model-writtenrun 9471 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - May 2019 ยท 17 entries
- Andres Freund opened the thread by criticizing undocumented FSM API parameters, caller-managed global state, heap-specific assumptions, and repeated filesystem calls; Tom Lane likewise favored storing the map per relation in relcache #1 #2.
- John Naylor and Amit Kapila explained the existing performance tradeoffs, while Andres argued that persistent cached state could reduce system calls and avoid the wasted space caused by checking only alternating pages #5 #6 #7 #8.
- Tom initially proposed reverting before beta, but Andres suggested allowing roughly two weeks for a cleanup proposal, and Tom agreed to resolve the issue by the end of April #9 #13 #14.
- John proposed removing the local map and checking only the final heap page as a simpler fallback, but Amit and Tom favored trying a per-relation cache, leading to Amit's first working prototype #12 #15 #16 #20.
- Review shifted the prototype toward retaining search state between insertions, checking every page, and invalidating cached information when space becomes available; Amit implemented the search changes while the choice between availability bits and actual free-space bytes remained unsettled #21 #23 #24 #26 #27.
- John supplied a delta that checks the cached FSM size before calling smgrexists(), then reported an alter_table regression failure when testing the relcache revision #29 #30.
- John separately proposed an API cleanup on HEAD that lets access methods supply their own FSM thresholds, restores the simpler GetPageWithFreeSpace() interface, and introduces GetAlternatePage(), while noting awkward frontend-header and pg_upgrade dependencies #32.
- John supplied an invalidation fix-up, Alvaro Herrera argued that vacuum should invalidate cached information whenever it frees space, and John reported that adding a clear/invalidation call in RecordPageWithFreeSpace() eliminated the regression failure #34 #35 #37.
- At the agreed decision point, Andres favored reverting for PostgreSQL 12, and John agreed that the replacement was not ready and should be revisited only with sufficient time and interest #41 #46.
- Amit identified that John's FSM-creation fix could fail to notify other backends when vacuum had no local map, then posted a consolidated revision with direct invalidations for FSM creation and newly available space #49 #51.
- John tested the consolidated revision with attached debugging instrumentation and found that it could try the same full block twice consecutively; he renewed his call for a revert, and Amit accepted that the redesign needed more time #52 #53.
- Amit posted a complete revert covering the original feature and its follow-up changes, while asking whether to retain the existing implementation for the release; John replied that the revert looked complete #54 #55.
- Andres and Tom reaffirmed their preference for reverting, and Robert Haas agreed that the feature-freeze deadline and remaining bugs made further rescue attempts inappropriate #57 #58 #59.
- Robert also questioned the cost of additional invalidation messages and suggested broader alternatives such as shared FSM storage, while Andres argued for amortizing searches with cached state and potentially reusing vacuum's existing invalidations #59 #61 #64 #65.
- Amit committed the revert from #54 unchanged as 7db0cde6b5, removing the small-relation optimization and leaving the relcache, API, and syscall-reduction proposals uncommitted.
- After the revert, discussion continued over whether backends should communicate cached free-space changes at all: Tom favored independent caches and questioned invalidation guarantees, while Andres stressed timely discovery of an on-disk FSM #67 #68 #69 #70.
- The thread ended without a replacement design agreement: Andres mentioned renewed work on shared relation-size tracking, Tom suggested refreshing FSM expectations during ordinary size checks, and Andres raised shrinking relations and prepared statements as remaining problems #71 #72 #73.