model-writtenrun 4654 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Melanie Plageman reported that passing a fractional vacuum delay to the integer-millisecond WaitLatch() interface truncated sub-millisecond waits to zero, despite configuration displays accepting them #1.
- Thomas Munro proposed accumulating fractional delays, Tom Lane considered where to retain that state, and Stephen Frost emphasized preserving postmaster-death detection when restoring short sleeps #2 #3 #4.
- Tom Lane supplied a static residual accumulator that periodically issued whole-millisecond waits; Melanie Plageman questioned behavior after configuration changes, and Tom explained that the residual remained below one millisecond #7 #10 #11.
- Nathan Bossart identified obsolete sleep-resolution commentary, while Thomas Munro's measurements showed accurate fractional sleeps on several modern systems; Tom Lane defended treating vacuum_cost_delay as an actual duration, and the discussion moved toward higher-resolution latch waits rather than accumulated rounding or feedback control #8 #12 #14 #16 #17.
- Thomas Munro posted experimental microsecond latch APIs, vacuum and walreceiver conversions, and a pg_usleep() fallback with periodic postmaster checks, but noted an unexplained CI failure and limited availability of the required Linux interface #18.
- Thomas Munro then isolated a minimal backpatchable vacuum fix using pg_usleep() and an explicit postmaster check after every sleep, deferring the latch API work to another thread and commitfest; Tom Lane accepted this approach #19 #20.
- Cross-platform measurements prompted Thomas Munro to expand the unchanged vacuum fix into a three-patch series with revised sleep-accuracy commentary and a nanosleep() implementation, retaining early return on interruption because reliable backend responsiveness requires latches #21 #22.
- Nathan Bossart supported nanosleep() but questioned the timing description and interruption explanation; Thomas Munro explained signal races and latch behavior, and Nathan supplied replacement accuracy wording #23 #24 #25.
- The vacuum component of #22 landed with unchanged code as 720de00af4, with backports through PostgreSQL 14.
- The accuracy-comment component of #22 landed using Nathan Bossart's replacement wording and author credit as e4da2a44c1, while its nanosleep() component landed with clearer signal-and-latch commentary as a948e49e2e.
- Thomas Munro confirmed the pushes and explained that NetBSD's extra timer tick accounted for the observed twenty-millisecond sleeps; the final three-patch series was complete, while the broader microsecond latch and walreceiver work remained deferred #26 #19.