Re: [Ice] ICE PAC: When to start the timer waiting for possible peer reflexive candidates?

[Christer Holmberg <[e-mail]>]

Hi,

I don’t think there will be any interoperability issues. At the end of the day PAC is only about how long to wait for candidates, so the worse thing that can happen is than an agent declares ICE failure too early.

And, no matter whether an agent knows that the peer supports PAC or not,  it should aim at sending it’s candidates to its peer as soon as possible, depending on whatever local policies. The agent should not delay sending candidates just because it assumes that the peer will anyway wait for them.

Regards,

Christer

From: Justin Uberti <[e-mail]>
Date: Thursday, 2 May 2019 at 22.28
To: Nils Ohlmeier <[e-mail]>
Cc: Christer Holmberg <[e-mail]>, Roman Shpount <[e-mail]>, "[e-mail]" <[e-mail]>
Subject: Re: [Ice] ICE PAC: When to start the timer waiting for possible peer reflexive candidates?



On Thu, May 2, 2019 at 12:22 PM Nils Ohlmeier <[e-mail]<mailto:[e-mail]>> wrote:



On May 2, 2019, at 12:13, Justin Uberti <[e-mail]<mailto:[e-mail]>> wrote:



On Thu, May 2, 2019 at 10:07 AM Nils Ohlmeier <[e-mail]<mailto:[e-mail]>> wrote:

>> I do think Nils' point is important though, i.e., if we have a bad server it will take a very long time to decide on 'last set of candidates',
>> which is probably not helpful. As such I think the potential positions we can take are:
>> a) Start the timer as soon as we have an answer, regardless of any candidates.
>> b) a) + receipt of at least one remote candidate (or remote EOC). (This is Nils' suggestion).
>> c) a) + sending at least one local candidate (or local EOC).

As we are mostly concerned about the remote side: 1) not providing us with candidates, or 2) providing us with unusable candidates or 3) providing us with candidates really late I don’t see how option c) would help in any of these scenarios.
From my point of view we should choose either a) or b).

c) is just a clarification of a), in that you can't expect to receive prflx candidates until you've at least provided the other side with a candidate, so that may be the right time for the timer to start. I don't feel super strongly about this though.

Ok. I hadn’t looked at it from that angle. So c) being a stronger a) I guess it would be okay.

I guess my only concern is that in Firefox we stopped doing a) because it caused to many problems. With that in mind would it cause interop problems if we leave up to the implementor to choose to implement either b) or c)?

I'd be fine with that, but I'd want to describe what to watch out for. Can you explain a bit more?



>> b) has a problem if the remote side doesn't send any candidates, which we want to explicitly allow.
>
> True.
Just to make sure we are all on the same page: b) is only a problem in the scenario where the remote side doesn’t send any candidates but also does not send EOC.

The EOC should allow agents which explicitly don’t want to provide candidate to get the timer started soon.
I think that leaves us with scenarios where the remote doesn’t provide host candidates, and it’s reflexive or relay candidates take for ever because of slow servers.

Correct, but we can't control which endpoints will send us an EOC or not. So that will always be a possibility.

Fair enough.



>> I tend to lean towards a) as the simplest option.
>
> Keep in mind that RFC 8445 is generic, so we need to to define what we mean by "answer". I guess it means some kind of indication that makes the agent assume that the remote peer has been contacted. In ice-sip-sdp we can then map that to an SDP answer.

Good point. We basically treat the SDP answer here to be something like an beginning of ICE, because we don’t have an explicit signal for that. I think in SDP based worlds there is no need for an extra signal like that. Not sure if other use cases of ICE would benefit from an explicit begin signal.

The answer in some ways is an explicit begin signal, because it contains the username/password information needed to start ICE checks.

Yeah I didn’t see your reply before hitting send on mine. Using the availability sounds like a good idea as the minimum gating function/signal.

Best
  Nils