model-writtenrun 4127 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov 2022 - Jan 2023 · 6 entries
- Thomas Munro opened the thread with direct-I/O support extracted from the asynchronous-I/O work, proposing three disabled-by-default developer settings plus stack, heap, and shared-memory alignment changes; he emphasized that production performance required future concurrency and read-ahead infrastructure #1.
- Justin Pryzby proposed one comma-separated setting, requested CI coverage, and found a BufFile alignment-sanitizer failure; Andres Freund suggested a bitmask assignment hook, a permanent TAP test, and separate buffer types, then supplied measurements supporting page-aligned buffered I/O #2 #4 #5.
- Jim Nasby raised work_mem, clock-sweep, and SLRU concerns, while Andres Freund argued that those broader improvements should remain separate from the already large asynchronous/direct-I/O effort #6 #7.
- Thomas Munro implemented the revised design with io_direct=data,wal,wal_init, PGIOAlignedBlock, a permanent test, and a throwaway CI default override; BufFile alignment remained deferred because padding could waste memory when many files exist #10.
- After David Rowley landed the allocator dependency separately, Thomas Munro resent the remaining series and fixed handling of systems without direct-I/O support #11.
- Bharath Rupireddy identified GUC-string mutation and allocation leaks and supplied a review delta; after a several-month gap, Thomas Munro incorporated it with co-author credit, repaired the test configuration, and removed misleading wal_sync_method documentation #12 #13.
Apr - May 2023 · 11 entries
- The alignment and io_direct patches from the final main submission landed as faeedbcefd and d4e71df6d7, with additional guards for small block sizes, an omitted WAL-buffer alignment change, and editorial cleanup; the CI-only default override was left out #13 #14.
- The new test exposed a pre-existing log_newpage_range() buffer-pin exhaustion problem, prompting Andres Freund's pin-limiting prototype and then his simpler temporary proposal to use wal_level=replica, which Thomas Munro accepted and committed as 980e8879f5 #17 #19 #20 #21.
- A separate buildfarm failure produced invalid pages and incorrect row counts on Andrew Dunstan's Btrfs system, leading Thomas Munro, Tom Lane, Andres Freund, and Andrew Dunstan to investigate parallel workers, compression, filesystem coherency, and copy-on-write behavior #15 #22 #24 #27.
- Andres Freund pointed out that hint-bit writes can modify a buffer during I/O, and Thomas Munro distinguished that checksum hazard from observed stale reads; Andrew Dunstan confirmed that disabling Btrfs copy-on-write stopped the test failures #34 #39 #42.
- Further experiments implicated page faults because huge pages or PostgreSQL checksums prevented reproduction, and Andres Freund located a pending Btrfs kernel fix; concurrent buffer modification during direct writes remained a separate concern for future work #51 #52 #53.
- Christoph Berg reported a startup failure on overlayfs backed by tmpfs, showing that an operating-system whitelist could not establish filesystem support; Thomas Munro proposed probing O_DIRECT from Perl and then identified tmpfs as the rejecting layer #56 #57 #58 #59.
- Andrew Dunstan and Dagfinn Ilmari Mannsåker reviewed the pre-flight patch, with Ilmari supplying the simpler constant check; Thomas Munro committed the lightly revised test as 6ca8df2d61, and Christoph Berg subsequently confirmed that his build skipped it successfully #60 #61 #64 #66 #67.
- Tom Lane found that GCC 4.2.1 could ignore stack alignment, sometimes without warnings, undermining a proposed warning-based configure check; the discussion settled on replacing the obsolete OpenBSD buildfarm configuration, and Mikael Kjellström brought up a new animal #68 #78 #79 #80 #90.
- Robert Haas asked for understandable portability failures, while Thomas Munro reaffirmed the experimental, opt-in scope and deferred untested Solaris support; later discussion explored buffer replacement and adaptive shared-buffer sizing, with Andres Freund opposing silently disabling requested direct I/O #91 #93 #94 #95 #97 #99.
- Noah Misch argued that io_direct sounded too attractive for an experimental feature, and Thomas Munro initially proposed a startup warning; Justin Pryzby and Tom Lane preferred a debug_ prefix because users might miss warnings or apply advice intended for future releases #102 #104 #105 #106.
- Thomas Munro replaced the warning proposal with the debug_io_direct rename, Noah Misch approved it, and 319bae9a8d landed with additional error-string updates while leaving internal hook names unchanged #107 #109 #110.
Aug 2023 · 2 entries
- In August, Peter Eisentraut supplied a follow-up to rename the hooks and backing variable and restore alphabetical header order, while Peter Geoghegan continued the broader discussion of correlated buffer references and replacement quality #111 #112.
- At the last message Thomas Munro accepted Peter Eisentraut's cleanup for the following day, and it subsequently landed unchanged as 4f3514f201; the developer-only feature and supplied follow-up fixes were landed, while the pin-limiting prototype had no demonstrated landing here and production I/O and buffer-management improvements remained future work #113 #19 #93 #112.