model-writtenrun 8838 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar - Jun 2016 · 3 entries
- Thomas Munro opened the thread with a kqueue implementation of WaitEventSet, carried over from earlier API work and adjusted for NetBSD's integer-valued kevent.udata field, and requested platform testing #1.
- Robert Haas and Alvaro Herrera questioned whether there was demonstrated benefit sufficient to justify a late release addition, and Thomas chose to leave the work for a future commitfest while fixing event-mask and filter bugs and reporting successful basic tests on four operating systems #4 #5 #8.
- Testing also exposed a separate NetBSD libc collision with PostgreSQL's strtoi function; Thomas supplied a rename, Tom Lane preferred the spelling strtoint, and Alvaro later confirmed its resolution before the next kqueue submission dropped that attachment and adapted to upstream MAXALIGN changes #8 #15 #16 #17.
Sep - Oct 2016 · 4 entries
- Marko Tiikkaja tested and reviewed the patch but questioned its EINTR guarantee; Thomas investigated the kernel implementations and removed the unnecessary exception because the change-only kevent call could not reach the interruptible wait #18 #19.
- Heikki Linnakangas found no benefit on a 40-core FreeBSD instance and returned the patch with feedback, while Andres Freund argued that more connections and future larger wait sets could expose its value; Tom Lane then measured roughly 5–7% gains on macOS, which Thomas and Michael Paquier corroborated #20 #22 #23 #30 #33 #34 #37.
- Keith Fiske's FreeBSD tests showed regressions and briefly stopped the effort, after which Thomas changed postmaster-death detection to EVFILT_PROC following Tom Lane's suggestion in another discussion, retaining the pipe only for an already-dead process #39 #40 #41.
- The process-notification version still produced mixed results: Matteo Beccati measured a NetBSD slowdown, and Thomas's summary of further FreeBSD testing found improvements in some cases but substantial variation and regressions in others #42 #44.
Jun 2017 · 1 entry
- After a long gap, Thomas preserved the work through rebases despite initially declining to resubmit it; later tests on a 64-vCPU FreeBSD machine showed modest gains, and reports of busy-server pipe contention renewed interest #46 #47 #48.
Apr - May 2018 · 1 entry
- Mateusz Guzik supplied stronger evidence from a remote-client FreeBSD benchmark, showing roughly 31% and 54% gains above the server's hardware-thread count; Thomas clarified that the accompanying pdeathsig patch was irrelevant to that workload and revived the kqueue patch for PostgreSQL 12 consideration #50 #51 #52.
Sep - Oct 2018 · 2 entries
- Andres Freund's code review led Thomas to add close-on-exec protection, centralize kevent.udata access, update Windows configuration, and replace the fallback pipe with an event deferred until the next wait #53 #55.
- Further virtual-machine benchmarks again exposed local-client slowdowns, while moving clients to another machine yielded NetBSD gains; Thomas reproduced a FreeBSD regression, Andres suggested wakeup or locking differences, and the work paused for kernel investigation after another configure rebase #56 #59 #60 #61 #62.
Dec 2019 - Feb 2020 · 6 entries
- Rui DeSousa revived the thread with severe poll-related system load on large FreeBSD servers, prompting Thomas to supply master and PostgreSQL 11 rebases; Rui's initial deployment then reported load falling from peaks above 500 to the low twenties #63 #64 #66.
- Peter Eisentraut found a recovery-test hang, which Tom Lane reproduced on macOS and FreeBSD; Thomas traced it to missing exit_on_postmaster_death handling, while Tom also requested removal of kqueue-specific work from the generic wait path and migration to Solution.pm #67 #68 #69 #70 #71.
- Thomas's final submission fixed process-exit handling, kept deferred reporting inside the kqueue implementation, and guarded against postmaster PID reuse; combining registration with waiting and improving WaitEventSet reuse remained separate possible optimizations #72.
- Matteo reported an unkillable NetBSD virtual-machine hang, but Tom completed check-world on macOS, FreeBSD, two NetBSD configurations, and OpenBSD and could not reproduce the hang on bare metal; his small-machine benchmarks still showed no clear gain #73 #74 #76 #77.
- Rui supplied application-load comparisons, Thomas reproduced reduced kernel contention and a throughput increase from 539k to 781k TPS on a 72-vCPU system, and Mark Wong added favorable large-server results; Thomas decided to proceed with WAIT_USE_POLL available as a platform fallback #78 #79 #80 #81.
- Thomas committed #72 as 815c2f0972 after light comment and error-message edits, closing the thread with the feature landed and a commitment to monitor the buildfarm for platform problems #82.