model-writtenrun 8727 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro opened the discussion by proposing direct per-operation prefetch limits instead of effective_io_concurrency's RAID spindle formula, with separate ordinary and maintenance settings and possible new names #1.
- Andres Freund opposed renaming and prematurely separating random from sequential I/O, while Tomas Vondra warned that retaining the name would silently change tuned values; Tom Lane rejected maintaining two interdependent compatibility GUCs #2 #3 #4.
- Michael Banck questioned whether the maintenance name implied utility commands only, and Thomas Munro explained that it covered work performed on behalf of many sessions, including proposed recovery prefetching and existing VACUUM prefetching #5 #6.
- Thomas Munro withdrew the random-page naming distinction and considered documenting a conversion formula for existing settings; he also explained that Evgeniy Shishkin's proposed io_capacity terminology suggested a system-wide budget rather than the intended per-operation limit #7 #8 #9 #10.
- Thomas Munro submitted two patches that retained effective_io_concurrency's name, used its value directly, and added maintenance_io_concurrency with a supported-platform default of 10 to replace the existing extra-ten prefetch rule in heap_compute_xid_horizon_for_tuples(); he supplied SQL for converting previously tuned values #11.
- Both components landed with light edits: b09ff53667 added header cleanup to the simplification, and fc34b0d9de refined documentation and descriptions for the maintenance setting; Thomas Munro announced the push in #12.
- After Peter Geoghegan asked about closing the remaining renaming open item, Thomas Munro said renaming did not seem worth the disruption and planned to mark it resolved unless objections appeared; both submitted patches were already committed #13 #14.