model-writtenrun 9574 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun - Oct 2018 · 8 entries
- Jerry Jelinek proposed a default-on wal_recycle setting to avoid expensive reads when overwriting old, uncached WAL files on ZFS, with synchronous-standby stalls providing the motivating workload #1 #8.
- Peter Eisentraut and Andres Freund requested stronger performance evidence and more cautious documentation, while Alvaro Herrera asked about disk-full behavior; Thomas Munro reproduced a substantial benefit on a small FreeBSD/ZFS machine, and Jerry supplied benchmark and disk-exhaustion results #2 #5 #10 #11 #17.
- Robert Haas considered automatic tuning, but Tom Lane argued that COW filesystems also made zero-filling questionable, and Andres Freund emphasized retaining fixed-size WAL segments; the discussion distinguished recycling policy from file initialization #20 #21 #25 #26.
- Jerry revised the documentation to explain the default and intended COW use, leaving the recycling implementation unchanged while discussion of alternative controls continued #33.
- While investigating whether existing WAL-size settings could disable recycling, Kyotaro Horiguchi found a separate checkpoint-pointer bug, posted a correction, and immediately revised its explanatory comment #36 #37.
- Kyotaro concluded that min_wal_size and max_wal_size could not cleanly disable recycling, and Robert explained that a minimum of zero merely permits zero recycled segments whereas the requested behavior requires an upper bound of zero #38 #47.
- Tomas Vondra ran an extended benchmark campaign across ext4, XFS, Btrfs, LVM snapshots, and ZFS on SSDs and rotating disks, finding workload-dependent results rather than a universal advantage, and also caught the sample configuration's incorrect default #49 #55 #58 #59.
- Peter requested that disabled recycling bypass its unused calculations, but the patch was returned with feedback after progress stopped in autumn 2018 #61 #63.
Feb - Apr 2019 · 8 entries
- Jerry revived the work in February 2019 after discovering that zero-filling new WAL files could trigger ZFS write throttling during autovacuum, replacing the recycling-only switch with wal_cow_fs to disable both recycling and zero-filling #64.
- Alvaro supplied a cleaned-up version named wal_cow_filesystem with revised documentation, and Jerry incorporated most of those edits and reported a clean build and passing make check #65 #67.
- Robert objected that a generic COW switch coupled independent behaviors and obscured their meaning; after discussion with Andrew Dunstan and Alvaro, Alvaro accepted separate controls and Jerry implemented default-on wal_recycle and wal_init_zero settings #69 #70 #75 #78 #81.
- Thomas recommended pg_pwrite for creating the sparse segment and questioned the documentation's safety rationale, then withdrew that objection after tracing it to historical filesystem bugs; Jerry adopted pg_pwrite and rebased the patch #82 #84 #85.
- Robert requested clearer behavioral documentation and project formatting, which Thomas supplied in the final attachment along with FreeBSD/ZFS tests of all four setting combinations that again showed benefits from disabling both options #86 #88.
- Peter approved the revised documentation and Jerry reviewed Thomas's version favorably, while Tomas offered further benchmarking with less RAM #89 #90 #91.
- Thomas committed 475861b261 from #88 after changing error handling to preserve errno immediately after writes and making further description, comment, and cast cleanups; he explicitly left the uncertain historical durability commentary for separate work #92.
- The thread ended with Jerry thanking the participants after both configuration controls landed; the separate recycling-calculation patches have no recorded disposition in the supplied discussion #93 #36 #37.