From 0e37c87c9aa246afbfd3d5ec8042f336ff40d8aa Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Thu, 22 Oct 2026 13:41:49 -0700 Subject: [PATCH] quest: apply maintainer fixes to merged quest plans - Delete jitter-density; #6359 bounded the jitter reference history. - announce-cursor-cpu, max-hops-setting, and drain-peer-redial ship with the next cut from main instead of a release backport. - Drop the release-audit child count so it cannot go stale. Co-Authored-By: Claude Opus 5.5 --- quest/m1/README.md | 12 ++++----- quest/m1/announce-cursor-cpu.md | 8 +++--- quest/m1/audio-playout-memory.md | 1 - quest/m1/drain-peer-redial.md | 4 +-- quest/m1/jitter-density.md | 42 -------------------------------- quest/m1/max-hops-setting.md | 5 ++-- quest/m1/release-audit/README.md | 2 +- 7 files changed, 16 insertions(+), 58 deletions(-) delete mode 100644 quest/m1/jitter-density.md diff --git a/quest/m1/README.md b/quest/m1/README.md index aa019494bf8..0b060e965be 100644 --- a/quest/m1/README.md +++ b/quest/m1/README.md @@ -63,8 +63,9 @@ padding and the browser's rung disable stay in m2, with only padding's lite-07 stream type here. Added 2026-10-22 (maintainer, clustering audit): three interim fixes for -announce churn on today's hop-list mesh, each backported to `release`, so -deployments get relief before the route layer ships; and stats groups as +announce churn on today's hop-list mesh, shipping with the next cut from +`main` (not backported to `release`), so deployments get relief before the +route layer ships; and stats groups as tracks, so per-group stats stop costing cluster announces. ## Required @@ -74,9 +75,9 @@ tracks, so per-group stats stop costing cluster announces. - [Readers follow restarts](/quest/m1/sink-restart-follow.md) - every moq-cli and moq-gst reader, moqsrc and `moq play` included, follows instances through `announced_broadcast` and an epochless `Restart` where its output can mark it; otherwise it stays on the old instance and fails once that route dies - [In-band auth](/quest/m1/auth/README.md) - a session tells its peer what it may publish and subscribe to, unions tokens presented in band, and fails loud on an out-of-scope publish - [Cluster routing](/quest/m1/cluster-routing/README.md) - prove shared routing contracts, implement per-origin routes, and connect outbound local clusters -- [Announce cursors stay linear](/quest/m1/announce-cursor-cpu.md) - a withdrawal costs each announce cursor O(K log K), and a no-op route change allocates nothing; backported to `release` -- [Path length setting](/quest/m1/max-hops-setting.md) - operators cap hop paths near their mesh diameter instead of 32; backported to `release` -- [Draining relays don't pull peers back](/quest/m1/drain-peer-redial.md) - cluster peers fail over instead of redialing a draining relay, and peer redials back off with jitter; backported to `release` +- [Announce cursors stay linear](/quest/m1/announce-cursor-cpu.md) - a withdrawal costs each announce cursor O(K log K), and a no-op route change allocates nothing +- [Path length setting](/quest/m1/max-hops-setting.md) - operators cap hop paths near their mesh diameter instead of 32 +- [Draining relays don't pull peers back](/quest/m1/drain-peer-redial.md) - cluster peers fail over instead of redialing a draining relay, and peer redials back off with jitter - [Stats groups as tracks](/quest/m1/stats-group-tracks.md) - one stable node broadcast per stats producer, with each group a produced track listed by TRACKS - [Routing guide](/quest/m1/routing-guide.md) - concepts and practical outbound federation setup - [Controlled time](/quest/m1/time/README.md) - every crate and package reads time through moq-clock or @moq/clock, with regression tests for the changed contracts @@ -112,7 +113,6 @@ tracks, so per-group stats stop costing cluster announces. - [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 - [Buffered backpressure timestamps](/quest/m1/audio-backpressure-timestamps.md) - buffered audio keeps holding frames past 2^31 microseconds instead of decoding ahead and dropping audio - [Audio playout memory](/quest/m1/audio-playout-memory.md) - a catalog's advertised jitter or delay cannot size decoded PCM; the lookahead stays encoded -- [Jitter estimator density](/quest/m1/jitter-density.md) - the estimator does O(1) work per frame however densely timestamps are packed, in JS and Rust - [AAC alias classification](/quest/m1/aac-he-alias.md) - every accepted AAC spelling gets the right frame duration from one classification - [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 diff --git a/quest/m1/announce-cursor-cpu.md b/quest/m1/announce-cursor-cpu.md index 5ad3d48c674..a06a8c11a3b 100644 --- a/quest/m1/announce-cursor-cpu.md +++ b/quest/m1/announce-cursor-cpu.md @@ -6,8 +6,9 @@ An announce cursor's work per event is independent of how many entries it holds back. When an origin holding K broadcasts goes away, each cursor does O(K log K) work, not O(K²). A route change that leaves a cursor's winning route unchanged allocates nothing for that cursor. Both properties have an -operation-count regression test and a benchmark, and the fix is backported -to `release`. +operation-count regression test and a benchmark. It ships with the next cut +from `main`; the maintainer decided on 2026-10-22 not to backport it to +`release`. ## Plan @@ -44,8 +45,7 @@ both fan-out axes: one origin withdrawing 1k and 10k broadcasts, and one broadcast's route churning, each under 1, 20, and 100 cursors, against `main` before the change. -Backport: per CONTRIBUTING, land on `main`, then port onto `release` as a -separate PR keeping the published API. No wire change, no public API change. +No wire change, no public API change. ## Related diff --git a/quest/m1/audio-playout-memory.md b/quest/m1/audio-playout-memory.md index dfe45d3abe7..66acb1e6e4e 100644 --- a/quest/m1/audio-playout-memory.md +++ b/quest/m1/audio-playout-memory.md @@ -48,4 +48,3 @@ any `js/watch` API change goes in its `MIGRATE.md`. - [Buffered backpressure timestamps](/quest/m1/audio-backpressure-timestamps.md) - the 32-bit truncation in the same `Backpressure`, landing first - [Audio jitter target](/quest/m1/audio-jitter-target/README.md) - the playout target this sizes memory for -- [Jitter estimator density](/quest/m1/jitter-density.md) - the CPU side of the same hostile-publisher concern diff --git a/quest/m1/drain-peer-redial.md b/quest/m1/drain-peer-redial.md index 75108547c46..e24745e3f8d 100644 --- a/quest/m1/drain-peer-redial.md +++ b/quest/m1/drain-peer-redial.md @@ -6,7 +6,8 @@ When a clustered relay drains, its cluster peers fail over to other routes instead of redialing it, and they reconnect once its replacement is up. Peer redials back off with jitter and a cap of tens of seconds, so a mesh recovering from an outage doesn't redial in lockstep and a recovered peer -is not left waiting minutes. Fixed on `main` and backported to `release`. +is not left waiting minutes. It ships with the next cut from `main`; the +maintainer decided on 2026-10-22 not to backport it to `release`. ## Plan @@ -39,7 +40,6 @@ relay while it drains, and one reconnect after it restarts. Backoff tests assert spread and cap under controlled time. Public API: none expected; report any connect config change. Wire: none. -Backport onto `release` as a separate PR. ## Related diff --git a/quest/m1/jitter-density.md b/quest/m1/jitter-density.md deleted file mode 100644 index 71fccb20f16..00000000000 --- a/quest/m1/jitter-density.md +++ /dev/null @@ -1,42 +0,0 @@ -# [M] The jitter estimator costs constant time per frame - -## Goal - -The arrival-jitter estimator in `js/hang` and `rs/moq-audio` does amortized -O(1) work per observed frame however densely a publisher packs media -timestamps, and produces the same target series for ordinary traces. A -publisher can no longer stall a viewer's main thread by sending many frames a -microsecond apart. - -## Plan - -Found in a 2026-10-22 media audit at -[main 1f5caa7](https://github.com/moq-dev/moq/commit/1f5caa7026bcfd8a2684da3712b384b5f74a7496) -and confirmed by source at -[main d4cdfe4](https://github.com/moq-dev/moq/commit/d4cdfe4368b17cbdf7d2a0c43d24e364a9626480). -`Jitter.observe` in `js/hang/src/container/jitter.ts` and `Estimator::observe` -in `rs/moq-audio/src/jitter.rs` keep every arrival inside a 2 s media-time -window and scan it on each observation for the fastest reference, so N frames -in one window cost O(N^2). The audit measured the JS estimator at 82 ms for -10,000 observations a microsecond apart and 4.51 s for 80,000; Rust was -source-verified only. Observation runs before decode or drop, so nothing -downstream bounds it. - -For an in-order frame the reference is the window's minimum of -`arrival - media`, which a monotonic deque gives in amortized O(1). A -reordered frame needs the minimum over the prefix at or before its timestamp, -which the deque no longer holds; measuring it against the window minimum -instead reads a reordered frame's delay as at least as large, which is -arguably the wait it really costs. Decide that against the conformance -corpus, and update `doc/concept/audio-jitter.md` (normative) in the same PR -if the definition changes. Both languages land together and keep passing the -shared corpus. - -Add a benchmark swept over observations per window (ordinary 20 ms cadence up -to microsecond spacing) in both languages, so a slope shows up. No API or -wire change. - -## Related - -- [Audio jitter target](/quest/m1/audio-jitter-target/README.md) - the estimator's owner and conformance corpus -- [Audio playout memory](/quest/m1/audio-playout-memory.md) - the memory side of the same hostile-publisher concern diff --git a/quest/m1/max-hops-setting.md b/quest/m1/max-hops-setting.md index 6c822f0851b..b0d577cd35e 100644 --- a/quest/m1/max-hops-setting.md +++ b/quest/m1/max-hops-setting.md @@ -5,7 +5,8 @@ A relay operator can set the longest hop path a relay accepts or forwards, with today's 32 as the default. On a mesh whose diameter is a few hops, a cap near diameter plus a margin stops lite-06 path hunting from exploring -long detours after a withdrawal. The setting is backported to `release`. +long detours after a withdrawal. It ships with the next cut from `main`; the +maintainer decided on 2026-10-22 not to backport it to `release`. ## Plan @@ -33,7 +34,7 @@ the residual counterexamples' final state. A path longer than the cap is dropped on receive and never sent. Docs: `doc/bin/relay/cluster.md` gains the setting. Public API: one relay -config field. Wire: none. Backport onto `release` as a separate PR. +config field. Wire: none. ## Related diff --git a/quest/m1/release-audit/README.md b/quest/m1/release-audit/README.md index 58c4f100770..1443e4dfc18 100644 --- a/quest/m1/release-audit/README.md +++ b/quest/m1/release-audit/README.md @@ -83,7 +83,7 @@ network-impairment performance result is implied by these checks. ## Required -The 21 children comprise 17 reproduced or source-confirmed findings, three +The children comprise reproduced or source-confirmed findings, contract/lifetime investigations, and one security-review pointer. Each child records whether its behavior is new on main or inherited from release.