Am Fr., 27. Jan. 2023 um 11:49 Uhr schrieb Henrik Boström <h...@chromium.org >:
> *Contact emails* > h...@chromium.org, h...@chromium.org > > *Background* > I attempted to remove this feature before but had forgotten to file an > intent to deprecate, background here > <https://groups.google.com/a/chromium.org/g/blink-dev/c/RsIktnGhHqw/>. > > *Specification* > The getStats() API spec is here <https://w3c.github.io/webrtc-stats/> and > it contains all the metrics. The deprecated metrics are also listed, but in > the obsolete section > <https://w3c.github.io/webrtc-stats/#obsolete-rtcmediastreamtrackstats-members>. > There's an open issue to remove obsolete metrics from the spec as soon as > they are unshipped from modern browsers. This is considered a blocker for > the document to reach Proposed Recommendation status. > > *Summary* > WebRTC is a set of JavaScript APIs (spec > <https://w3c.github.io/webrtc-pc/>) that allow real-time communication > between browsers. For the relevant metrics being removed, we're only > talking about the WebRTC use case that is sending or receiving audio or > video (typically Video Conferencing use cases), not the data channel use > cases that is a popular WebRTC use case, since data channel only use cases > would never have any tracks/streams. > > RTCPeerConnection.getStats() returns a map of string-to-objects, where > each object is one of the dictionaries defined in the stats spec. The > reason an app calls getStats() is mostly to report quality metrics (send > and receive resolutions, bitrates, glitches, video QP, etc) which can be > important for A/B experimentation. It can also be used in a way that > impacts app logic or even UX inside the app. Most common use case I can > think of: poll getStats() at 10 Hz and render volume bars for each > participant based on volume levels from stats objects. > > The deprecation in question is to remove some stats objects that were made > obsolete several years ago: all metrics on the "track" dictionary have been > moved to non-obsolete objects ("inbound-rtp", "outbound-rtp", > "media-source"). Reasons for wanting to deprecate include: > > - Spec-compliance: needed for browser implementations to align and for > the spec to become Proposed Recommendation. > - Web compat: Firefox never implement "track" or "steam" > > <https://wpt.fyi/results/webrtc-stats/supported-stats.https.html?label=experimental&label=master&aligned> > due > to them being obsolete. > - Performance: the duplicated metrics make up ~40% of the stats report > size, which can be a significant number of bytes in larger meetings and it > is common for apps to poll getStats() 10 times per second. > - Tech debt: unblock removal of 1400 LOC. > > In the meantime, the obsolete metrics is duplicated in several places of > the stats report. > > *Risks* > *- Impossible to properly measure usage* > Because stats objects are exposed as JavaScript dictionaries, and because > apps have to iterate all objects of the stats report in order to find the > ones they are interested in or if they just dump all the data without > filtering, there is no way to measure how big the dependency is on track in > the real world. > > While we lack use counters, we have some positive signs: > > - Because Firefox does not have "track" or "stream" stats, any app > that can run on Firefox already exercises the paths of these not existing. > > > - An experiment to "unship deprecated metrics" has been running at 50% > Canary since October > > <https://groups.google.com/a/chromium.org/g/blink-dev/c/RsIktnGhHqw/m/3iqjODsMBwAJ>, > giving developers testing Canary a heads-up. Nobody complained until the > experiment reached Stable. > - We got to 50% Stable in M109 and while we're in the process of > rolling it back now due to breaking twilio-video.js > <https://github.com/twilio/twilio-video.js/issues/1968>, it's > interesting to note that this is the only breakage we are aware of (that > does not mean there aren't more breakages, but I believe this at least says > something about the severity). > > *- Selenium et al typically starts browsers from fresh profiles and hence > does not know the finch trial seed* > The most likely explanation for breakage is not testing Canary or test > environments not having access to Finch experiments. This makes the > behavior on Stable a surprise. > > *- To have a Reverse Origin Trial or not to have a Reverse Origin Trial?* > Migrating should require so few lines of code (look for stats.type == > 'inbound-rtp' instead of stats.type == 'track', for example) that it seems > to be a bigger hurdle for a developer to enroll in the trial than to simply > fix their code. > > *- Compatiblity risk* > There is one particular metric out of all metrics that, if you run Safari, > does not exist on "inbound-rtp" yet. This can be a problem, but again is > probably not a big problem because this particular metric was never > implemented on Firefox so apps already need to survive without it, and it > is very easy to write a fallback path for the Safari case: > > let trackIdentifer = null; // In Firefox this will never be set > regardless. > if (inboundRtp.trackIdentifier) { > // Spec-compliant browser. > trackIdentifier = inboundRtp.trackIdentifier; > } else if (inboundRtp.trackId) { > // Fallback-path for Safari or 1+ year old Chromium browsers. > trackIdentifier = report.get(inboundRtp.trackId).trackIdentifier; > } > > *Proposal* > Rollback to 0% Stable but keep the "unship deprecated" experiment at 50% > Canary/Beta. Wait for Twilio to fix their issue and do another rollout > attempt. Keep a slower rollout pace next time. > > I see limited amount of value in a Reverse Origin Trial since it appears > to be more effort to register to the trial than to fix the issue, if you > are affected. > > I do prefer to have the feature enabled-by-default in M111+ and overwrite > that default via Finch rather than the other way around as to not "turn off > the fire alarm" for non-Finch testing environments. I realize that is not > perfect (what if you run in a non-Finch environment) but it would reduce > overall risk. > Thank you Henrik. I agree with one suggestion: only do default-off in M112+ (which is branching so you would just need to revert this commit <https://chromiumdash.appspot.com/commit/3a4d52f365df03413a856ea20366b36e8fb8ea0b> on the M111 branch). This gives developers another month to update (which itself should be quick) and then rolling it out to their customers and users (which takes time). I hope that the 50% rollout caused enough incidents (even though you may never hear about some of them) to get the fixes in ASAP. cheers Philipp -- > You received this message because you are subscribed to the Google Groups > "blink-dev" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to blink-dev+unsubscr...@chromium.org. > To view this discussion on the web visit > https://groups.google.com/a/chromium.org/d/msgid/blink-dev/5ecf1ea6-c16c-464a-b529-439e05e4feedn%40chromium.org > <https://groups.google.com/a/chromium.org/d/msgid/blink-dev/5ecf1ea6-c16c-464a-b529-439e05e4feedn%40chromium.org?utm_medium=email&utm_source=footer> > . > -- You received this message because you are subscribed to the Google Groups "blink-dev" group. To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscr...@chromium.org. To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CADxkKi%2B0M2bDXM9GDVSDhUrt9bUbXRnOOGC24fqTb3GS-uc3xA%40mail.gmail.com.
