Repository navigation
quest: plan follow-ups from the m0/m1 spawn run #6186
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Changes from all commits
1a554d2
64e9a6a
98d395e
6558d04
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,22 @@ | ||
| # [S] Audio device tests pass on ARM | ||
|
|
||
| ## Goal | ||
|
|
||
| `just rs audio-devices` passes on the hosted ARM runner, as it does on | ||
| x86_64 Linux, so its nightly lane is a signal rather than a known red. | ||
|
|
||
| ## Plan | ||
|
|
||
| The recipe first ran nightly on 2026-10-20 and failed all five playback | ||
| tests on ARM; nothing owned it | ||
| ([moq-dev/moq#6155](https://github.com/moq-dev/moq/pull/6155)). | ||
| [moq-dev/moq#6175](https://github.com/moq-dev/moq/pull/6175) then fixed | ||
| PulseAudio's 20 ms playback period for these tests. Check a nightly after | ||
| #6175 first: if ARM passes, delete this quest. Otherwise reproduce on ARM | ||
| and fix the cause, not the timing. | ||
|
|
||
| Public API: none. Wire: none. | ||
|
|
||
| ## Related | ||
|
|
||
| - [PulseAudio capture periods](/quest/m2/pulseaudio-capture-period.md) - the capture half of the same period fix |
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,36 @@ | ||
| # [S] Binding auth polish | ||
|
|
||
| ## Goal | ||
|
|
||
| Three gaps left by [moq-dev/moq#6137](https://github.com/moq-dev/moq/pull/6137) | ||
| close: OBS reconnects with the token the user last presented, moq-ffi wraps | ||
| one auth handle type instead of two, and Swift exposes grants as an | ||
| `AsyncSequence`. | ||
|
|
||
| ## Plan | ||
|
|
||
| Decided 2026-10-21 with the maintainer: one m1 quest. | ||
|
|
||
| - **OBS reconnect.** A reconnect dials the URL's original token, not the one | ||
| presented mid-stream through `MoQOutput::PresentToken`. Once tokens can | ||
| ride SETUP, the presented token replaces it on reconnect. | ||
| - **FFI dispatch.** `rs/moq-ffi/src/auth.rs` wraps both | ||
| `moq_net::auth::Handle` (server sessions) and moq-tokio's | ||
| `connection::auth::Handle` (client sessions). With one handle spanning a | ||
| connection, collapse the enum. #6137 decided whichever of it and | ||
| [moq-dev/moq#6109](https://github.com/moq-dev/moq/pull/6109) lands second | ||
| does this; skip it if that already happened. | ||
| - **Swift.** Grants arrive as an `AsyncSequence`, once the generated types | ||
| are `Sendable`. If they still are not, say why in the PR rather than | ||
| forcing it. | ||
|
|
||
| Public API: Swift additive. Wire: none. | ||
|
|
||
| ## Required | ||
|
|
||
| - [Auth outlives its session](/quest/m1/auth/issuer.md) - one handle per connection, which the enum collapse needs | ||
| - [Token in band](/quest/m1/auth/token-in-band.md) - tokens that ride SETUP, which the OBS reconnect needs | ||
|
|
||
| ## Related | ||
|
|
||
| - [First dial failures in Go](/quest/m2/go-first-dial-failure.md) - the m2 half of this binding cleanup |
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,29 @@ | ||
| # [M] Binding checks cost less | ||
|
|
||
| ## Goal | ||
|
|
||
| An `rs/moq-ffi` change's binding checks finish markedly faster, with the | ||
| same coverage. `moq.cpp` compiles once per C++ standard instead of once per | ||
| consumer build, and Kotlin reuses a warm Gradle daemon. The PR reports the | ||
| before and after. | ||
|
|
||
| ## Plan | ||
|
|
||
| From the CI audit in [moq-dev/moq#6155](https://github.com/moq-dev/moq/pull/6155), | ||
| decided 2026-10-21 with the maintainer as an m1 quest. Measured there: C++ | ||
| plus the OBS plugin take about 4.5 min because the generated `moq.cpp` | ||
| compiles six times at `-O0`, and Kotlin spends about 65 s of its 30 to 110 s | ||
| starting a cold Gradle daemon. Those two are most of the binding time. | ||
|
|
||
| - Build `moq.cpp` into one library per C++ standard and link every consumer | ||
| against it. | ||
| - Keep a Gradle daemon warm where the checks run, without letting it leak | ||
| state between builds. | ||
| - Measure locally and on the queue runner, before and after. | ||
|
|
||
| Public API: none. Wire: none. | ||
|
|
||
| ## Related | ||
|
|
||
| - [CI workflow audit](/quest/m1/audit-ci.md) - judges every job's cost against its signal | ||
| - [CI cost and latency](/quest/m0/ci/README.md) - runner and cache infrastructure |
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,33 @@ | ||
| # [XS] CI housekeeping | ||
|
|
||
| ## Goal | ||
|
|
||
| Three small CI changes land together: | ||
|
|
||
| - The merge queue's nextest runs with `NEXTEST_TEST_THREADS=24`. | ||
| - Nightly `benchmark-site` and smoke `relays-report` delete an earlier | ||
| attempt's stale `github-pages` artifact before uploading, as interop's | ||
| report job does. | ||
| - `.github/actionlint.yaml` drops its ignore for the `queue` key once an | ||
| actionlint release after 1.7.12 is in the dev shell. | ||
|
|
||
| ## Plan | ||
|
|
||
| - **Threads.** Recommended in | ||
| [moq-dev/moq#6174](https://github.com/moq-dev/moq/pull/6174): the suite | ||
| needs about 200 CPU-seconds, and most slots wait on sockets, subprocesses, | ||
| and timeouts. Pinned to 8 CPUs it took 105 to 113 s at 8 threads and 47 s | ||
| at 24. Set it for the queue job only, after a few loaded | ||
| `just check --all` runs at 24; a flake there points at a test to fix. | ||
| - **Pages artifact** and **actionlint** both come from | ||
| [moq-dev/moq#6180](https://github.com/moq-dev/moq/pull/6180). A re-run of | ||
| either nightly job fails on the stale artifact today. The actionlint fix, | ||
| [rhysd/actionlint#661](https://github.com/rhysd/actionlint/pull/661), is | ||
| merged but unreleased; if no release reaches nixpkgs by the time the rest | ||
| is ready, land the rest and leave this bullet. | ||
|
|
||
| Public API: none. Wire: none. | ||
|
|
||
| ## Related | ||
|
|
||
| - [CI workflow audit](/quest/m1/audit-ci.md) - the broader per-job review | ||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,24 @@ | ||
| # [S] The GOAWAY drain deadline starts at the trigger | ||
|
|
||
| ## Goal | ||
|
|
||
| A session's GOAWAY drain deadline counts from when the drain was triggered, | ||
| not from when the GOAWAY reached the wire, in moq-net (lite and IETF) and | ||
| js/net. A peer that stops reading can no longer stall the GOAWAY write and | ||
| hold the session open past its deadline. | ||
|
|
||
| ## Plan | ||
|
|
||
| [moq-dev/moq#6144](https://github.com/moq-dev/moq/pull/6144) found that | ||
| moq-net arms its GOAWAY timer only after the message is written. A peer that | ||
| keeps the transport open but stops reading stalls that write forever, on lite | ||
| (open, flush, FIN ack) and IETF (`encode_message(..).await`). #6144 fixed the | ||
| relay only: its drain observer closes each session at the absolute deadline. | ||
| Library users of moq-net and js/net still hang. | ||
|
|
||
| - Arm the deadline at the trigger and race the write against it, so the | ||
| relay's own deadline close becomes a backstop rather than the only bound. | ||
| - Port #6144's `a_stalled_goaway_write_is_still_closed_at_the_deadline` to | ||
| moq-net and js/net, on a mocked clock, for lite and IETF. | ||
|
Comment on lines
+21
to
+22
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more.
At c51941f, js/net has no production outbound GOAWAY path to which this regression can be ported: AGENTS.md reference: AGENTS.md:L62-L62 Useful? React with 👍 / 👎. |
||
|
|
||
| Public API: behavior only. Wire: none. | ||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,31 @@ | ||
| # [S] Timed moq-transport tracks cross a lite-07 mesh | ||
|
|
||
| ## Goal | ||
|
|
||
| A timed track ingested over moq-transport keeps its timestamps across a | ||
| lite-07 hop to a peer that negotiated the `ietf` extension, so a relay mesh | ||
| carries them end to end. A peer without the extension still gets the track | ||
| untimed. | ||
|
|
||
| ## Plan | ||
|
|
||
| Decided 2026-10-21 with the maintainer: m1, under moq-transport passthrough. | ||
|
|
||
| [moq-dev/moq#6150](https://github.com/moq-dev/moq/pull/6150) marks a timed | ||
| track received over moq-transport with the crate-private | ||
| `track::Info::unchecked` and serves it untimed to moq-lite, so its | ||
| timestamps stop at the first lite hop. | ||
| [moq-dev/moq#6107](https://github.com/moq-dev/moq/pull/6107) gives every | ||
| moq-transport track `track::Info::ietf()`, its ingress version. Derive | ||
| `unchecked` from `ietf()` instead of keeping a second flag, and let the lite | ||
| publisher serve such a track timed to an `ietf` extension peer, which knows | ||
| the timestamps came from moq-transport and are unchecked. | ||
|
|
||
| Mirror it in js/net, where `Info.unchecked` is `@internal`. If the timed | ||
| encoding on that hop needs a wire change, update | ||
| `drafts/draft-lcurley-moq-lite.md` (or the extension's draft) in the same PR. | ||
|
|
||
| ## Required | ||
|
|
||
| - [Received timestamps are media time](/quest/m1/lite-time-provenance.md) - adds `unchecked` and the untimed serving this changes | ||
| - [Ingress-version provenance](/quest/m1/ietf-extension/lite.md) - adds `Info::ietf()`, which `unchecked` derives from |
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,27 @@ | ||
| # [S] Interop covers a clean broadcast finish | ||
|
|
||
| ## Goal | ||
|
|
||
| The interop matrix has a media "finish" case: a publisher in each | ||
| implementation finishes its broadcast cleanly, and every subscriber drains | ||
| the tail and then sees the broadcast end, rather than an abort. A regression | ||
| in any implementation's linger shows up as a red cell. | ||
|
|
||
| ## Plan | ||
|
|
||
| From [moq-dev/moq#6151](https://github.com/moq-dev/moq/pull/6151), which | ||
| gave js/net `Broadcast.Producer.finish()`. Rust, JS, and the bindings each | ||
| have a finish once their linger quests land. Drive it from the existing | ||
| interop clients rather than adding a new harness, and assert both halves: | ||
| the late tail arrives and the end is clean. | ||
|
|
||
| Public API: none. Wire: none. | ||
|
|
||
| ## Required | ||
|
|
||
| - [Binding linger](/quest/m1/linger-ffi.md) - bindings finish a broadcast, so the binding interop clients can | ||
| - [JS linger](/quest/m1/linger-js.md) - js/net finishes a broadcast, so the JS interop client can | ||
|
|
||
| ## Related | ||
|
|
||
| - [Interop report](/quest/m1/interop-report.md) - where the new case's cells show up |
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,24 @@ | ||
| # [S] A parked JS stream open gives up when its group expires | ||
|
|
||
| ## Goal | ||
|
|
||
| A js/net publisher's pending stream open ends as soon as its group expires | ||
| or is evicted, as Rust's parked machines do with | ||
| [moq-dev/moq#6140](https://github.com/moq-dev/moq/pull/6140). A stalled | ||
| downstream no longer makes a JS publisher hold one pending open per group. | ||
|
|
||
| ## Plan | ||
|
|
||
| Decided 2026-10-21 with the maintainer: m1. JS lite and IETF open each | ||
| group's stream with `waitUntilAvailable: false`, so an over-limit open fails | ||
| at once on a transport that rejects it. A WebTransport implementation that | ||
| parks it instead leaves one pending open per group, bounded only by the 10 s | ||
| `OPEN_TIMEOUT_MS` or the subscriber leaving. | ||
|
|
||
| - Cancel the pending open when its group expires or is evicted, mirroring | ||
| Rust's `poll_abandoned`. Then decide whether `OPEN_TIMEOUT_MS` still earns | ||
| its place, and delete it if not. | ||
| - Mocked-time test: a transport that parks opens, a group that expires while | ||
| its open is parked, and the open released at expiry. | ||
|
|
||
| Public API: none. Wire: none. |
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,20 @@ | ||
| # [XS] JS redirect keys strip URL userinfo | ||
|
|
||
| ## Goal | ||
|
|
||
| js/net's reconnect (`js/net/src/connection/reload.ts`) keys an endpoint | ||
| without its URL userinfo, as Rust does, so a redirect loop is detected | ||
| whatever credentials the URL carries and none of them end up in the key. | ||
|
|
||
| ## Plan | ||
|
|
||
| `endpoint()` clears only `search` and `hash` today. Rust's redirect-key rule | ||
| from [moq-dev/moq#6162](https://github.com/moq-dev/moq/pull/6162) also drops | ||
| userinfo; match it exactly, and test both a userinfo URL and a redirect loop | ||
| that differs only in userinfo. | ||
|
|
||
| Public API: none. Wire: none. | ||
|
|
||
| ## Required | ||
|
|
||
| - [Rust redirect loops](/quest/m1/tokio-redirect-loop.md) - lands the Rust rule this mirrors |
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,21 @@ | ||
| # [XS] JS answers SUBSCRIBE_START like Rust | ||
|
|
||
| ## Goal | ||
|
|
||
| The js/net lite publisher (`js/net/src/lite/publisher.ts`) answers | ||
| SUBSCRIBE_START with the lower of the first group it serves and where its | ||
| source starts, as Rust does. A JS publisher sending a descending backlog no | ||
| longer names its newest group, which a resumed idle copy misjudges as a gap. | ||
|
|
||
| ## Plan | ||
|
|
||
| Found in [moq-dev/moq#6183](https://github.com/moq-dev/moq/pull/6183). | ||
| Decided 2026-10-21: match Rust in m1. A test publishes a descending backlog | ||
| from JS and checks the START names the oldest group. | ||
|
|
||
| Public API: none. Wire: none, only the value JS sends. | ||
|
|
||
| ## Related | ||
|
|
||
| - [lite-05/06 gap judgement](/quest/m2/lite-legacy-gap.md) - the same misjudgement on versions whose START cannot say it | ||
| - [Resume through an outage](/quest/m1/resume-outage.md) - the resume work that found it |
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,31 @@ | ||
| # [M] moq-net internals after transport erasure | ||
|
|
||
| ## Goal | ||
|
|
||
| With the transport boxed, the lite publisher stops cloning the session and | ||
| group per group, and the lite and IETF internals lose the transport type | ||
| parameter they keep only for test mocks. A `just bench-runtime` run on | ||
| io_uring measures the cost of the `Local` thread check. | ||
|
|
||
| ## Plan | ||
|
|
||
| Follow-ups from [moq-dev/moq#6176](https://github.com/moq-dev/moq/pull/6176), | ||
| decided 2026-10-21 with the maintainer. | ||
|
|
||
| - **Per-group clones.** The lite publisher clones its session and group for | ||
| each group it serves. Borrow or share them instead, and show the win on the | ||
| session delivery benches. | ||
| - **Type parameter.** Lite and IETF internals stay generic over the | ||
| transport only so tests can inject mocks. Mock behind the boxed transport | ||
| instead, and drop the parameter. | ||
| - **`Local` check.** Run `just bench-runtime` on io_uring and report what the | ||
| `Local` thread check costs; act on it only if it shows. | ||
|
|
||
| Dropped: renaming the over-long `session_delivery_*` bench IDs, since it | ||
| loses their history for a cosmetic gain. | ||
|
|
||
| Public API: none. Wire: none. | ||
|
|
||
| ## Required | ||
|
|
||
| - [moq-net compiles once](/quest/m1/net-erase-transport.md) - boxes the transport these cleanups build on |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Split the nextest concurrency tuning, Pages artifact cleanup, and actionlint configuration into independent quests. They have separate validation and readiness conditions, and the actionlint item is explicitly blocked on an external release; the later plan to “land the rest and leave this bullet” would leave this quest only partially completed rather than keeping each implementation PR focused and independently landable.
AGENTS.md reference: AGENTS.md:L22-L24
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Declining: the CI grouping was the maintainer's call in the 2026-10-21
/quest-planinterview (R3 CI, "4 quests in m1"). The three changes are each a few lines, and the Plan already handles the external actionlint release: if it has not reached the dev shell, the rest lands and the quest remains as that single waiting bullet, which is the outside-party shape the quest guide asks for.(Written by Claude Opus 5.5)