specific transaction type name in #42; Michael accepted the latter two in #43 while preferring the existing allocation design for clarity.
- Bertrand measured the possible pending-allocation savings but agreed they did not justify reduced clarity, approved the revised patch, and requested a comment correction in #44; Michael supplied the wording in #45.
- Michael committed #43 with that comment correction as f2ccb8e6c5 and confirmed completion in #46, leaving smaller index pending allocations open for future discussion and turning back to relfilenode statistics.