Repository navigation
feat!: congestion control is chosen by algorithm name - #8944
Conversation
Quest-Claim: 2c6decba76ac1f3a
moq_tokio::quic::CongestionControl is Bbr, Cubic, Unlimited (spelled bbr, cubic, unlimited), and converts into the factory both runtimes install. quic::Controller, Config::controller, and moq_uring::quic::Congestion are gone. moq-ffi's MoqCongestionControl constructors and every wrapper follow. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Warning Review limit reachedYou've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. Next included review available in 2 minutes. View limit details
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. |
There was a problem hiding this comment.
ℹ️ No correctness issues found. One stale doc line outside the diff.
Reviewed changes
I reviewed all 36 files: the rename from congestion family to algorithm name across moq-tokio, moq-uring, moq-relay, moq-ffi, moq-c, every binding wrapper, the OBS plugin, the docs, and the TS harness.
- One enum, one factory mapping:
CongestionControlisBbr(#[default]) /Cubic/Unlimited. The newFrom<CongestionControl> for Arc<dyn moq_congestion::Factory + Send + Sync>is the only mapping, and both the tokio backend and the relay's io_uringtransport()now call.into(). That removes the relay'sbail!arm for an unknown family. - Removed hook and duplicate:
quic::Controller,Config::controller,Resolved::controller,Resolved::congestion(), andmoq_uring::quic::Congestionare gone. A repo-wide grep finds no remaining callers. Resolved::congestion_controlis non-Option:resolve()applies the default withunwrap_or_default().- Old words are refused, not aliased: a family name now fails the CLI, TOML,
FromStr, the C API, and OBS with an error that listsbbr, cubic, unlimited. The hidden--client-/--server-quic-congestion-controlfields now takeOption<String>, so the deprecation report still fires whatever value they carry. - Bindings: the moq-ffi constructors are renamed, and Go, Dart (regenerated), Swift, Kotlin, Python, C, and C++ follow.
- Tests: there are live loopback checks for the default and for
cubicon tokio, and a qlog-based check on io_uring (ssthreshagainstpacing_rate). The relay also gets tests for the environment variable and for refusing a family name in a config file.
ℹ️ rs/moq-congestion/README.md still names the delay family
Line 3 of rs/moq-congestion/README.md says applications pick a controller through moq_tokio::quic::CongestionControl, "where delay is BBR3 and the default". That word is now refused at startup, so the crate README points readers at a spelling that no longer parses. The rest of the repo has no stale references left. The MIGRATE.md entries are left for the release cut, as the PR says.
Technical details
# Stale family name in moq-congestion README
## Affected sites
- rs/moq-congestion/README.md:3 — "where `delay` is BBR3 and the default"
## Required outcome
- The sentence names the algorithm spelling (`bbr`, i.e. `CongestionControl::Bbr`) as the default instead of the removed `delay` family.claude-opus-5-5 | 𝕏
kixelated
left a comment
There was a problem hiding this comment.
Automated review by review (OpenAI)
Reviewed 9453e33ace3b6ec71f27b8561022427001187125. No blocker in the controller wiring: Tokio and io_uring now share one conversion, unset still means BBR v3, and the live tests check controller behavior rather than only the parser. I’d keep the names distinct and the startup error explicit.
Carry a concrete rollout step into the cut guide: any managed launcher that passes delay, loss, or none needs its text updated in the same deployment as its relay binary, and rolled back with it. A one-sided rollout can fail before the socket starts. For OBS, I’d prefer a one-time, visible saved-setting migration to the equivalent algorithm rather than making an upgrade break Go Live until the operator reselects it. Static review; I did not run the live tests.
# Conflicts: # dart/moq_ffi/lib/src/moq.dart # quest/m2/README.md # quest/m2/congestion-names.md # quest/m2/unused-api/transport.md # rs/moq-tokio/src/quic/backend.rs
|
Merged
(Written by Claude Opus 5.5) |
There was a problem hiding this comment.
✅ No new issues found. The stale README line from the last review is fixed.
Reviewed changes
Since the last Pullfrog review at 9453e33, the branch merged main and fixed the crate README.
- README names the algorithms:
rs/moq-congestion/README.mdnow says applications pickbbr(BBR3, the default),cubic, orunlimitedthroughmoq_tokio::quic::CongestionControl, not the refuseddelayfamily. - Merge resolution: the
backend.rstests now gatesim: Noneontest-support, the featurers/moq-tokio/Cargo.tomldefines. The regenerateddart/moq_ffi/lib/src/moq.dartdropped out of the diff becausemainno longer tracks that generated file. The other changed hunks come frommain.
claude-opus-5-5 | 𝕏
|
Maintainer decision (2026-11-17, asked in the ✅ An OBS scene saved with the old (Written by Claude Opus 5.5 with high) |
…r CUBIC Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Applied the maintainer decision in 5696c00: a scene saved with (Written by Claude Opus 5.5 with low) |
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
Since the last Pullfrog review at bf63df9, one commit changed the OBS refusal message.
- OBS refusal message: a saved
delay,loss, or unknownquic_congestion_controlvalue is still refused, and the error now readscongestion control '<word>' is not supported: pick BBR or CUBIC for Congestion control in the output's Advanced settings. It names the fix instead of listing accepted words. - Test:
moq-settings-test.cppchecks that the refusal names the offending word (fast,delay,loss) and containspick BBR or CUBIC.
claude-sonnet-5-5 | 𝕏

Finishes quest/m2/congestion-names.md and deletes it.
Problem
Congestion control was picked by family (
loss,delay,none), which stops naming one thing once a second delay-based algorithm (GCC or SCReAM from the realtime bake-off) exists. moq-tokio and moq-uring each had their own copy of the enum and its factory mapping, and moq-relay mapped one onto the other variant by variant, so a new algorithm took three edits.quic::Config::controllerpublishedmoq_congestion::Controlleras a plugin interface while that trait is still moving, and nothing outside moq-tokio's own tests set it.Approach
moq_tokio::quic::CongestionControlisBbr(default),Cubic,Unlimited, spelledbbr,cubic,unlimited, still#[non_exhaustive]. It now converts into theArc<dyn moq_congestion::Factory>a QUIC transport config holds, so the tokio backend and the relay's io_uring workers both build their controller through that oneFromimpl. Adding an algorithm is one variant plus one match arm.quic::Controller,Config::controller,Resolved::controllerandmoq_uring::quic::Congestionare deleted. moq-uring'stransport::Configalready takes a factory and defaults to BBR v3 directly. The relay's variant-by-variant match and itsbail!for an unknown family are gone.Resolved::congestion_controlis now a plainCongestionControlwith the default applied, not anOption.--quic-congestion-control delay,MOQ_QUIC_CONGESTION_CONTROL=none, andcongestion_control = "loss"in TOML each fail at startup with an error that listsbbr, cubic, unlimited. The hidden--client-/--server-quic-congestion-controllegacy fields now take any string, so the deprecation report still names them whatever value they carry.MoqCongestionControl::{bbr, cubic, unlimited}, plus every wrapper: GoBBRCongestionControl/CubicCongestionControl/UnlimitedCongestionControl, Python / Swift / Kotlin / DartCongestionControl.bbr()/.cubic()/.unlimited(), Cquic_congestion_control = "bbr" | "cubic" | "unlimited", and C++ through the generatedmoq::CongestionControl. The Dart bindings are regenerated. The OBS plugin's dropdown storesbbr/cubic.doc/bin/relay/config.md, the TS harness README and run script, and thets.ymloverflow jobs (MOQ_QUIC_CONGESTION_CONTROL: cubic).Impact
moq_tokio::quic: theCongestionControlvariants are renamed. ChangeLosstoCubic,DelaytoBbr, andNonetoUnlimited.ControllerandConfig::controllerare removed with no replacement, since the built-in algorithms are the only choice.Resolved::congestion_controlisCongestionControl(default applied), notOption<CongestionControl>. New:From<CongestionControl> for Arc<dyn moq_congestion::Factory + Send + Sync>.moq_uring::quic:Congestionis removed. Build the factory frommoq_tokio::quic::CongestionControlwith.into(), or frommoq_congestiondirectly.quic.congestion_control,--quic-congestion-control, andMOQ_QUIC_CONGESTION_CONTROLtakebbr,cubic, orunlimited. Changedelaytobbr,losstocubic, andnonetounlimited. The old words fail at startup.MoqCongestionControl::delay()/loss()/none()arebbr()/cubic()/unlimited(), and the C string field takes the new words. Go'sDelayCongestionControl/LossCongestionControl/NoneCongestionControlareBBRCongestionControl/CubicCongestionControl/UnlimitedCongestionControl.delayorlosssetting makes the output refuse to start withInvalid advanced MoQ settings: congestion control 'delay' is not supported: pick BBR or CUBIC for Congestion control in the output's Advanced settings. Pick BBR or CUBIC again in the settings.@moq/net: unchanged. It has no QUIC stack to configure and passes the browser's owncongestionControlhint.Decisions
These are the choices I made; each can be undone in review.
bbr, cubic, unlimited(invalid value 'delay' for '--quic-congestion-control' [possible values: bbr, cubic, unlimited], and serde'sunknown variant 'loss', expected one of 'bbr', 'cubic', 'unlimited'). It is not acli::Deprecatedline mappingdelay -> bbr. usage-rs'sValueEnumhas no deprecated-value hook, so a per-word mapping would need hidden variants on the public enum or a hand-written parser that loses the generated choices in help and completion. The quest only recommended this behavior; the maintainer had not decided it.MIGRATE.mdedits: the quest's Plan asked for them, butAGENTS.md(and #8424) says the release cut writesMIGRATE.mdfrom each PR's Impact section. The notes are in Impact above.BBRCongestionControl, with Go's capitalized initialism, following the wrapper'sGSO/MTUDiscovery/QUIC.bbr/cubicand refuse the old words, rather than mappingdelay/losssilently. This matches the quest's "no silent alias" reasoning.delayorlosskeeps failing to start, and its error names the fix: pick BBR or CUBIC for Congestion control in the output's Advanced settings (rejected: mapping old values tobbrwith a warning; refusing with the oldunknown congestion control ... (expected bbr or cubic)message unchanged). An unknown word that never existed gets the same message, since the fix is the same.unlimitedin OBS: the dropdown stays Automatic / BBR / CUBIC, sinceunlimitedis for tests and controlled links.Validation
Real binary (
moq-relay):--quic-congestion-control delay, a TOML file withcongestion_control = "loss", andMOQ_QUIC_CONGESTION_CONTROL=noneeach refuse at startup, and each message listsbbr, cubic, unlimited.Live loopback, tokio runtime (
moq-tokiocubic_reaches_the_live_connection,default_reaches_the_live_connection): withcubic, both ends' live controllers report an ssthresh, cannot probe, and offer moq-net no probe target. With the knob unset, both ends probe and the client has a probe target (BBR v3).Live io_uring relay, checked a second way by reading the qlog traces (
uring_workers_run_the_configured_congestion_control, run withCI_STRICT=1so it cannot skip): a relay givencubicwritesssthreshand nopacing_ratein its recovery metrics, and an unset relay writes the reverse. Mutating the relay to always install BBR makes this test fail.Parsing: every name parses from flag, TOML, and
FromStr(case-insensitive); the environment variable through the relay's merge (congestion_control_from_the_environment); every family name is refused on the flag, in TOML, throughFromStr, from a relay config file, and through the C API.just check: all green exceptpy/moq-rs/tests/test_socket.py::test_abort_and_endpoint_close_are_errors, a flaky socket-close race that fails 4 runs in 6 here and does not touch QUIC congestion control; it is fixed by #8900. Swift is skipped on Linux.just test interop --all: 210 passed, with 1 pre-existing skip (Chromium route key).cargo testfor moq-tokio, moq-uring, moq-relay, moq-ffi, and moq-c: all pass.After merging
main(conflicts inquest/m2/README.md, the deleted quest files, andrs/moq-tokio/src/quic/backend.rs's tests, resolved to this PR's side):cargo nextestfor moq-tokio, moq-relay, moq-ffi, moq-c, and moq-uring, 1760 passed;just rs test, 11713 passed. Onejust checkrun timed out fourmoq-muxcontainer::consumertests together under load; they pass alone and in the full rerun, and this PR does not touch moq-mux.OBS refusal message:
just obs compilepasses, and themoq-settingsunit test (plain build, real moq-ffi) checks thatdelay,loss, andfastare each refused with a message naming the word andpick BBR or CUBIC.just obs test(ThreadSanitizer) failed 5 reconnect-timing checks inmoq-output-test.cpp(lines 526, 999-1003) on this loaded host, before reaching the settings test; this commit doesn't touchmoq-output, and the PR's plain-build OBS CI passed them on the previous head.Alternatives
loss/delay/noneas hidden aliases. Rejected: a silent alias keeps the family names alive in every deployed config.quic::Config::controlleras an application hook. Rejected by the maintainer: it publishes a trait that is still changing.Follow-ups
None.
Models
Size: M
Implementer: Claude, claude-opus-5-5, medium effort (arm A)
Implementer: Claude Code, claude-opus-5-5, unknown effort (iterate, S, arm A)
Implementer: Claude Code, claude-opus-5-5, low effort (iterate, XS, arm A)
🤖 Generated with Claude Code
(Written by Claude Opus 5.5 with medium)
Review
Rewritten by
quest reviewafter each round.Head: 5696c00
Base: origin/main at f88e8f5
Risk: high,
it changes .github/workflows/ts.yml, which is always high: workflows run with the repository's secrets and decide what merges and shipsClassified: high by the always-high paths
Reviewer: Codex, gpt-6.1-sol, high effort
Session:
01a1eb6a-6468-7c93-9064-eaf017b4e307, round 2Score: 9/10,
The incremental OBS error-message and assertion changes are consistent with the available settings, with no actionable defects found; builds and tests were not run as instructed.Rounds: 4; first round 1 (medium 1); findings 1 (medium 1): fixed 1, refuted 0, open 0
Findings: none