model-writtenrun 9656 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro proposed removing the vestigial smgr SQL type because forthcoming storage managers for undo and SLRU data did not need a user-visible catalog type, attaching the removal and regression-test adjustments in #1.
- Tom Lane asked how future storage managers would be selected, and Thomas explained that special database OIDs could select implementations without catalog lookups, avoiding circular dependencies when reading low-level transaction data #2 #3 #5.
- Tom favored special tablespaces and warned about fake database entries, while Robert Haas argued that reserved database OIDs could preserve real tablespace selection and that removing the SQL type was independent of this broader design decision #8 #9 #11.
- Tom explicitly accepted the removal but identified obsolete storage-manager README text, providing the documentation correction incorporated into the committed patch #12.
- Shawn Debnath relayed Anton Shyrabokau's suggestion to reuse fork-number bits, Tom favored that approach over consuming database or tablespace identifiers, and Andres Freund explained why widening buffer tags could hurt performance while discussing a future buffer-mapping redesign #13 #15 #17 #18.
- Thomas explored adding an explicit storage-manager selector to smgropen(), while Andres and Tom considered small built-in OIDs that could fit in a split fork-number field; these remained proposals beyond the type-removal patch #19 #22 #23.
- Shawn proposed reorganizing SMgrRelation into a generic handle with implementation-specific state and offered a separate patch, while Thomas advocated leaving md.c intact initially and deferring separation of file layout from actual I/O to later work #24 #25 #26 #27.
- Konstantin Knizhnik described compression, encryption, snapshot, and raw-storage motivations for separating layout from I/O, and Thomas discussed the associated space-management tradeoffs without expanding the submitted patch #28 #29.
- Thomas committed #1 as 91595f9d49 with README cleanup and a catalog-version bump; his final message confirmed the removal had landed and said he was experimenting with fork-number bits for subsequent storage-manager work #30.