Re: What is missing for building "real" services?

Good points, all. I second all of Alex's points.

Gili

On 07/01/2014 8:50 PM, Alexandre GOUAILLARD wrote:
> here are a few proposition on things that are really biting us, and 
> how to (perhaps) make it easier:
>
> - bandwidth control
> 1. It seems that the number one sdp munging cause is the now infamous 
> B=AS: line to put a cap on bandwidth. Since that capacity exists in 
> the underlying code, it would be great to have an API that can help us 
> put caps, either on each stream, and/or on the full call.
> 2. I also see that there is a "auto-mute" feature being implemented 
> that depend on an arbitrary threshold. It might be interested (but 
> overkill?), to give user the capacity to set that limit (currently 50k 
> I guess) somehow.
> 3. Additionally, and perhaps not unrelated, we would alike to be able 
> to decide what happen when bandwidth goes down. Right now it feels 
> like the video has the priority over the audio. We would like to be 
> able to explicitly set the audio priority higher than the video in the 
> underlying system, as opposed to implement a stats listener, which 
> triggers re-negotiation (with the corresponding O/A delay) when 
> bandwidth goes below a certain threshold.
>
> - call controls like mute / hold
> Right now, you can mute a local stream, but it does not seem to be 
> possible to let the remote peers know about the stream being muted. We 
> ended up implementing a specific off band message for that, but we 
> believe that the stream/track could carry this information. This is 
> more important for video than audio, as a muted video stream is 
> displayed as a black square, while a muted audio as no audible 
> consequence. We believe that this mute / hold scenario will be 
> frequent enough, that we should have a standardized way of doing it, 
> or interop will be very difficult.
>
> - screen/application sharing
> We are aware of the security implications, but there is a very very 
> strong demand for screen sharing. Beyond screen sharing, the capacity 
> to share the displayed content of a given window of the desktop would 
> due even better. Most of the time, users only want to display one 
> document, and that would also reduce the security risk by not showing 
> system trays. Collaboration (the ability to let the remote peer edit 
> the document) would be even better, but we believe it to be outside of 
> the scope of webRTC.
>
> - NAT / Firewall penetration feedback - ICE process feedback
> Connectivity is a super super pain to debug, and the number one cause 
> of concern.
> 1. The 30s time out on chrome generated candidate is biting a lot of 
> people. The time out is fine, but there should be an error message 
> that surfaces (see 5)
> 2. Turn server authentication failure does not generate an error, and 
> should (see 5)
> 3. ICE state can stay stuck in "checking" forever even after all the 
> candidate have been exhausted
> 4. Not all ICE states stated in the spec are implemented (completed? 
> fail?)
> 5. It would due fantastic to be able to access the list of candidates, 
> with their corresponding status (not checked, in use, failed, ….) with 
> the cause for failure
> 6. In case of success, it would be great to know which candidate is 
> being used (google does that with the googActive thingy) but also what 
> is the type of the candidate. Right now, on client side, at best you 
> have to go to chrome://webrtc-internals, get the active candidate, and 
> look it up from the list of candidates. When you use a TURN server as 
> a STUN server too, then the look up is not an isomorphism.
>
> right now, the only way to understand what's going on is to have a 
> "weaponized" version of chrome, or a native app, that gives you access 
> to the ICE stack, but we can not expect clients to deploy this, nor to 
> automate it. Surfacing those in an API would allow one to:
> - adapt the connection strategy on the fly in an iterative fashion on 
> client side.
> - report automatically the problems and allow remote debug of failed 
> calls,
>
>
>
> On Tue, Jan 7, 2014 at 2:15 AM, Eric Rescorla <ekr@rtfm.com 
> <mailto:ekr@rtfm.com>> wrote:
>
>     On Mon, Jan 6, 2014 at 10:10 AM, piranna@gmail.com
>     <mailto:piranna@gmail.com> <piranna@gmail.com
>     <mailto:piranna@gmail.com>> wrote:
>     >> That's not really going to work unless you basically are on a
>     public
>     >> IP address with no firewall. The issue here isn't the properties of
>     >> PeerConnection but the basic way in which NAT traversal algorithms
>     >> work.
>     >>
>     > I know that the "IP and port" think would work due to NAT, but
>     nothing
>     > prevent to just only need to exchange one endpoint connection data
>     > instead of both...
>
>     I don't know what you are trying to say here.
>
>     A large fraction of NATs use address/port dependent filtering which
>     means that there needs to be an outgoing packet from each endpoint
>     through their NAT to the other side's server reflexive IP in order to
>     open the pinhole. And that means that each side needs to provide
>     their address information over the signaling channel.
>
>     I strongly recommend that you go read the ICE specification and
>     understand the algorithms it describes. That should make clear
>     why the communications patterns in WebRTC are the way they
>     are.
>
>     -Ekr
>
>