Web Real-Time Communications Working Group Teleconference

11 Jun 2012


See also: IRC log


Harald_Alvestrand, Stefan_Hakansson, Magnus_Westerlund, Ted_Hardie, Tim_Terriberry, Anant_Narayanan, Dan_Burnett, Dan_Druta, Dominique_Hazael-Massieux, Cullen_Jennings_(remote), Justin_Uberti_(remote), Adam_Bergkvist, Jim_Barnett, Neil_Stratford, (remote), Stephan_Wenger, Gang_Liang(remote), Salatore, Loreto, Jonathan_Lennox, EKR_(remote), Randell_Jesup_(remote), Maire_Reavy_(remote), Mary_Barnes, Gonzalo_Camarillo, Richard_Ejzak, Dan Romascanu (remote)
Harald_Alvestrand, Stefan_Hakansson
anant, burn, DanD, stefanh_, dom, stefanh


<dom> (there are more people in the room, but I can't identify them visually; if you know any of the missing names, please type "Present+ Name")

<dom> JSEPified PeerConnection API (slides)

<dom> Scribe: anant

<dom> WebRTC open action items

<dom> ACTION-11?

<trackbot> ACTION-11 -- Daniel Burnett to add Constraints API to API spec -- due 2012-01-12 -- OPEN

<trackbot> http://www.w3.org/2011/04/webrtc/track/actions/11

Welcome to the W3C interim! Coffee at 10:30, lunch at 12:30 CET

Administrivia, going through action items

stefanh_: Action 11 is ongoing, more to be discussed today

<dom> ACTION-12?

<trackbot> ACTION-12 -- Daniel Burnett to add Stats API to API spec -- due 2012-01-20 -- OPEN

<trackbot> http://www.w3.org/2011/04/webrtc/track/actions/12

stefanh_: ACTION 12

<dom> ACTION-12: Harald's proposal http://lists.w3.org/Archives/Public/public-webrtc/2012Jun/0040.html

<trackbot> ACTION-12 Add Stats API to API spec notes added

harald: I put some comments, feedback welcome

<Martin_> I'm not sure that the mic is working

burn: capabilities was discussed in terms of the constraints. there needs to be a quick check on the list before we can put it in

dom: I agree. there is the sysapps WG whose one of the charter items is to define how web applications should be given access to privileged APIs

burn: it would be good to look at that group. I want to make sure we don't wait on a model from that group before being able to put capabilities in our document

<dom> ACTION-16?

<trackbot> ACTION-16 -- Eric Rescorla to propose how to tie into identity frameworks for comms partner verification -- due 2012-01-12 -- OPEN

<trackbot> http://www.w3.org/2011/04/webrtc/track/actions/16

stefanh_: next action, putting identity information

ekr: I am a little behind on that, working on it now. I can have that by 2 weeks or so

<dom> ACTION-16 due June 25

<trackbot> ACTION-16 Propose how to tie into identity frameworks for comms partner verification due date now June 25

stefanh_: next action (25), belongs to the media capture task force. to draft initial requirements

stefanh: propose to not discuss it further here

fluffy: I had a question about that, what is the plan here?

stefanh_: the plan is to move it from this tracker and opened into the mediacap tracker

<dom> ACTION-29?

<trackbot> ACTION-29 -- Cullen Jennings to change all numeric constants to be enumerated strings -- due 2012-06-15 -- OPEN

<trackbot> http://www.w3.org/2011/04/webrtc/track/actions/29

fluffy: largely we've taken the first stab at moving most of the stuff, but we're still waiting for the respec2 move

burn: there is no need to move respec2. dom made changes to the gUM doc, but we can do the same changes to the webrtc doc
... respec 3 = respec 2 + extra module

<dom> ACTION-29: mostly done, waiting for new respec version

<trackbot> ACTION-29 Change all numeric constants to be enumerated strings notes added

<dom> +1 on moving WebRTC to respec v3

burn: it worked well with gUM, so it's worth trying with the webrtc document

stefanh_: action 42 is also mediacap
... shall we move on to the next part of the agenda? JSEPified PeerConnection API

adambe has slides on the discussion

<dom> JSEPified PeerConnection API (slides)

JSEP in PeerConnection

adambe: slides about our current PeerConnection API, I didn't really know in what form to do this. start with a simple example
... the code in the example has never been run, so there could be issues. but here's my view of how this API could work right now

as few lines as possible, the straightest line possible between a two-way audio/video call

the overview shows each step in the process

subsequent slides will go in detail for each part

who you are calling is left to the web application. variable signalingChannel is a way to send data to the other side (somehow)

create a peerconnection, a way to handle ice candidates as they come in, use signalingChannel to send the candidate from the event over to the other side

handler for handling what happens when you get a stream from the other side. in this case, we simply show the video in a video element

<dom> (as far as I can tell, the current editors draft doesn't allow "null" for the IceServers configuration param in PeerConnection param)

part2: use getUserMedia to get access to the local media and create an offer or answer (based on role)

part3: handling incoming messages sent through the signalingChannel. three types of messages: "offer", "answer" and "candidate"

anant: what is the SessionDescription constructor?

adambe: that's how it is in the spec right now, it converts the string to an object

<dom> anant: we can't add "SessionDescription" as a global object; we could either make it a sub-interface of PeerConnection, or avoid a constructor altogether by using a string

<Martin_> mic

adambe: we have a constructor to go one way, and stringifier to go another wa

the object is the place to add those, and it's the placeholder

<nstratford> Would it be possible for someone to relay slide numbers into the chat for those of us trying to follow along without video or slides in webex?

<stefanh_> harald switched to the right slide in hang-out

<DiMartini> this is the last slide

dom: I don't know if there's another WebAPI that does that

harald: get it out of the global namespace as an action item, and we can do the specific proposal

adambe to consult with dom and make a proposal

<dom> ACTION: adam to move SessionDescription and IceCandidate out of the global namespace [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action02]

<juberti> PeerConnectionSessionDescription

<trackbot> Created ACTION-43 - Move SessionDescription and IceCandidate out of the global namespace [on Adam Bergkvist - due 2012-06-18].

<dom> juberti, I think the idea was more to have PeerConnection.createSessionDescription() or something

adambe: let's see what happens when someone calls. the start method takes a boolean in this example (true for outgoing, false for incoming)

anant: on the receiving side, you're calling navigator.getUserMedia, and then createAnswer, without having received the offer

adambe: the caller side has called start(true); the callee side hasn't done anything yet

<fluffy> If this helps at all, I have a bit of a call flow at

<fluffy> https://github.com/fluffy/webrtc-w3c/raw/master/call-simple.png

<fluffy> it has some known problems

ekr: what is the sequence of events on the answerer side?

juberti: [answering, but not audibly]

fluffy: I suspect this hasn't been carefully thought about

fluffy: originally we didn't have the split and we didn't have setRemoteDescription

juberti: do we still need pc.remotedescription as an argument to createAnswer?

fluffy: what will drive this requirement is a rollback on update, but so far we don't have call flows that require this

adambe: you do addstream, and when you call createAnswer, you use the added stream on the pc as the source of information, and then you pass a separate offer as another input
... it might make better sense to get both added streams and remote description from the pc

<ekr> My general ask to the authors is that they need to provide a definitive answer to every such questoin.

harald: the difference I remember is that with createAnswer(arg) you have not committed to accept from the other side

<Martin_> There is a potential bug in the example code if the browser calls the getUserMedia callback inline because the remote description wont have been set when that happens.

i.e. no guarantee to call setRemote after createAnswer

ekr: I'd like to hear definitely whether or not I should call setRemoteDescr...

juberti: you don't right now

fluffy: this is clearly an issue where I haven't heard strong arguments one way or another. but obviously need to be defined

lots of different alternatives

if people have a preference for one or the other it would be great to hear

ekr: one of the things I need to do is to introspect the offer. is one of the ways I do this is via setRemote that's fine, but we want to try and limit side effect to calling setRemote

juberti: my expectation is that it doesn't really have a lot of side effects. unless there's both local & remote description nothing will happen

fluffy: yeah

juberti: why would we not automatically generate an answer when calling setRemoteDescr…? the answer lies is some outlying cases where there might be modifications required. but if we are always passing in the same description, then we should.

<dom> ACTION: cullen to deep dive on setRemoteDescription with justin [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action05]

fluffy: what happens when you call setRemote? are any added stream callbacks called?

<trackbot> Created ACTION-44 - Deep dive on setRemoteDescription with justin [on Cullen Jennings - due 2012-06-18].

juberti: the stream stuff is pretty clear, when setRemote is called, it triggers the callback

fluffy: we might need two different callback. onstreamproposed/onstreamaccetped?

harald: is it even meaningful to reject a stream? if the other end is sending me a stream, I can either take the data, or cause the data to not be sent

ekr: I don't reject the idea, but the spec talks quite a bit about what happens when a stream is permanently dead
... when I get told there's a stream on the other side, I can either accept, get media, or I've been told there is no more media

stefanh_: but you do hear about all this stuff. you get events on the stream

ekr: what about the event where I get a video with h.264 but I constraint to only vp8 via setLocal

juberti: you plugged in setRemote, got streams with audio/video. I don't know if you have track event, but you get stream event, you would then listen for onended, and these things are not negotiated

ekr: conversely, when you accept, what happens?

juberti: we talked about having an event where media start arriving

stefanh_: you get an unmuted for incoming data

ekr: what's the UI? in incoming call request, want to display audio/video, open a screen big enough, but don't want to do it until we know we can display it. but also before media actually arrives

harald: tentatively, I think, when you get media stream (onstreamadded), and then you get media track events saying that the stream has ended because it cannot be delivered

fluffy: using muted is bogus, because the other side may actually be muted so there's no data

ekr: I tend to agree with cullen, but I can be convinced that it can be made to work

juberti: we need to know one way or another if the negotiation completed or not

ekr: that may imply the main thread has to block until the negotiation finished?

so we'll need another event

fluffy: so we'll add another event

juberti: we need a state machine for streams

<ekr> harald++

harald: we need a state machine for streams and a state machine for tracks. audio track will be perfectly fine if we can't agree on the video codec.

fluffy: the per track state. do you want us to do that as an extension of the tracks in this document or in the gUM document?

dom: I think it belong in webrtc document

harald: we have to have a state machine in the gUM document, but extend with more event and state in this document

first we should figure out what events and states are

fluffy: the high level use case is to see if negotiation failed or suceeded

juberti: media arriving and muting should all be explicit

adambe: as in the spec right now, things are very fluffy "when a stream has enough information to know it succeeded it should unmute"
... addtrack event is after setRemoteDescr… is called. is that really a stream in that case? it's something to be negotiated. I think we should have a mediastream only when the negotiation has completed

fluffy: this ia an alternative approach, but we can make it work. there are some corner cases that we need to handle
... I think we can deal through all those issues.
... in the media stream's object it won't be in the video or audio track list (for smellivision)

ekr: what's going to happen when we add non video/audio tracks to streams?

<dom> partial interface MediaStream { attribute Smell smelltracks; }

fluffy: does it buy you anything to have separate "tracks" attribute? and not just video/audio track sets

stefanh_: it was to align with the media elements spec

fluffy: a track should tell you what the type is

harald: we discussed it a lot, and there was no case where it was simpler to have one set of tracks than 2 sets of tracks. I don't want to reopen that discussion
... ekr I think that if you want to have a dictionary instead of attributes, throw yourself at it

<Martin_> var audiotracks = alltracks.filter(function(x) { return x.type === 'audio'; })

<dom> +1 to harald

<stefanh_> +1 to harald

harald: I think it is actually extensible enough that we can add later when we need them, but adding earlier than needing them is not pleasant

ekr: I'm not suggesting that we add them right now, as a programmer it's not ideal to have things named like this

<ekr> I usually argue that the valid number of objects is: 0, 1, and infinite

<juberti> audiotracks = pc.tracks("audio")

<ekr> juberti++

jim barnett: we need a JS API to introspect the offer, and you don't get a stream object until you accept

richard: you still need a way for the browser of responding to an offer based on what it's capabilities are. my interpretation is that createAnswer is the way to do that

the question is, if there is enough information to the JS for it to know if it can accept the answer or not. there's a little bit of a chicken and egg there

<Martin_> there is a problem with the haptic track

<hta> ekr, I wouldn't mind too much if we defined tracks as tracks { audio[], video[] } rather than audiotracks[], videotracks[]. That's what I was driving at with "dictionary".

<ekr> hta, that would be preferable to me.

<Martin_> +

<ekr> how do we make it so

<juberti> that works for me too

<hta> ekr, type up the IDL you want and send it in.

<ekr> Willdo.

adambe: we can discuss this issue separately later

<dom> ACTION-44: Adam can help with when streams should be dispatched

<trackbot> ACTION-44 Deep dive on setRemoteDescription with justin notes added

[time for coffee! 15 minute break]

<nstratford> Hangouts still doesn't work for many of us - please don't turon off WebEx!

adambe: continuing example slide

after setRemoteDescription is called, on the callee side, getUserMedia is called to select a local source

if we're lucky, media will start flowing and the streams can been displayed

ekr: there are a lot of events been thrown out of this API. only the ICE event fire in the example, what about the others?

adambe: there's a subsequent slide that discusses the ICE events

ekr: but there's also PeerConnection events. it's unclear to me when all these events fire

<dom> (and do we need all these events?)

adambe: I don't think it's clear when some of these events fire

ekr: I can go in and enumerate when I think these events fire, do people want that?

<juberti> ekr: perhaps you could mark up the sample i wrote up with the event times? https://docs.google.com/document/d/1L3lMBINuLn2S7EO4APWTUaPcUMV0Ke1q2zCzZaYYba8/edit

fluffy: proposal, why don't ekr and I take as an action item to annotate when the events fire
... sometimes ICE is per track and other times it's for the peerconnection, deliberately not cleared up in the spec yet

<dom> ACTION: cullen to annotate the callflow diagram with events fired [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action06]

<trackbot> Created ACTION-45 - Annotate the callflow diagram with events fired [on Cullen Jennings - due 2012-06-18].

<juberti> ekr, https://docs.google.com/document/d/1L3lMBINuLn2S7EO4APWTUaPcUMV0Ke1q2zCzZaYYba8/edit

<ekr> juberti, thanks

adambe: regarding events and examples, most of the events, we don't really need

fluffy: it's true you don't need them, but the customer for the spec is the browser implementers who need to know when to generate them

adambe: I have suggested some discussion topics
... how to tell we have enough candidates? in the trickle case

adambe: do we need to talk about that before we decide what the events are associated with?

juberti: the "null" event may have independently of a candidate

fluffy: in nearly every I think that is the case

ekr: it is passing effectively a domstring to the interface for each candidate, then null may be okay, but for JSON, what would you do

juberti: if some ICE candidate is an object, then this would have the m-line for SDP

and real objects can be null in the DOM

ekr: there are fast path lines, there is the case where I have two interfaces but can't write to one of them

harald: for the implementation, we faked it by just doing a timeout. the browser should not decide when it's enough that we got all the candidates

either we define what's enough, or we leave it to the application

ekr: but the browser does know!

fluffy: I think harald makes a good point, enough is not the right thing here, the question is when is ICE done? there are no more candidates

ekr: the technical state in which every candidate fails or succeeds, happens about 40 seconds later...

fluffy: that's the event we are talking about

juberti: the application needs to either have its own timeout, but if it gets told if I have everything ahead of that, I'm not going to wait that long

fluffy: so we need an event that does this ICE session is done

ekr: the relevant event here is: I have now received STUN answers or given up on every possible candidate. min time: 0, max time: ~40-60 seconds after

<JonLennox> The max time depends on your rtt estimate to your stun/turn server, doesn't it?

adambe: so we need something in between 1 candidate and 40 seconds later.

ekr: how would you tell the browser this

I would set a timer at the beginning for roughly at 4-5 seconds, and timer or callback firing would send out the offer

juberti: I think it should be really short, if it takes more than 5 seconds for a candidate, you probably don't want to use that candidate

<fluffy> I think the action here should be that we define what happens when the browser wishes to indicate that it it is not expecting any more candidates to be produced - at that point it will indicate it by doing the following ?

ekr: one thing it might be relevant here, would it make sense for the application to control
... there are only two relevant events, I got 1 candidate, or I'm done.

juberti: anything in between is hard to specify

if you using trickle candidates, that will always work better than timeouts

fluffy: so null or events?
... it seems to me that the code you want to write is different in the case where you get a real candidate than when you get this event, you might want two different callbacks

juberti: there's gathering, connecting.. there's really not any linear state progression in ICE. you can't get away from an explicit callback

<JonLennox> I think there are separate state machines for gathering and connecting

adambe: to sum up, people seem to agree that there should be some information on the candidates so the app can decide when to send something off. is the 40 second event useful to anyone?

<Martin_> JonLennox, I agree

<Martin_> the new issue is whether there are different state machines for the multiple different flows that might be created

juberti: the last event is when the browser has finished ICE and it's got all the candidates that is can get

fluffy: that often happens in <50ms

so we definitely need that event

juberti: the middle event is not needed when you use trickle candidates, and it's essentially a timeout

jonathan: one issue where the 40second timeout is potentially interesting, trying to connect and nothing is working

at the NULL you can switch from pinwheel to failure message

juberti: that makes even more argument for the "now I think it's a good time callback"

stefanh_: I have a problem with this in-between event because it depends on the other side

adambe: so theres an event for every candidate, and one final event

<JonLennox> You can always try the host candidates — you never know, they might work.

<Martin_> and there are always host candidates

stefanh_: resolution is: there will be 1 event for each candidate, and one event for "no more candidates".

<dom> PROPOSED RESOLUTION: there will two kind of events: one for each candidate (to allow trickling), one when the browser has exhausted all possibilities

the middle event is left up to the application

<dom> RESOLUTION: there will two kind of events: one for each candidate (to allow trickling), one when the browser has exhausted all possibilities

adambe: next topic is the renegotiation event

<dom> [Shouldn't we make PeerConnection derive from EventTarget, to make it possible to use addEventListener/removeEventListener in addition to on... functions?]

the idea here is to have a callback or an event that would help the developer to know when to actually create a new offer on answer

<anant> +1 dom

ekr: I don't know if this is needed, but I have a question, but will this be fired whenever addstream is called?

fluffy: yes, that's the current thinking

adambe: the name should probably be negotiationneeded instead of renogotiationneeded since it can happen the first time too.

ekr: so if I add two streams, I get two of these callbacks?

juberti: the callback only fires when it's actually needed

adambe: this is quite a big topic, I don't know if we have enough to discuss it here. should the callback be triggered or not depending on the state...

ekr: this is problematic in the naive implementation of gUM that is calls onaddstreams when it responds. now I call gUM twice if the negotiationadded is called twice

juberti: only setLocalDescription changes the state, so calling createOffer without setting it won't call it

adambe: if you do two addstreams in the same event loop iteration, it should only result in 1 event

Martin_: but this would be in the gUM callback which almost certainly isn't in the same event loop iteration

ekr: but what happens when I get this callback when I'm still waiting for createOffer to return?!

adambe: perhaps we need more call flows & examples before we can dig into this

stefanh_: we're only discussing here for tracks or streams, does this also happen when hardware is removed/added, or there is a browser-level mute?

<ekr> Executive summary: I'm worried about race conditions.

but we haven't decided which way to do these, and lot of other things to consider

ekr: i'm not against this functionality just that it's defined in a way that doesn't result in problems

juberti: this call flow seems to make sense to developers on webrtc-discuss

adambe: to make the API easy to use, this is important, but it's not crucial for the functionality

Paul_Kyzivat: : do you have a any notion of replacing a stream or changing the characteristics? it introduces nasty issues

it's not unheard of to replace a m=audio line with another completely different line

you have to keep both streams live and then decide which one to keep after a while

the question is: the model you're talking about, maybe it's not rich enough to handle those cases? what do you do to your stream to change a codec?

burn: as far as constraints are defined now, the browser can change the stream midway as long as it satisfies the constraints, even if it need a codec change

fluffy: one use case is when the server switching to a narrow/wideband

richard: if we just look at need to renegotiate in SDP, I don't think we want to support changing the media type for an m-line, in WebRTC. it would be OK for us to say once you've defined characteristics with a media line then make it be immutable

<Martin_> port => 0, or a=inactive

richard: if you need to renegotiate in order to add a new media line, you also want to list all codecs that are present in other lines, when creating an offer you want a list of all capabilities

juberti: in some cases you do, in some cases you don't. in the JSEP draft I say the cases where you'd need a full offer

but for some cases where you are only adding one track you don't need the full offer

richard: doesn't the application need to be able to define that/

Andrew Hutton: does this renegotiation happen, for instance, when direction of an m-line is changed?

juberti: the only way to change the direction is via client setLocalDescription
... the whole idea is that when you get this negotiation callback, the developer creates an offer and ships it off

stefanh_: I would like to conclude this discussion… we are moving into IETF territory

the consensus seems to be that we need this callback, but editors need to define in what cases

adambe: other topics: constraints that we can add, new global object IceServers, createProvisionalAnswers, ICE restart
... when we get a stream, how many places in the API can I have an effect on the workings of the system? what are the possibilities of introducing conflicting constraints

if we count tweaking sdp from string and back to object, we have 5 places, and it feels like a lot of places where we can tweak

juberti: 1, 2, 3, 4 are all needed, and 4 and 5 seem the same to me

ekr: is there a 4 in the spec? do we need it?

<ekr> sdp.tweakOffer = function(f) { this = f(this); }

juberti: there will always be cases where we won't provide what the application wants (and they have to do it by hand), but for streams and gUM they are seperate.

adambe: I agree that the intention is to modify separate things, but we have to be careful that we don't introduce conflicting constraints

the reason or adding #4 is that, we should provide APIs to tweak the SDP

<ekr> my brain seems to be failing: where is the constraints algorithm currently defined?

fluffy: I agree no-one should parse the SDP on their own, but I'm hoping that constraints will cover all the things we need to do

<dom> ekr, in getUserMedia

<dom> ekr, http://dev.w3.org/2011/webrtc/editor/getusermedia.html#methods-3 more specifically

stephanh: wouldn't it be confusing if one constraint in getUserMedia could also be set/overriden in addStream?

ekr: I'm less concerned about 1vs2 than I am about 2vs3

<ekr> oh, I see, it's just not where I expected. Thanks

juberti: we need to have a clear indication about what constraints go into which API calls. you can't pass ICE restart into gUM

<dom> (I think this means the constraints registry should make which constraint for which context abundantly clear)

juberti: there's a 2nd parameter to the constructor where you put ICE constraints

fluffy: some are perfectly willing to put relays for audio, but not putting video. constraints will be different for two different cases
... let's do an easy one like aspect ratio. If I set aspect ratio in #1, will that be remembered, or do I call it everytime?

juberti: 1 gets carried over to 2; but if I add the stream to two different streams, then I can override

<ekr> correction: 5 doesn't exist

harald: this particular point illustrates that setting constraints have to fade at some point, because in the current setup it is easy to define conflicting constraints

ekr: 5 exists, 4 doesn't

<ekr> oh, you're right

<fluffy> Ted has got the video reflected into webex for the folks on webex

burn: I think that there will be subtle differences in interpretations of constraints in the different cases unless we define the context

dom: does the registry ask for context for constraints?

burn: currently doesn't but we can add it once we know what we want

adambe: : for ICEServers, we have two suggestions: list of string, list of list of strings

<ekr> so, I think we still didn't work out the merge algorithm

dom: first easy change is to make it a dictionary

<ekr> OR when it's needed.

<fluffy> we can't hear whoever that was

adambe: I think you're right

<ekr> CAn the chairs keep this issue open?

<stefanh_> it will be kept open

<ekr> stefanh_, thanks

<dom> (the actual syntax would be "DOMString[] servers", not "DOMString servers[]")

adambe: do we have any requirements of different ICE constraints on different servers

harald: that might make sense

<Martin_> can someone explain how SRV interacts with this while you are at it?

<hta> ACTION: anant to write up a spec for IceServer object, and compare [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action07]

<trackbot> Created ACTION-46 - Write up a spec for IceServer object, and compare [on Anant Narayanan - due 2012-06-18].

dom: in your example, PeerConnection has null as the value, the draft doesn't allow null.

anant: I think we should allow null and the browser should have defaults.

adambe: for createAnswer, do we need the offer argument or can it automatically grab it from the pc?

<JonLennox> Martin_, I'd think that'd be defined by the STUN/TURN URI definition?

<dom> (on top of make it nullable, we should also make it optional then)

fluffy: no-one could come any reason for why we couldn't remove the argument

<Martin_> JonLennox, it's pretty vague in the STUN URI draft

<dom> ACTION: Anant to provide a code example showing continuation for createAnswer [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action08]

<trackbot> Created ACTION-47 - Provide a code example showing continuation for createAnswer [on Anant Narayanan - due 2012-06-18].

<JonLennox> Martin_, should be fixed there then

<Martin_> JonLennox, I'll take it up with the authors

harald: this the 3rd redesign in 6 months, and I dont' want a redesign without a compelling reson

6 months ago, I would settle for appealing reasons, but at this point I'd rather have a compelling reason

fluffy: we haven't designed error handling yet, this may fall in this category

harald: want to get into SdpType before lunch. having them twice is wrong, we should settle that

ekr: we should have a new method call

<ekr> that was sarcasm

harald: we should try the polling method. who would like to have SdpType inside or outside?
... 1st question: do you have an opinion?

7 opinions

how many prefer to have the type inside the sdp object: 5

how many prefer to be outside: 2

conclusion: put the sdptype inside, remove the additional parameter

harald: we can have the discussion about mutability later

<Martin_> You missed the fourth and fifth questions, which are who thinks that the colour of the bike shed doesn't matter

ekr: certain things are errors, but mutating it to wrong values is an error


<juberti> derf i would prefer that it be mutable, but yes, that could be a less elegant workaround

<juberti> it shall be green: http://mamdblueroom.files.wordpress.com/2010/11/bikeshed2.jpg

<burn> scribe: burn

Statistics API proposal (http://www.w3.org/2011/04/webrtc/wiki/images/7/7d/June_11_Stats.pdf)

hta: vital need for statistics, but often left until the last minute, so i wrote something

<dom> Stats API proposal, from Harald

hta: statistics not intended for end user, mainly for service provider. Is everything actually still working?
... since service provider's only access is API, stats should be there
... should reuse meanings in other statistics collection approaches
... MediaStreamTrack is the core unit for collecting stats. Feedback from recipient to sender is important.
... since all of the data we care about is time-varying, need to timestamp everything
... means we will need to sync clocks (or equivalent), but lots of world knowledge here.
... user JS calls GetStats() on pc, then callback returns info
... model includes a pointer to track, local/remote data sets, data items are key/value pairs with keys in a new (?) registry
... define some MTI stats such as packets and bytes, IP:Port
... anyone can propose new statistics for registry. Need to distinguish between unsupported statistics data item and no result for that item.
... need aggregated statistics (MediaStream, all PC)
... maybe schedule periodic callbacks as well. The latter two may not need to be in version 1
... one challenge is that not all info is known to browser

<fluffy> one comment on OS audio path, echo cancelation often estimates the round trip

hta: another is that synchronized stats are needed for aggregation, but can't always exactly correlate sender and recipient data
... (jumps to "issues solved elsewhere") JS solves this

anant: setInterval doesn't control when callbacks occur

dom: you made this async because collection can take time?

hta: if i can't guarantee getting back to you within 10ms, i shouldn't block. sometimes may need to call out to external module that could take time, although usually it won't.

adam: can you say "collect for 10 secs"

hta: don't want to. count in the core and use callbacks to compare and do the calculation

dom: in zakim, eg, can ask who is making noise and it will wait for 10 secs

hta: should be done at JS level

stefan: have you been thinking about the data channel?

hta: no

stefan: i don't think we should have stats

cullen: web sockets doesn't' have stats but is visible to browser

randell: info is useful to app. bytes queued are available in websockets

cullen: at least need bytes xmitted and received

randell: per data channel, or global?

cullen: not sure

hta: difference from media is in data channel app sees the bytes, but not for media

cullen: want to know what happened on network

randell: there could be other useful info

magnus: about data channel, also have partial reliability option. may need to know reliability stats

hta: RFC for STCP MIB exists?

<DanRomascanu> nobody implements that AFAIK

hta: (continuing with slides) another challenge is model problems
... eg, where to count in FEC streams, where stats go for removed streams, how you count for multi-stream tracks

<dom> SCTP Management Information Base (MIB)

adam: where are counters in the first place?

hta: conceptually they are attached to a MediaStreamTrack. You need a handle to the track to get data

dom: why not put the stats method on the track object itself?

adam: +1

adam: it can remain as an ended or finished track

<fluffy> I like HTA idea of never removing a track

jonathan lennox: there are post-repair stats for RTCP

scribe: there are also multiple remotes. result of tomorrow's discussions may make this more complex

hta: don't want to support transport relays on multicast in v1 or rule out doing it in the distant future
... with multi-stream tracks, how do I count only once even though only sent once

ted: just count once. if you count for a particular track, you are right. However, adding up counts for all tracks will not add up to the number of bytes sent. Not a problem as long as app author knows what they did

justin: track in multiple streams might be sent more than once due to different encodings

randell: could be different processing on tracks too

justin: should show up multiple times

hta: maybe instead of MediaStreamTrack as selector, could query track for what to query to find out about its stats. Then ask PC for the info.

anant: what is same stream/track is added to multiple peer connections

cullen: sounds too complicated. better just to know what are all the objects to query

stefan: why can't this go on the track?
... its all on the receiving side

(several): disagree

stefan: then the sides need to agree in advance on this info

hta: yes, RTCP

magnus: need a clear model for how to handle multiple encodings of same media source.

justin: on remote side, what would they see if you had different encondings? Two tracks, right? Because different SSRCs. Maybe then we need to clone track rather than using multiple times

cullen: this would get with propagating use up to gUM for camera changes, etc.

(missed some)

randell: adding add'l semantics on top of media stream tracks that already exist.
... network media tracks add info on local streams/tracks
... tracks in PC are not necessarily the same as those returned from getUSerMedia

anant: make media stream tracks immutable so you can't change their characteristics after creation. it has fixed properties. if you want to display different resolutions in different images, then those are different tracks. can derive one track from another.

justin: but if want to change resolution, will need to create a brand new track.

ekr: what if other side changes resolution

justin: benefit of making immutable? 1-1 identity is nice, but why does that mean you can't change an existing track

anant: avoids having to change constraints that may conflict for derived tracks, where we would have to distinguish between changeable params and others that arent

<Martin_> I was observing that there are four MediaStream sub-types; LocalIdealMediaStream, LocalPacketizedMediaStream, RemoteIdealMediaStream, RemotePacketizedMediaStream

anant: can deal with remote changes differently

randell: if track is sourced from video element, source-encoded, then you change the track?

derf: this could happen at every keyframe!!!

anant: should be forced to create a new track if characteristics change

justin: can happen just by grabbing scroll handle

randel: encoder might do this itself

anant: SDP doesn't have all that?

(several): no

jimb: perhaps anything is SDP shouldn't be changeable, but everything else is okay?

cullen: SDP does specify an envelope within which you can operate. I would still expect to be able to change SDP

randell: request resolution changes may be able to happen without SDP changes, sometimes might.

ekr: benefit of immutable?

anant: video has fixed size. video doesn't know hat resolution is being received on track. more complex now in fixed output if track is changing under the covers.

randell: already handled today

justin: happens for html you download too
... want to avoid downscaling

randell: always latency between UI resize and change in the source. ALso may not cause a resize (say if different parties have different sizes for same stream)

justin: may go from small to large display and need fuller sending, but that doesn't change other small images.
... many reasons for this

<fluffy> I want to insert myself on Q

stefanwenger: may or may not be value of renegotation for change of resolution, but there are *many* SDP params that can change (framerate) during stream lifetime
... idea that stuff that sits in SDP without renegotiation not true for 264 and, i believe, VP8

cullen: we agree that two different windows is two tarkc objects. we just don't agree with immutability of a track

jimb: what is immutability? can a track change from audio to video? of course not, so that's one kind of immutability

hta: will modify proposal to have another layer of indirection so that in simple case we can get just one piece of info back but to allow more complexity

dom: question about privacy. some of the info available (remote ip and port) might be additional.

hta: don't see anything yet that hadn't already been exposed
... did say that data must be possible to be anonymized

anant: API is getStats, callback. Perhaps instead should be event that can be registered for regular returns

hta: concerned about timers that no one is still around to listen to

richard: RTCP also has application data that should be returned / received

randell: data channel API would be better way to transmit such info.

hta: if we find later that there is other info available in browser that other browser needs, RTP may be way to communicate it
... application data has multiple meanings

<fluffy> +1 lenox

lennox: app data is stuff for your app, not something standardized. if standardized, not "application data"

ddruta: question about remote sources for stats. where does app connect.

hta: whatever is sending RTCP reports .

druta: should we have param that specifies URI?

hta: perhaps could extend that way, but I need to see the use case before we go beyond remote browser

stefan: what's next?

hta: will come up with new proposal that can handle multiple stats per track.

dom: will be separate spec, or part of main one?

hta: if quick, should be part of main doc

<scribe> scribe: DanD

Data API

<dom> P2P Data API slides

adambe: [Showing example from the slides]

adambe: example creating a datachannel with an active peerconnection

fluffy: We need to add the same thing that we do for media for data

jesup: there will be no offer answer for datachannel

fluffy: I'm on board with this proposal

anant: Complicates the case as it combines the everything in one connection

adambe: we talked about negotiation call back
... you will only have to create an offer for the first channel

Richard: Why isn't data treated like the other media?

hta: We had this discussion on the mailing list

dom: there are differences between media and data

hta: I proposed for unichannels for datachannel

Richard: It seams to be the need to create a construct datachannels

fluffy: we need to write down and we need to negotiate the lines in SDP. We're going in the right direction

adambe: You are right. It can be a container for multiple datachannels

fluffy: how do I know how to receive datachannels?

<fluffy> @dan - you get a callback on the PeerConnection that tells you there is a new data stream

<fluffy> you need some out of band info to know what it might contain

<fluffy> I think we can do a little better than that

Ted: I agree with Cullen. Designing it on the fly in the room is not productive

jesup: I can write up a proposal

adambe: we have a facility but is not in Javascript

<dom> ACTION: Jesup to write up possible directions for datachannels in peerconnection and relationship with media streams/tracks [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action09]

<trackbot> Sorry, couldn't find user - Jesup

burn: It seams that we're treating datachannel as a track
... we don't have a container to hold all the datachannel

<dom> ACTION: Stefan to pester Jesup to write up possible directions for datachannels in peerconnection and relationship with media streams/tracks [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action10]

<trackbot> Created ACTION-48 - Pester Jesup to write up possible directions for datachannels in peerconnection and relationship with media streams/tracks [on Stefan Håkansson - due 2012-06-18].

justin: datachannels are very application specific

fluffy: I'd like to challenge this. CLUE might be able to use this

hta: We need to add the use case for data channel standardidation

jesup: going over the slides

<dom> Data Channel Issues, slides by Jesup Randell

jesup: Open Issues are when can you send data on the datachannel

jesup: Second issue is when can we call create datachannel

adambe: how can I connect datachannel if I don't have a peerconnection?

dom: p2p data is very useful for developers with or without media
... we should not make the assumption that media is used

jesup: proposal to create offer
... to create datachannel before createoffer

ekr: We need a datachannel container as burn suggested

ekr: It is an expessive task

jesup: renegotiation need is application specific

Stefan: you cannot treat renegotiation needed with delay

Richard: If we don't have a construct for data channels
... first datachannel is special

<dom> (note that data channels have at least two different types: reliable and non-reliable; I'm not sure how that is dealt with when some channels are reliable, and others are not)

Ted: We have to consider resource utilization (radio) when keeping these datachannels alive

jesup: If you decide you're done with the datachannel you can drop it
... when there's no data it makes sense to shut it down. If you do shut it down you're left with nothing. Back to square 0
... I don't have an objection

Paul: to support exposing this object. If there are errors there's no place to report them

fluffy: agreed with the error handling and add statistics to the case

burn: I'd like to see this explicit object.

<Martin_> from far enough away, everything looks the same

burn: from an API perspective it looks like a track

hta: doesn't really match

JonLennox: You need to know that you can't create the objects

<dom> have we come to a conclusion about the mystery data track object? is this discussion part of Randell's previous action item

jesup: THe question is when can you call Send (from the slide proposal)
... if we allow before send we can reuse code written for websockets

fluffy: I'm not worried about interoperability with websockets. More interested on error handling

jesup: being application specific, application can figure out

hta: if app really needs this it can build it. If you don't have early data it can fake it. I don't favor early data

JonLennox: it's not clear to me what's the different between I'm connected and I can't send data to I just can't send data

Ted: There's no such thing as early data. It's just data

jesup: I you can create the connection before, better

hta: should we poll for this?
... a lot of people have oppinions
... decision not to support early data
... coffee break

Stefan: there was support for container

<dom> ACTION: Adam to work with Randell on a proposal for a data channel container [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action11]

<trackbot> Created ACTION-49 - Work with Randell on a proposal for a data channel container [on Adam Bergkvist - due 2012-06-18].

<stefanh_> scribe: stefanh_

First topic after coffee:

Report on status Audio WG.

WebRTC & testing slides

dom: THere has been some controversy over what API to pick from two proposals.

However, now the group has agreed on one API: the Web Audio API

WebRTC & testing

dom: Next steps as we continue develop the APIs.

Document stages FPWD LCWD (several of them usually) CR

At CR we have to prove that the spec is implementable

and that different implementations implement the spec in the same way

testsuites are created for this purpose

one or more testcases for each MUST in the spec

<burn> scribe: burn

dom: similarly for MUST NOT
... why do we need to do this? of course the process requires it, but more importantly interoperability is crucial for adoption and success of standards
... additionally, writing test cases *REALLY* exercises the spec language, pointing out where interpretations need to be clarified
... Although test cases are required for Candidate Recommendation, it's best to start as soon as the spec begins to stabilize. There is an obvious trade-off between getting it done early and being forced to update tests often as the spec changes.
... but tests can be written for stable parts of the spec. Some people/orgs are test-driven, requiring a test to be provided for every change request, but this can result in many changes.
... Best is not to wait too long. We should set up the testing framework before Last Call, and ideally begin writing tests as well.
... Often no one in the group wants to write tests. However, often others outside the group find it fun. It is a great way to improve the specification and does not require agreeing to the intellectual property statements that members must agree to.
... It's also a good way to really understand how the spec works -- if you can't write a test for it, the problem may be with the spec.
... Best practice is to have one or more test facilitator(s) per spec to oversee work. The facilitators do not have to write all the tests, just ensure they are written properly, getting done, etc.
... Most JS-based working groups now use testharness.js (assertion-building primitives), with a repository per spec in dvcs.w3.org. Each group needs to decide on the process for submission and review.
... Process could be "submit, review, approve" or "submit, approved" until proved wrong. If there is a formal review process details about the review need to be defined in advance.

burn: review process does not have to be laborious or complex. can just have writers review other writers' tests, and vice versa.

dom: (now showing test case(s) he wrote for getUserMedia)

dvcs.w3.org/hg/media-capture/file/de85fe3f590f/submitted/W3C/ (if I got it right)

(now looking at dvcs.w3.org/hg/media-capture/file/de85fe3f590f/submitted/W3C/video.html)

dom: library provides two different kinds of tests: synchronous and asynchronous

dom: in this example, he calls getUserMedia and verifies three assertions: there is a LocalMediaStream, no audio tracks were returned, and at least one video track was returned.

anant: why do you call t.step inside the callback?

dom: that might be a bug.

hta: what's the procedure for running these against implementations?

dom: browsers usuallly run the tests on their own. If they don't pass and they think the test or the spec is wrong, they then contact the WG
... also, the second js library allows for integration into various test frameworks for automated testing (for tests that do not require human judgement)
... Now for specifics for WebRTC. First, how do you test constraints interoperable? Second, how do you have peers to connect to? Also server-side components that we may need ref implementations for. We also need to make sure there is not a failure in the protocol itself (beyond the API).

JonLennox: if ICE connection fails, need to do XXX. These kinds of tests are needed as well.

dom: yes, network conditions need to be simulated as well.

<stefanh_> scribe: stefanh_

juberti: should we create a sw test harness with virtual input devices virtual network etc.?

hta: dom is already in contact with chrome test people

ekr: we will do this for firefox

cullen: when the discussion starts we can contribute

<dom> Scribe: dom

hta: we're expecting a Mozilla volonteer for testing!

Back to JSEP

adambe: we talked about sdptype on media description
... you could set the type as provisional either as a param to createAnswer, or by setting the attribute in the generated answer

justin: as far as I know, the only meaning of provisioning vs final answer,
... the final answer ends the offer/answer exchange
... it only affects the state machine, not the actual offers/answers that are generated
... so the only effect of that parameter would be to set the type to pranswer
... based on previous discussions, we have already identified that the type attribute needs to be mutable
... I also object to this ad-hoc parameter on the method

ekr: I think I agree with Justin here

cullen: setLocal would behave different with pranswer
... I would put it as a constraint

richard: there seems to be a potential need for the answer to inform the offer
... whether or not the intention behind it is provisional or not

martin: the decision is always made by the application

justin: it actually matters: there are some cases in which treating an answer as a pranswer is ok, but it's not ok to treat a pranswer as an answer

richard: OK from which perspective?

justin: at the callee side, the person generating the answer, the app decides whether to mark it as a pranswer or an answer
... the caller receives something; if he deals with a pranswer as an answer that's bad

justin: it's probably OK in the reverse

richard: in SIP, pranswers are not exposed

justin: if a caller treats an answer as a pranswer, then the callee assumes that the state machine is in a stable state when it is not

adambe: to summarize, we can either treat is as a constraint, or use the fact that the type attribute is mutable in the offer object
... so, should we have a constraint for it?

justin: a constraint would probably be fine

dom: what would we need several ways to do this?

cullen: linked to error handling
... this depends on things we haven't looked at, so I don't think we can really make a decision

ekr: if it turns out we need to know that type, I don't think we should stuff into constraints
... It really doesn't seem like a constraint

<ekr> What I'm saying is that if we do decide we need this, putting it in a constraint seems pretty gross

<ekr> it's not clear to me why it's any better than an extra argument

<ekr> Obviously, it's just a taste issue

adambe: so, we remove the additional argument; if we need it as a constraint, we'll add it back later

[discussion about the value of constraints as a host for this]

justin: I would prefer we avoid a bunch of positional parameters
... a dictionary with options would be much better

dan: constraints were not designed for parameters

adambe: yeah, I think we should have a settings dictionary

<scribe> ACTION: adam to look at replacing mediaconstraints in createAnswer with a settings dictionary [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action12]

<trackbot> Created ACTION-50 - Look at replacing mediaconstraints in createAnswer with a settings dictionary [on Adam Bergkvist - due 2012-06-18].

harald: what on earth does it mean for the error callback to be optional?
... I see no reason to make it optional since the app stops when error occurs

anant: continuation would help here as well

martin: this is similar with things done e.g. in XHR

<Martin_> setTimeout

tim: making it required would at least raise the chances that people copy & pasting the code would deal with error

anant: another approach is to deal with errors as part of a single callback signature à la node.js

<Martin_> node.js uses doSomething(function(err, value) { }); It's a nice pattern.

adambe: moving on to ICE Restart
... should we have an explicit updateIce() method to reset the IceServers configuration

justin: in RFC@@@ says that restarting ICE is done by changing @@@

<stefanh> scribe: stefanh

<JonLennox> RFC 5245, changing ufrag and password

discussion on restart ice

usernam+password change

(scribe a bit lost)

general design: most apps will never call update ICE

but what drove is that an app might be after a while willing to supply non-realay candidates

Ted: is there not a need to be able to restart ICE but the app does not supply username+frag

justin: what we need

api call "generate new one and restart ice"

the new username+password must be supplied to the server

adambe: can the server even generate all info?

does it have all info (like msid)?

thompson: an advanced server can do this

justin: we don't need the extra parameter

magnusw: can someone tell me how this works if it is the browser that detects that an ICE restart is needed.

cullen: "onrennegotiaonfeedback" signals this.

lennox: new I/F available: should signal to app

what if you have a perfectly usable 2G connection but moves into WiFi coverage

what should happen

should be discussed tomorrow

justin: what should happen when new candidates are trickled 10min after start?

cullen: what is the difference betwenn a mandatory constraint and a setting?

<JonLennox> The logic I understood of ICE was that once you converge, the way you change in the future is to do an "ICE Restart". The old selected pair is still live until a new pair is selected.

cullen asking for guidance on settings/constraints/dictonaries

ekr: should we replace parameters with dictonaries

cullen: editors will take liberties and wait for yelling

hta: chairs to bring back to rtcweb that how interface changes happen is unclear

RESOLUTION: IceRestart to be removed

<JonLennox> RFC 5245 "Existing Media Streams with ICE Running" is equivalent to trickle candidates before ICE has completed; "…with ICE Completed" says you have to send the existing selected candidate unless you're doing an ICE Restart.

<JonLennox> 9.1.1 "ICE Restarts" says "during the restart, media can continue to be sent to the previously validated pair."

<JonLennox> So adding a candidate is an ICE restart; you keep using the old selected pair until the restart succeeds.

<Martin_> JonLennox, does this imply that you need to gather on the existing network interfaces, or retry connectivity checks on previously failed candidates?

<JonLennox> You can reuse the existing gather state if you want for the successful candidates, or re-gather. Whether you re-check previously failed candidates is a local decision, depending on whether you have some reason they'll start working now.

setRemote/setLocal should accept the union of object andf string

<Martin_> correct :)

<JonLennox> What candidates to gather is the part of ICE that's the most subject to implementation choice

<JonLennox> But the point is that once you're in the "ICE Completed" state the only way to change your set of candidates is through an ICE Restart.

<JonLennox> From a w3c pov the interesting question is whether it's the application or the browser that needs to decide whether and when to do a re-gather.

<JonLennox> (And how)

<Martin_> The next trick is working out a) how to trigger ICE restart and b) how to discover that an ICE restart is needed...

<Martin_> I think we have a, but I think we realize that we also need b

<JonLennox> needed in a broad sense, including "possibly desirable"

DanB: you usuall have to touch the SDP when interoprating

<Martin_> exactly

anant: important to define for the normal web developer.

hta: we need to know what SDP things you'd like to munge before starting design an API for it

<Martin_> of course, if you go to the trouble of enumerating your use cases so precisely, you might as well drop the SDP altogether and build APIs for each use case. Understanding the use case is the hard part, designing APIs is easy.

Summary of Action Items

[NEW] ACTION: 39 to repsec2 move [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action01]
[NEW] ACTION: adam to look at replacing mediaconstraints in createAnswer with a settings dictionary [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action12]
[NEW] ACTION: adam to move SessionDescription and IceCandidate out of the global namespace [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action02]
[NEW] ACTION: Adam to work with Randell on a proposal for a data channel container [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action11]
[NEW] ACTION: Anant to provide a code example showing continuation for createAnswer [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action08]
[NEW] ACTION: anant to write up a spec for IceServer object, and compare [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action07]
[NEW] ACTION: cullen to annotate the callflow diagram with events fired [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action06]
[NEW] ACTION: cullen to deep dive on setRemoteDescription with justin [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action05]
[NEW] ACTION: Jesup to write up possible directions for datachannels in peerconnection and relationship with media streams/tracks [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action09]
[NEW] ACTION: juberti to deep dive on setRemoteDescription with cullen [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action03]
[NEW] ACTION: justin to deep dive on setRemoteDescription with cullen [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action04]
[NEW] ACTION: Stefan to pester Jesup to write up possible directions for datachannels in peerconnection and relationship with media streams/tracks [recorded in http://www.w3.org/2012/06/11-webrtc-minutes.html#action10]
[End of minutes]

Minutes formatted by David Booth's scribe.perl version 1.136 (CVS log)
$Date: 2012/07/02 13:03:56 $