W3C

– DRAFT –
WebRTC November 19 2024 meeting

19 November 2024

Attendees

Present
Bernard, Carine, Dom, Florent, Guido, Harald, Henrik, PatrickRockhill, TimP, Youenn
Regrets
-
Chair
Bernard, Guido, Jan-Ivar
Scribe
dom

Meeting minutes

Recording: https://www.youtube.com/watch?v=eHjg9figxLk

Slideset: https://docs.google.com/presentation/d/1qFWqfbOVv3mHlYCBtnmXtLdpMkk-pXf--cGoEQBCyi4/ (archived PDF copy)

Timing Control in MediaStreamTrack Insertable Media Processing using Streams 🎞︎

Issue #80: Expectations/Requirements for VideoFrame and AudioData timestamps 🎞︎

[Slide 11]

Jan-Ivar: what we want to do is to write tests to see how browsers proceed today, and ideally prescribe that

Bernard: but where? in HTML video element?

HTA: I think it'd belong there; what do video elements do when they get the same timestamps?

Youenn: each sink is supposed to define how it handles videoframes
… In Safari, we use the time at which the frame at which it is committed the sink
… webrtc-pc should define what it does when sending frames
… the HTML video element should define it, although we could do in mediacapture-main if it's easier

Bernard: that's a good point - the track could go to a peerconnection where ti would be weird to have these duplicate timestamps

Youenn: we need to test, but I suspect they're using the time at which frames are submitted

Guido: +1 to youenn
… this would also mean specifying for mediatrackrecorder, mediatrackprocessor

youenn: let's make sure to file issues for all these specs

PROPOSED: Add to mediacapturem-main extensiblity to make sure sink define their behavior on frame timestamps and file issues on sink specs accordingly

TimP: if you were to feed a track with null timestamps to a video element, would it be different if it had gone through a null transform? that would be weird

hta: a null transform would do nothing

TimP: right; but an identity transform shouldn't change how the track gets rendered

dom: as long as we're defining at the sink, this means the processing pipeline shouldn't change the outcome

youenn: it may be that we can define a default behavior for sinks (likely, changing the timestamps would have no effect)

bernard: that's my hope too; PC is an interesting case

Jan-Ivar: in the case of a transform, a Processor WOULD be a sink - so we should be careful we keep the invariant of identity

RESOLUTION: Add to mediacapture-main extensiblity consideration to make sure sink define their behavior on frame timestamps and file issues on sink specs accordingly

Issue #114: VideoTrackGenerator/MediaStream should buffer the current frame 🎞︎

[Slide 12]

Youenn: +1 on not buffering; I'm surprised that captureStream() is different
… we should check if there isn't a misunderstanding on the behavior of captureStream()

Jan-Ivar: it would be nice if we could keep them consistent, although I understand the cost of buffering

Youenn: but I don't think it's even buffering, it's only a timing question

Henrik: clearly we shouldn't buffering in the traditional sense; but in terms of remembering the last frame, in the case of a variable framerate track, is delivery guaranteed?

Bernard: that may deserve another issue

Henrik: +1 on this being the expected behavior

HTA: the <canvas> element exists no matter, so a frame can be generated at any time; I think this is expected behavior. Seeing if we can align the polyfill with this, that would be best and informative

RESOLUTION: Not buffering the last frame is the expected behavior

Guido: we can recommend to the developer that they can buffer the frame in the application and send it to the generator after the track is associated to the element

Youenn: note that this problem only occurs when using VTG in the main thread

Issue #20: Add “real-time” warning/note to `MediaStreamTrackGenerator` 🎞︎

[Slide 13]

Youenn: with VTG's writable stream, the promise could be awaited as a cue to when a new frame can be submitted
… not sure if it's needed or still web-compatible

Bernard: what happens if you oversubmit? is it only they're not being displayed by the sink?

HTA: I think they would be thrown away

Jan-Ivar: a writable stream normally comes with a queue with a highwater mark; how should the VTG infers its framerate from the pace of frames it is being fed? Maybe we should define a max framerate over which it gives backpressure
… or maybe it just drop frames over that rate if we're sticking to real-time behavior
… the stream can either be push- or pull-based; for pull-, we need to determine where the framerate gets identified

youenn: the write-frame algorithm returns a promise, which resolves the writable write(), which gives a path for backpressure
… at the moment, the promise always resolve at the end of the microtask, but we could change it to allow for backpressure
… at the high level, I don't think VTG should be lossy, while sinks might be

Guido: +1 to youenn
… if you don't wait for the promise, you have no guarantee that the frame will be processed

henrik: the promise helps with avoiding pushing too much data
… but for lossy sinks, there will still be no way to make sure the frame got processed

dom: ideally, lossy sinks would be able to transfer backpressure via VTG

youenn: this would be possible for PC; possible for recorder; not sure about media elements, although in case of too many frames they should be able to at least render the last
… I think the invariant is that VTG is not lossy, but some sinks will be lossy, and we should say that the last frame should win for renderers

jan-ivar: a writable always has a queue, even if the sink doesn't have one
… we need to understand how the framerate gets determined for a pull-based stream

Bernard: this will depend on pre-recorded vs live

Issue #86: Playback and sync of tracks created by `VideoTrackGenerator` 🎞︎

[Slide 14]

HTA: looking at the reported bug, the person reporting it had specific ideas on how the video element work, e.g. with playout points

Jan-Ivar: I don't know the answer, but writing tests would help determine if timestamp has any effect, in which case browsers are likely optimizing local playback

[Slide 15]

[Slide 16]

Youenn: Tracks are sent synchronized by the PC, send asked to render ASAP given the real-time context; likewise for local capture; I don't thnk timestamps are playing any role in it

Henrik: I thought presentationTimestamp was being used in Chrome?

Jan-Ivar: that should be testable

WebRTC API 🎞︎

Issue 229: receiver.hardwareAcceleration attribute instead of disableHardware[En|De]coding() static methods 🎞︎

[Slide 20]

Youenn: not sure whether it's worth exposing just that; there are proposals to expose webcodecs data to the PC
… I'm not sure if the implementation will be able to determine if switching to software or hardware is adapted given impact on latency, drops given the opacity of the PC flow; this is much different from the WebCodecs context where everything is happening under the control of the JS
… I would prefer we spend more time on exposing WebCodecs rather than providing this hard-to-exploit toggle to developers
… I'm not more in favor of the new API than the current one

Bernard: the original motivation was to turn-off hardware acceleration to avoid encoder bugs
… this new approach gives control which as Youenn describes wouldn't be operational

Henrik: I think this new proposal is best because it allows to have the remote party signal to the specific client they should disable their hardware encoder

HTA: I recall as Bernard that this was a toggle to remove hardware encoder on detected failure pattern
… this new proposal is more flexible; the "preferSoftware" might be too soft for the original purpose though

[Slide 21]

Jan-Ivar: one of the reason of making it softer is to allow UA to override the hint when e.g. they've been blocklisted due to issues in earlier versions of the browser

HTA: hearing two proposals: Youenn suggest deleting the old API altogether; the other is to add this new API

Bernard: with encoded sources, we're heading towards WebCodec control which will give you the granular control if you need it

Youenn: +1 to Bernard
… encoded sources bring that to the encoder side; we don't have the equivalent on decoding yet, but would be a reasonable direction to explore

Henrik: how about my suggestion to rely on the remote corruption metrics?

Youenn: you could hot-download a WASM decoder when you detect that situation

Jan-Ivar: not hearing a lot of consensus on the proposal; I expect Mozilla's position will return to negative to the existing API

Repository: w3c/webrtc-pc

Issue #3017: Reconsider replaceTrack's blocking on the operations queue 🎞︎

[Slide 22]

HTA: +1 after putting more thoughts on this

Youenn: sounds fine, and helps getting a bit faster with limited compat issue

[Henrik: +1 on reaction]

RESOLUTION: move with forward with PR to align spec with Chrome behavior

Issue #3022: Clarify order of events when the SCTP transport is closed 🎞︎

[Slide 23]

[Slide 24]

Youenn: I think we should enqueue a task per event; in terms of order for data channels, I'm not sure if there an API with an order - there used to be a getDataChannels() ordering which we could consider reusing
… re closed transports, my intuition we should fire on the SCTP transport first, but no concrete reasoning

Jan-Ivar: we could use the orders of the internal slot, although not sure how that intersects with datachannels that are transfered

Florent: the sctpReason property gives indication about what caused the closure of the channel

TimP: I could live with getting notifications only at the transport level when the full association goes down

Jan-Ivar: if we fire first on SCTP, then you would know why all the channels close subsequently

TimP: right; in particular, you don't want to just re-open the channels if the transport itself has failed

Youenn: I think it's useful to still emit the close events on channels in particular when they've been transfered to a worker

HTA: note that there is also an error event fired

RESOLUTION: Proceed with PR to fire sctp closed on transport object and then on individual channels

Add event for CSRC change 🎞︎

[Slide 30]

[Slide 31]

Bernard: the original use case wasn't to have extensive polling; there was also no expectations to manage as large meetings as is the case today
… I think the main use case is for audio rather than video to surface where sound is coming from
… there was concerned about the volume of events that could be fired from this as well

Henrik: this would be only when the source changes, so it shouldn't a lot

Jan-Ivar: my concern is the relationship to media flowing, and the linkage to the mute event which is already being fired

Henrik: I asked and they indicated they're interested in source information

Jan-Ivar: but I'm concerned in introducing an API that would cement the interop gap we have today

Jan-Ivar: this could be expressed in the standard's position; I'm trying to make sure we distinguish implementation issues from use case discussion

Youenn: the idea is that when a CSRC or SSRC change is detected, a task would be queued to fire an event with the changed value?
… if so, seems OK with me

Jan-Ivar: I'm not sure if the csrc/ssrc names are evocative enough for the use case

Henrik: this matches the existing names of the getters

Bernard: for the video, would this be coming from the video compositor? or if the mixer mixes audio from different sources?

Henrik: it would probably send an array with the various sources?

HTA: we know there are apps that do this today via polling; what we're looking for is exposing this in a more ergonomic and efficient way

Jan-Ivar: re the use case - polling isn't necessarily bad

Henrik: you would still need polling e.g. to animate an audio volume
… but having information asap on a change to e.g. switch the main video is time critical
… we've received several reports where this polling is creating glitches and impacts performance

Bernard: it's kind of a new use case from the original one that was focused on audio - but now this also extends to video, is that common now?

Henrik: I know at least of a Google product doing this

Youenn: this aligns with usual Web expectations in terms of signaling changes to states

Jan-Ivar: my support is tentative, given my concern about the lack of interop of mute event

Henrik: we can make that clear in our intent to ship for Chromium

Youenn: implementing this new API could help improve implementations of the mute event; for Safari, we would like to align with spec and I think this would help

Henrik: fixing onunmute along with implementing this makes a lot of sense to me
… onunmute makes sense regardless since not all apps used virtual ssrcs
… the lack of support has mostly been a question of prioritization

Guido: a lot of the mute issues were linked to the ambiguity in the mediacapture spec that crept in multiple specs; this is something we want to fix
… see e.g. the proposal I made to expose additional stats to help with transition

Jan-Ivar: onunmute also useful as trigger for gathering stats

RESOLUTION: Rough consensus on moving forward with events for csrc/ssrc changes, with concerns noted on related onunmute interop

Repository: w3c/webrtc-pc

Issue #3020: Codec negotiation and level IDs 🎞︎

[Slide 32]

[Slide 33]

[Slide 34]

Bernard: you said before negotiation, 180 throws because you have no codec; this relates to what constitutes the media format
… for H265, it's profile-id and tier only, so setting level-id to 180 should not throw

Henrik: I think the current spec says "equals" which could be clarified
… but even with that, there is still another issue that my next slide details

Jan-Ivar: I think it's OK for the code to fail before negotiation

Henrik: you're in proposal B camp then :)

[Slide 35]

Henrik: I personally prefer proposal A :)

Bernard: me too

Jan-Ivar: A makes sense after hearing your argument that it could produce unexpected results in prod

HTA: proposal A, and we should document that when you can't negotiate the codec, it should throw

RESOLUTION: Proposal A

Repository: w3c/webrtc-pc

[Youenn departs]

Remote channelCount + Stereo Opus #3010 #3011 🎞︎

[Slide 36]

[Slide 37]

Jan-Ivar: re how to test it, with Web Audio

Henrik: still I think it would better to add it for symetry

Jan-Ivar: adding it seems OK to me

RESOLUTION: proceed with adding .channelCount on remote tracks

[Slide 38]

[Slide 39]

TimP: I'm almost sure that the SDP for Opus is always stereo

Henrik: I think the spec says it's always stereo capable; but the SDP parameter only expresses a preference

TimP: but Opus on the wire I think will always be stereo

Henrik: I think it can be mono on the wire (since I understood it can impact bitrate)

Jan-Ivar: it's hard to detect a mono signal from a stereo track; having that exposed in channelCount would be useful
… re channelCount, would this trigger onconfigurationchange?

Henrik: I haven't thought about it

HTA: I talked with Jean-Marc about mono & stereo; if you know a signal is mono you can encode/decode with less CPU power, which is good
… re testing, generate a track with different signals (e.g. a waveform on one channel and silence on the other) and see whether that's preserved

Jan-Ivar: re proposal 2, "MAY" is not great

Henrik: I suspect implementations would align with what libwebrtc does

Jan-Ivar: let's confirm it and standardize that
… if the libwebrtc behavior isn't satisfactory, we can revisit that

Summary of resolutions

  1. Add to mediacapture-main extensiblity consideration to make sure sink define their behavior on frame timestamps and file issues on sink specs accordingly
  2. Not buffering the last frame is the expected behavior
  3. move with forward with PR to align spec with Chrome behavior
  4. Proceed with PR to fire sctp closed on transport object and then on individual channels
  5. Rough consensus on moving forward with events for csrc/ssrc changes, with concerns noted on related onunmute interop
  6. Proposal A
  7. proceed with adding .channelCount on remote tracks
Minutes manually created (not a transcript), formatted by scribe.perl version 235 (Thu Sep 26 22:53:03 2024 UTC).