From 7ba4b7a0ef73ae4a96cab457ae9158b6b81db0f1 Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Thu, 22 Oct 2026 08:54:37 -0700 Subject: [PATCH 1/3] quest: plan the repository improvement audit's findings Six findings from the 2026-10-22 audit become quests: tooling tests in the merge gate, Signals structural equality, the SendBuffer lookup sweep (m1), and fuzz corpus persistence, crates.io trusted publishing, and Firefox publishing coverage (m2). Binary attestations were declined. Co-Authored-By: Claude Opus 5.5 --- quest/m1/README.md | 3 +++ quest/m1/gh-test-gate.md | 27 ++++++++++++++++++++++ quest/m1/quic-send-lookup.md | 34 ++++++++++++++++++++++++++++ quest/m1/signals-equality.md | 28 +++++++++++++++++++++++ quest/m2/README.md | 4 ++++ quest/m2/crates-trusted-publisher.md | 19 ++++++++++++++++ quest/m2/fuzz-corpus.md | 26 +++++++++++++++++++++ quest/m2/publish-firefox.md | 33 +++++++++++++++++++++++++++ quest/m2/quic-egress-profile.md | 1 + quest/m2/release-rs-oidc.md | 24 ++++++++++++++++++++ 10 files changed, 199 insertions(+) create mode 100644 quest/m1/gh-test-gate.md create mode 100644 quest/m1/quic-send-lookup.md create mode 100644 quest/m1/signals-equality.md create mode 100644 quest/m2/crates-trusted-publisher.md create mode 100644 quest/m2/fuzz-corpus.md create mode 100644 quest/m2/publish-firefox.md create mode 100644 quest/m2/release-rs-oidc.md diff --git a/quest/m1/README.md b/quest/m1/README.md index c46e8eec290..30c0198c3c2 100644 --- a/quest/m1/README.md +++ b/quest/m1/README.md @@ -95,6 +95,7 @@ lite-07 stream type here. - [Audio jitter target](/quest/m1/audio-jitter-target/README.md) - the playout target is default in both languages; a manual browser proof and a native trace replay remain - [A/V clock](/quest/m1/av-clock.md) - the audio playhead drives Sync.reference while audio plays, through per-track sync handles - [Publish channel count](/quest/m1/publish-audio-channel-count.md) - forcing a channel count on an Audio.Capture stops costing the subscriber gaps of silence, and the media fixture's cold-start audio-tone flake is fixed at its cause +- [Signals notify on every structural change](/quest/m1/signals-equality.md) - a renamed `undefined` key or a changed sparse array length notifies, checked by a fast-check model of the notification contract - [Container reads drop backfill](/quest/m1/read-backfill.md) - the JS container consumer drops groups below its read cursor like Rust - [qmux no-wait opens](/quest/m1/qmux-no-wait.md) - the in-tree @moq/qmux rejects an over-limit create when waitUntilAvailable is false, like Chrome - [TS fidelity and recovery](/quest/m1/ts-reliability/README.md) - both modes have explicit profiles, automated validation, loss tests, and a dedicated guide @@ -107,10 +108,12 @@ lite-07 stream type here. - [Track live signal](/quest/m1/track-live.md) - a reader can wait for its track's feed to be live, so a player skips old cached groups while a fresh upstream subscription starts - [moq-net compiles once](/quest/m1/net-erase-transport.md) - the session drivers stop monomorphizing per transport and per crate, saving several MiB of the relay, FFI library, `libmoq` and CLI - [Audio devices on ARM](/quest/m1/audio-devices-arm.md) - `just rs audio-devices` passes on the hosted ARM runner +- [Tooling tests gate every merge](/quest/m1/gh-test-gate.md) - every queue Check runs `just gh test`, so a change to the CI selector or release tooling cannot merge without its own tests - [CI housekeeping](/quest/m1/ci-housekeeping.md) - validate queue nextest at 24 threads and remove the actionlint queue ignore after its release - [Interop audio delivery](/quest/m1/interop-audio.md) - every interop subscriber decodes and checks the publishers' 440 Hz tone, and the Rust publisher sends Opus and AAC - [Native audio decode conformance](/quest/m1/audio-decode-conformance.md) - Opus and AAC decode through every moq-audio backend, moq-c, and moq-ffi is checked against known tones and clips, like moq-video's suite - [moq-net internals](/quest/m1/net-internals.md) - no per-group session clones, no transport type parameter in lite and IETF internals, and the `Local` check measured +- [Send buffer lookups stay linear](/quest/m1/quic-send-lookup.md) - a segment-count sweep decides whether `SendBuffer::get`'s front-to-back scan needs a cursor - [Main is no slower than release](/quest/m1/release-perf/README.md) - every benchmark where main is slower than release is fixed or accepted with a reason, checked by a final paired run before the cut - [Subscribe update churn](/quest/m1/subscribe-update-churn.md) - a viewer sends SUBSCRIBE_UPDATE only on a meaningful change, measured after the publisher stall fix - [Interop coverage](/quest/m1/interop-coverage/README.md) - the riskiest moq-transport features moq-net exposes get cross-implementation tests over the even drafts 14 to 22 diff --git a/quest/m1/gh-test-gate.md b/quest/m1/gh-test-gate.md new file mode 100644 index 00000000000..ef1865d8f10 --- /dev/null +++ b/quest/m1/gh-test-gate.md @@ -0,0 +1,27 @@ +# [XS] Tooling tests gate every merge + +## Goal + +Every merge-queue Check runs `just gh test`, the tests of the CI selector, +release, size, and benchmark tooling, so a change to the machinery that picks +coverage cannot merge without its own tests. Nightly keeps running it as part +of the full gate. + +## Plan + +Decided 2026-10-22 from the repository improvement audit: run it on every +Check rather than select it by path. The whole recipe took about 10 s locally +(about 3% of a median 5.5-minute queue entry), and path selection would let a +broken selector skip the very tests that catch it. + +- The run that planned this found `gh test` red on `main`: + `sh/gh/setup.test.ts` expects Check to be the last step of its job, and + [moq-dev/moq#6155](https://github.com/moq-dev/moq/pull/6155) appended + "Upload build timings" after it. It stayed unnoticed because no PR runs the + suite. [moq-dev/moq#6325](https://github.com/moq-dev/moq/pull/6325) fixes + that fixture; land after it, or fix it here if it has not merged. +- Run it from the `gh` module's check in `sh/dispatch.sh` (or an always-selected + module) and add any tools it needs (`python3`) to that module's tool list. +- Update the recipe comment in `.github/justfile` and the `gh test` comment in + Nightly's gate matrix, which both say no pull request runs it. Keep it in + Nightly only if the full gate there does not already cover it. diff --git a/quest/m1/quic-send-lookup.md b/quest/m1/quic-send-lookup.md new file mode 100644 index 00000000000..5829c6b9a1b --- /dev/null +++ b/quest/m1/quic-send-lookup.md @@ -0,0 +1,34 @@ +# [S] Send buffer lookups stay linear + +## Goal + +A measured verdict on `moq-quic`'s `SendBuffer::get` cost as retained +segments grow, and, if the cost grows with segment count, a lookup whose +per-packet cost does not. + +## Plan + +Decided 2026-10-22 from the repository improvement audit; m1 because it sits +on every relay's egress path and is cheap to settle before the cut. + +`SendBuffer::get` (`rs/moq-quic/src/connection/send_buffer.rs`) walks the +retained segments from the front on every call, and packetization calls it +repeatedly, so sending N retained segments in order visits O(N^2) segments. +The audit's loop replay counted 8,256 visits for 128 segments of 64 bytes and +524,800 for 1,024. With a 1 MB window of ~1 KB writes, each packet walks about +500 segments. Those are visit counts, not measured time. The send window +already bounds retained bytes, so this is CPU only, not memory. + +- Measure first: sweep the existing `rs/moq-quic/benches/streams.rs` over + segment size and count, with ACKs delayed so segments stay retained. Count + instructions as the [performance audit](/quest/m1/audit-perf.md) does. +- Only if the sweep shows a slope, compare a sequential cursor (the next + unsent offset's segment index, with a fallback scan for retransmits) + against indexed segments (offsets searchable by binary search). Keep + each write as the caller's `Bytes`, with no coalescing copy. +- If the sweep is flat within noise, record that here and delete the quest. + +## Related + +- [Egress profile](/quest/m2/quic-egress-profile.md) - the broader send-path profile; its `SendBuffer` pool lever shares this type +- [QUIC I/O boundary](/quest/m2/quic-io-boundary.md) - reworks the same transmit buffers diff --git a/quest/m1/signals-equality.md b/quest/m1/signals-equality.md new file mode 100644 index 00000000000..ec9de399ca3 --- /dev/null +++ b/quest/m1/signals-equality.md @@ -0,0 +1,28 @@ +# [S] Signals notify on every structural change + +## Goal + +A `Signal` holding a plain object or array notifies whenever a `set` changes +its structure, including a renamed key with an `undefined` value +(`{left: undefined}` to `{right: undefined}`) and an array whose length +changes without enumerable entries (`Array(1)` to `Array(2)`). Class +instances keep identity equality and notification coalescing is unchanged. + +## Plan + +Decided 2026-10-22 from the repository improvement audit, which reproduced +both missed notifications against `isEqual` in `js/signals/src/index.ts`: it +compares enumerable key counts and values, but not key membership or array +length. + +- Fix `isEqual`: compare array `length`, and require each key of one side to + be an own key of the other. +- Add a fast-check model-based test (fast-check is already used by + `js/json`'s snapshot properties): commands for subscribe, set (including the + cases above), microtask flush, unsubscribe, and close, checked against a + small reference model of which subscribers fire and when. It covers the + notification and lifetime contract generally, not only this bug. +- Out of scope: an immediately disposed `Signal.watch()` still runs its queued + initial callback. No consumer depends on either behavior, and changing it is + an API timing decision; leave it unless the model test shows a consumer + hazard. diff --git a/quest/m2/README.md b/quest/m2/README.md index d5d9f8cc95f..c73f056e2e2 100644 --- a/quest/m2/README.md +++ b/quest/m2/README.md @@ -67,6 +67,7 @@ ranked just below deterministic TS legs, which it requires. - [Media stats](/quest/m2/stats/README.md) - publishers announce a stats track in the catalog, viewers answer a soliciting catalog through a per-catalog `.echo` broadcast, and one model turns both into a health verdict and a preflight report - [Group storage benchmark](/quest/m2/group-storage-bench.md) - numbers pick the append-only log, a ring buffer, or sorted runs for `GroupState` before either lands - [Perf](/quest/m2/perf/README.md) - eliminate measured hot-path costs across moq-uring, kio, and the moq-net model +- [Nightly fuzzing keeps its corpus](/quest/m2/fuzz-corpus.md) - each fuzz target resumes from the previous Nightly's minimized corpus instead of starting from seeds - [Nightly benchmark regressions](/quest/m2/bench-ci.md) - each nightly compares the Criterion benches against the previous nightly's commit, publishes it to the git.moq.dev report, and alerts on a regression - [Relay session bench](/quest/m2/bench-relay.md) - the same scenario through moq-relay's own connection handling - [Transport x version benches](/quest/m2/bench-transport-sweep.md) - the session benches sweep transports as well as versions @@ -94,11 +95,14 @@ ranked just below deterministic TS legs, which it requires. - [The catalog publish path compiles once](/quest/m2/catalog-publish-once.md) - the catalog producer and its JSON diff codegen in `moq-mux` instead of once per importer crate, about 1 MiB of the CLI - [Zero-copy byte returns](/quest/m2/ffi-bytes-return.md) - returned payloads reach the host without a copy, if the payload numbers justify it - [Namespaced external types](/quest/m2/uniffi-namespaces.md) - the UniFFI fork lets two crates export the same type name, reopening the per-layer crate split +- [crates.io trusts the workflow](/quest/m2/crates-trusted-publisher.md) - maintainer action: every published crate gets a trusted publisher for `release-rs.yml` +- [Rust releases publish through OIDC](/quest/m2/release-rs-oidc.md) - release-plz publishes with trusted publishing and `CARGO_REGISTRY_TOKEN` is revoked - [Go module proxy](/quest/m2/go-module-proxy.md) - `moq.dev/moq-ffi` ships from a static GOPROXY we host, so its staticlibs stop growing git history by ~61 MiB per release - [Go per-platform lib modules](/quest/m2/go-ffi-platform-modules.md) - a Go build downloads only the host's staticlib, and `go mod vendor` works - [Go first dial failure](/quest/m2/go-first-dial-failure.md) - Go reports why a first dial fails, and the Kotlin and Dart docs name `client.failure()` correctly - [Dart on iOS](/quest/m2/dart-ios.md) - prove the shipped iOS native asset actually loads on a device, which no CI can - [Mobile completion](/quest/m2/mobile-completion.md) - verify the Swift and Kotlin bindings over Rust-owned capture +- [Browser publishing paths under test](/quest/m2/publish-firefox.md) - the media harness publishes from Firefox too, covering the worker capture and video-element fallback Chromium never takes - [Bindings in the interop matrix](/quest/m2/interop-bindings.md) - Kotlin, Swift, and Dart publish and subscribe against every other interop client - [WebGPU HDR](/quest/m2/webgpu-hdr.md) - HDR renditions play as HDR where the browser and display can show it, and tone-map to SDR elsewhere - [One port](/quest/m2/one-port/README.md) - one UDP port carries QUIC, SRT, and WebRTC media through an embedder hook, and one TCP port carries HTTP, RTMP, and RTMPS diff --git a/quest/m2/crates-trusted-publisher.md b/quest/m2/crates-trusted-publisher.md new file mode 100644 index 00000000000..5d909f0c27f --- /dev/null +++ b/quest/m2/crates-trusted-publisher.md @@ -0,0 +1,19 @@ +# [XS] crates.io trusts the Rust release workflow + +## Goal + +Condition: every crate that `release-rs.yml` publishes has a crates.io trusted +publisher for `moq-dev/moq`, workflow `release-rs.yml` (and its environment, +if the job uses one). This is a maintainer action in each crate's crates.io +settings and blocks +[Rust releases publish through OIDC](/quest/m2/release-rs-oidc.md). + +Check: for each crate that release-plz publishes, its crates.io settings page +lists that configuration. Delete this quest once every crate has one. + +## Plan + +Decided 2026-10-22: the human step is split from the workflow change so it +stays visible as ready work. A crate that has never been published still needs +its first publish done by hand before it can have a trusted publisher; list any +such crate here. diff --git a/quest/m2/fuzz-corpus.md b/quest/m2/fuzz-corpus.md new file mode 100644 index 00000000000..679aa8b9ae7 --- /dev/null +++ b/quest/m2/fuzz-corpus.md @@ -0,0 +1,26 @@ +# [S] Nightly fuzzing keeps its corpus + +## Goal + +Each Nightly fuzz target starts from the corpus the previous run evolved, so +coverage found in one 300-second run compounds instead of being discarded +with the runner. A second run demonstrably loads the first run's discoveries. + +## Plan + +Decided 2026-10-22 from the repository improvement audit. Nightly runs twelve +targets for 300 s each on fresh hosted runners and uploads only crash +artifacts. `sh/rs/fuzz.sh` already keeps a writable +`rs//fuzz/corpus/` separate from generated seeds, so only the +workflow has to persist it. + +- Store it in `actions/cache`: restore by a per-target key prefix, then save + under a run-unique key after the run. Nightly runs daily, so the 7-day + eviction never fires. Rejected: chaining artifacts between runs (needs + cross-run lookup) and committing to a branch (needs a writer job with push + credentials). +- Minimize with `cargo fuzz cmin` before saving, so the corpus stays bounded. +- Save only after a successful run; a crash input already reaches the + failure artifact. +- Verify it: the second Nightly's log reports more initial corpus entries than + seeds alone (libFuzzer's `INITED` line), for at least one target. diff --git a/quest/m2/publish-firefox.md b/quest/m2/publish-firefox.md new file mode 100644 index 00000000000..1c36f285bf0 --- /dev/null +++ b/quest/m2/publish-firefox.md @@ -0,0 +1,33 @@ +# [M] Browser publishing paths under test + +## Goal + +The interop media harness publishes from Firefox as well as Chromium, so the +production capture paths Chromium never takes run in CI: the worker +`MediaStreamTrackProcessor` and the video-element fallback. + +## Plan + +Decided 2026-10-22 from the repository improvement audit. The harness +(`test/interop/clients/js/media.ts`) always launches Chromium for the +publisher, and its capability probe requires a window-level +`MediaStreamTrackProcessor`. `js/publish/src/video/processor.ts` uses that +only on Chromium; elsewhere it spawns a capture worker and falls back to a +video element. + +- Add a selectable publisher engine to the harness, and make the capability + probe check the path production would take (worker support, not the window + global). +- Cover the worker path end to end in Firefox: frame transfer, source + replacement, cancellation with a read pending, and that every `VideoFrame` + is closed. +- Cover the video-element fallback by forcing it (for example, a worker + refused by CSP), in either engine. +- Firefox only, on Linux. Playwright WebKit and a macOS Safari lane stay with + the m3 Safari work; Playwright's WebKit is not Safari. +- The test-runner option is the only interface change; document it in + `test/interop/README.md`. + +## Related + +- [Safari WebTransport](/quest/m3/safari-webtransport.md) - the later Safari lane diff --git a/quest/m2/quic-egress-profile.md b/quest/m2/quic-egress-profile.md index 3d0adabca7a..192a2f3c12f 100644 --- a/quest/m2/quic-egress-profile.md +++ b/quest/m2/quic-egress-profile.md @@ -51,6 +51,7 @@ hosts need. ## Related +- [Send buffer lookups stay linear](/quest/m1/quic-send-lookup.md) - settles `SendBuffer::get`'s per-packet scan first - [QUIC I/O boundary](/quest/m2/quic-io-boundary.md) - redesigns the transmit buffers the pool lever would feed - [#3201](/quest/m3/3201-moq-uring-use-sendmsg-zc-for-large-udp-gso-trains.md) - zero-copy on the same trains diff --git a/quest/m2/release-rs-oidc.md b/quest/m2/release-rs-oidc.md new file mode 100644 index 00000000000..f7417e37723 --- /dev/null +++ b/quest/m2/release-rs-oidc.md @@ -0,0 +1,24 @@ +# [S] Rust releases publish through OIDC + +## Goal + +`release-rs.yml` publishes crates with crates.io trusted publishing through +release-plz's native OIDC support, and the long-lived `CARGO_REGISTRY_TOKEN` +secret is removed and revoked. + +## Plan + +Decided 2026-10-22 from the repository improvement audit. release-plz (0.3.172 +in the dev shell) exchanges the GitHub OIDC token itself, so no separate +authentication Action is needed. + +- Grant the publishing job `id-token: write` and drop `CARGO_REGISTRY_TOKEN` + from its environment. Keep the existing release guards. +- Verify on the next approved release that every crate published, then delete + the repository secret and revoke the token on crates.io. +- This removes a reusable credential; it does not protect against a + compromised authorized workflow. + +## Required + +- [crates.io trusts the workflow](/quest/m2/crates-trusted-publisher.md) - every published crate has the trusted publisher configured From e021b325fceda50965789d769f6fb46287142929 Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Thu, 22 Oct 2026 09:00:46 -0700 Subject: [PATCH 2/3] quest: keep release-rs write permissions beside id-token Co-Authored-By: Claude Opus 5.5 --- quest/m2/release-rs-oidc.md | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/quest/m2/release-rs-oidc.md b/quest/m2/release-rs-oidc.md index f7417e37723..57a020659a4 100644 --- a/quest/m2/release-rs-oidc.md +++ b/quest/m2/release-rs-oidc.md @@ -13,7 +13,10 @@ in the dev shell) exchanges the GitHub OIDC token itself, so no separate authentication Action is needed. - Grant the publishing job `id-token: write` and drop `CARGO_REGISTRY_TOKEN` - from its environment. Keep the existing release guards. + from its environment. Keep the existing release guards. A job-level + `permissions` block replaces the workflow's `contents: write` and + `pull-requests: write`, which release-plz needs for tags, releases, and + PRs, so add `id-token: write` beside them or repeat all three. - Verify on the next approved release that every crate published, then delete the repository secret and revoke the token on crates.io. - This removes a reusable credential; it does not protect against a From 3c5a760fb8659954215cdf46709da4f68354f9a4 Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Thu, 22 Oct 2026 10:23:07 -0700 Subject: [PATCH 3/3] quest: promote the survey's send-fragment quest to m1 The performance survey (#6318) landed the same SendBuffer finding with a fuller plan, so drop the audit's duplicate and move that quest to m1. Co-Authored-By: Claude Opus 5.5 --- quest/m1/README.md | 2 +- .../quic-send-fragments.md | 5 +++ quest/m1/quic-send-lookup.md | 34 ------------------- quest/m2/performance-survey/README.md | 1 - quest/m2/quic-egress-profile.md | 2 +- 5 files changed, 7 insertions(+), 37 deletions(-) rename quest/{m2/performance-survey => m1}/quic-send-fragments.md (87%) delete mode 100644 quest/m1/quic-send-lookup.md diff --git a/quest/m1/README.md b/quest/m1/README.md index 1dc772df7cc..30eae6d4e9c 100644 --- a/quest/m1/README.md +++ b/quest/m1/README.md @@ -112,7 +112,7 @@ lite-07 stream type here. - [Interop audio delivery](/quest/m1/interop-audio.md) - every interop subscriber decodes and checks the publishers' 440 Hz tone, and the Rust publisher sends Opus and AAC - [Native audio decode conformance](/quest/m1/audio-decode-conformance.md) - Opus and AAC decode through every moq-audio backend, moq-c, and moq-ffi is checked against known tones and clips, like moq-video's suite - [moq-net internals](/quest/m1/net-internals.md) - no per-group session clones, no transport type parameter in lite and IETF internals, and the `Local` check measured -- [Send buffer lookups stay linear](/quest/m1/quic-send-lookup.md) - a segment-count sweep decides whether `SendBuffer::get`'s front-to-back scan needs a cursor +- [Fragmented QUIC sends](/quest/m1/quic-send-fragments.md) - a chunk-size by retained-flight sweep decides whether `SendBuffer::get`'s front-to-back scan needs a cursor or index - [Main is no slower than release](/quest/m1/release-perf/README.md) - every benchmark where main is slower than release is fixed or accepted with a reason, checked by a final paired run before the cut - [Subscribe update churn](/quest/m1/subscribe-update-churn.md) - a viewer sends SUBSCRIBE_UPDATE only on a meaningful change, measured after the publisher stall fix - [Interop coverage](/quest/m1/interop-coverage/README.md) - the riskiest moq-transport features moq-net exposes get cross-implementation tests over the even drafts 14 to 22 diff --git a/quest/m2/performance-survey/quic-send-fragments.md b/quest/m1/quic-send-fragments.md similarity index 87% rename from quest/m2/performance-survey/quic-send-fragments.md rename to quest/m1/quic-send-fragments.md index de657c7125a..e157c81b4e0 100644 --- a/quest/m2/performance-survey/quic-send-fragments.md +++ b/quest/m1/quic-send-fragments.md @@ -7,6 +7,11 @@ chunks, if benchmarks establish a material cost. ## Plan +Decided 2026-10-22 (maintainer): moved from the m2 performance survey to m1, +since it sits on every relay's egress path and is cheap to settle before the +cut. A duplicate m1 quest planned from the repository improvement audit was +dropped in its favor. + At the audit revision, [SendBuffer::get](/rs/moq-quic/src/connection/send_buffer.rs) starts at the first unacknowledged segment for each lookup (lines 146-162). [Packetization](/rs/moq-quic/src/connection/streams/state.rs) calls it again diff --git a/quest/m1/quic-send-lookup.md b/quest/m1/quic-send-lookup.md deleted file mode 100644 index 5829c6b9a1b..00000000000 --- a/quest/m1/quic-send-lookup.md +++ /dev/null @@ -1,34 +0,0 @@ -# [S] Send buffer lookups stay linear - -## Goal - -A measured verdict on `moq-quic`'s `SendBuffer::get` cost as retained -segments grow, and, if the cost grows with segment count, a lookup whose -per-packet cost does not. - -## Plan - -Decided 2026-10-22 from the repository improvement audit; m1 because it sits -on every relay's egress path and is cheap to settle before the cut. - -`SendBuffer::get` (`rs/moq-quic/src/connection/send_buffer.rs`) walks the -retained segments from the front on every call, and packetization calls it -repeatedly, so sending N retained segments in order visits O(N^2) segments. -The audit's loop replay counted 8,256 visits for 128 segments of 64 bytes and -524,800 for 1,024. With a 1 MB window of ~1 KB writes, each packet walks about -500 segments. Those are visit counts, not measured time. The send window -already bounds retained bytes, so this is CPU only, not memory. - -- Measure first: sweep the existing `rs/moq-quic/benches/streams.rs` over - segment size and count, with ACKs delayed so segments stay retained. Count - instructions as the [performance audit](/quest/m1/audit-perf.md) does. -- Only if the sweep shows a slope, compare a sequential cursor (the next - unsent offset's segment index, with a fallback scan for retransmits) - against indexed segments (offsets searchable by binary search). Keep - each write as the caller's `Bytes`, with no coalescing copy. -- If the sweep is flat within noise, record that here and delete the quest. - -## Related - -- [Egress profile](/quest/m2/quic-egress-profile.md) - the broader send-path profile; its `SendBuffer` pool lever shares this type -- [QUIC I/O boundary](/quest/m2/quic-io-boundary.md) - reworks the same transmit buffers diff --git a/quest/m2/performance-survey/README.md b/quest/m2/performance-survey/README.md index de7cc34cd6c..52e00217885 100644 --- a/quest/m2/performance-survey/README.md +++ b/quest/m2/performance-survey/README.md @@ -112,7 +112,6 @@ No duplicate quest or workflow edit belongs here. ## Required -- [Fragmented QUIC sends](/quest/m2/performance-survey/quic-send-fragments.md) - packetization cost under a retained fragmented flight - [Task capacity after bursts](/quest/m2/performance-survey/task-capacity.md) - polling and memory after peak task occupancy subsides - [Bandwidth grants](/quest/m2/performance-survey/bandwidth-grants.md) - repeated full allocation in Rust and JS - [Cached JS group joins](/quest/m2/performance-survey/js-cached-groups.md) - repeated sorting while replaying retained groups diff --git a/quest/m2/quic-egress-profile.md b/quest/m2/quic-egress-profile.md index 192a2f3c12f..6b3e05060d8 100644 --- a/quest/m2/quic-egress-profile.md +++ b/quest/m2/quic-egress-profile.md @@ -51,7 +51,7 @@ hosts need. ## Related -- [Send buffer lookups stay linear](/quest/m1/quic-send-lookup.md) - settles `SendBuffer::get`'s per-packet scan first +- [Fragmented QUIC sends](/quest/m1/quic-send-fragments.md) - owns `SendBuffer::get`'s per-packet segment scan - [QUIC I/O boundary](/quest/m2/quic-io-boundary.md) - redesigns the transmit buffers the pool lever would feed - [#3201](/quest/m3/3201-moq-uring-use-sendmsg-zc-for-large-udp-gso-trains.md) - zero-copy on the same trains