diff --git a/quest/m0/nightly/splice-stall-starve.md b/quest/m0/nightly/splice-stall-starve.md new file mode 100644 index 0000000000..99ac913d49 --- /dev/null +++ b/quest/m0/nightly/splice-stall-starve.md @@ -0,0 +1,47 @@ +# [M] The ring6 mesh replay starves no reader under four CPUs + +## Goal + +`taskset -c 0-3 just test mesh-replay --mesh test/mesh-replay/ring6.toml --broadcasts 32`, +with release-built relays, starves no reader, so Nightly's mesh replay stays +green now that no `[[known]]` entry covers a stranded or starved reader. + +Non-goal: the same run on two CPUs, unless the four-CPU run passes and the +two-CPU starvation turns out to be the same cause. + +## Plan + +Found while working [mesh-replay-violations](https://github.com/moq-dev/moq/pull/9236), +on `main` at `9828a20cab`, before +[moq-dev/moq#9231](https://github.com/moq-dev/moq/pull/9231) landed +([evidence](https://github.com/moq-dev/moq/pull/9231#issuecomment-7049596622)): + +- 4 CPUs: `reopen every link of b (1/2)` starved 21 readers (`d` on + `replay/b/*`, 75s). +- 2 CPUs: after `kill and restart b`, reader `a` read nothing on `replay/d/*` + and `replay/e/*` for 150s while `replay/c/0` on the same relay read 10 + groups/s. + +\#9231 finished the splice-stall quest (a lite-07 relay holds a request for a +claim it shows available) and dropped its `[[known]]` entries, so nothing owns +this starvation if it survives that fix. + +Decided 2026-11-20 (maintainer, `/quest-plan` after a `/quest-run`): ✅ rerun +first, then fix (rejected: include the two-CPU case in the goal; plan nothing +and let the Nightly triage decide). + +- Rerun the recipe on `main` once #9231 and #9236 (release-built replay + relays) have both landed, three times. If no reader starves, delete this + quest in a PR that records the runs. +- Otherwise find which relay holds the starved readers' tracks and why it + stops feeding them, from the replay's relay and reader logs, and fix the + cause in moq-net. A starved reader that reads nothing while its relay still + shows the broadcast is the same failure the deleted splice-stall quest + described (`git show 1a34abe437:quest/m2/multicast/splice-stall.md`). +- CPU pressure is how the bug shows itself, not the cause: no new `[[known]]` + entry for load, and no longer deadline. + +Caused by: unknown until reproduced; check whether #9231 is the last change +to the starved path before blaming it. + +Public API: none expected. Wire: none expected.