Re: Is "Version Greasing" a new benfit or a new obstacle?

["Brian Trammell (IETF)" <[e-mail]>]

> On 15 Apr 2019, at 11:28, Mikkel Fahnøe Jørgensen <[e-mail]> wrote:
> 
> QUIC is not going to be used only for HTTP/3 and I don’t see how TCP fallback will work for the general case.
> 
> A lot of games, VR applications, streaming services etc. are going to build semi-proprietary QUIC versions, or at least use QUIC with customised application protocols where TCP fallback is simply pointless because it doesn’t support the framing needed and if it were possible, why would adapt an application layer protocol to work on TCP if you don’t have to - of course you have to if it gets blocked - hence try to avoid that.

So you see the deployment strategy for these applications to be "zero availability on non-UDP-accessible networks"? That seems bold to me unless you're talking about QUIC being used as a replacement for "application-provided transport layer over UDP", which is certainly a big part of the market but not all of it.

You could always tunnel UDP over TCP of course, but from a transport layer perspective (and a transport-visible wire image perspective), that's a TCP fallback.

> I don’t have any numbers on transparency, but given the reach of voice call applications, you can probably safely assume that UDP will be supported through most networks.

When we looked at this, admittedly a few years ago, we saw UDP blocking on par with the QUIC failure rates reported by Google in 2015 - O(3-6%) of paths. That's one nine of availability pathwise, which is... not great.

> The problem again is if only some UDP is allowed through traffic inspection or at least port numbering.
> 
> It can be a good thing that a QUIC version is detectable if the alternative is that anything unknown gets blocked. That is the big unknown.

As I said in my other message this morning, I don't think version negotiation greasing is useful at this time because there is no reasonable model of a wire-visible version difference where an on-path device might want to let through one greased-version and not another greased-version, because the wire-image profiles and semantic properties of those two versions are by definition of greasing identical. 

At this point, given connection-availability as the metric to optimize for, it seems like reducing features that might be mistaken for QUIC-distinguishers in the wire image would be useful, and version greasing seems like one such feature.

> Will it help to be is opaque as possible, or not? Some of the scenarios I posted earlier show that no matter what you do, something will get in the way.

This all comes down to the threat model, and the model of incentives for on-path devices.

I personally think the correct design target is to put trivial QUIC-discriminability into the wire image, since there will be a lot of pressure to distinguish QUIC from not-QUIC traffic regardless of what we do, and if we don't put grippy bits in the wire image, the distinguishing functions will grip on to the slippy bits, with difficult to predict outcomes. I recognize I am in the minority (of people who speak up) on this issue, though.

Cheers,

Brian

> On 15 April 2019 at 11.20.25, Brian Trammell ([e-mail]) wrote:
> 
>> hi Mikkel,  
>> 
>> perhaps a bit late to this party, I'm not sure I understand the process you envision by which the TCP fallback will go away.  
>> 
>> Once it's in, it'll get used in a high enough proportion of cases that the availability risk of turning it off, balanced against the fact that it already works (as it's pretty much mandatory for deploying in today's 95%-UDP-transparent Internet), will go away very slowly indeed. 
>> 
>> Cheers, 
>> 
>> Brian 
>> 
>> > On 11 Apr 2019, at 10:20, Mikkel Fahnøe Jørgensen <[e-mail]> wrote: 
>> >  
>> > In a not so distant future, a fallback to TCP would not be a viable alternative for a number of services. Think WebRTC. 
>> > If QUIC and other encrypted protocols become dominant, just blocking everything would literally block everything. 
>> > Of course, http(s) as we know it, it will continue to live on, but many mobile apps would not. 
>> >  
>> >  
>> > On 11 April 2019 at 10.03.14, Mirja Kuehlewind ([e-mail]) wrote: 
>> >  
>> >> Hi Mikkel,  
>> >>  
>> >> I guess it is much more likely that any of those actors would just block QUIC all together (as there is actually a fallback to TCP).  
>> >>  
>> >> Mirja  
>> >>  
>> >>  
>> >>  
>> >> On 10.04.19, 21:57, "QUIC on behalf of Mikkel Fahnøe Jørgensen" <[e-mail] on behalf of [e-mail]> wrote:  
>> >>  
>> >>  
>> >> Let’s say China makes a deal with Google to allow their search engine (a not entirely unreasonably proposition).  
>> >> The Great Firewall of China (GFC) is reconfigured to accept traffic from a specific IP range for QUIC protocol v1.  
>> >>  
>> >>  
>> >>  
>> >> Google later upgrades to v2. GFC is updated. Later, out of ignorance, some other department within said company decides that its browsers should all use forced version aliases as a rule. GFC breaks. (Assuming Googles servers are not located within the demilitarised  
>> >> zone and users use Chrome).  
>> >>  
>> >>  
>> >> Alternative 1: Google up front decides to deploy forced version aliasing. China rejects the deal because they do not want to support random traffic through GFC.  
>> >>  
>> >>  
>> >> Alternative 2: Google up front decides to deploy forced version aliasing, and China does not reject the deal, out of ignorance. Later some department in China GFC oversight realises the deal and forces Google to remove forced version aliasing, or shut down  
>> >> its service.  
>> >>  
>> >>  
>> >> Alternative 3: Google up front decides to deploy forced version aliasing and China says, of course, we trust everything from Googles IP range.  
>> >>  
>> >>  
>> >> Alternative 4: Google realises that forced version aliasing does not work in the general case, and makes a special case for users within a certain geographical area.  
>> >>  
>> >>  
>> >> Disclaimer: I’m not affiliated with Google and this might not at all be how things would work, but it does highlight some of the challenges.  
>> >>  
>> >>  
>> >> Of course, to be fair, China is not the only government entity that might have a vested interest. Let's take a fictional example of US a real estate investor with heavy ties to the eastern european block who somehow finds himself elected as president needing  
>> >> to find leverage on trade negotiations without effectively hurting personal finances. It turns out that blocking certain digital services from said eastern european block is the perfect tool in the trade negotiations.. Advisers point out that those services  
>> >> run heavily encrypted on the dark web with perpetual circulation of IP ranges so, ignoring any legal concerns, it would not be practically possible. Eventually someone figures out that these services all usually are at the forefront of technology and currently  
>> >> use QUIC v3.2 and no-one else has deployed that version yet, so it would suffice to ask NSA to tap into the backbone and disrupt specific packets. Of course, this ends up taking down the Bavarian local government election process in Germany where they are  
>> >> the first to use a new digital election system. Not that the is an issue, since trade negotiations are also running hot in that area, so that is just an accidental bonus.  
>> >>  
>> >>  
>> >> Or an endless number of other developments over the next few decades if the past few years is anything to go by.  
>> >>  
>> >>  
>> >>  
>> >>  
>> >>  
>> >> On 10 April 2019 at 20.13.59, Gorry (erg) ([e-mail]) wrote:  
>> >>  
>> >> Thanks Mikkel, I do understand that various actors intentionally drop - but are you saying these actors would specifically choose to block a new version of QUIC ... I do not understand that assertion.  
>> >>  
>> >>  
>> >> Gorry  
>> >>  
>> >> On 10 Apr 2019, at 18:25, Mikkel Fahnøe Jørgensen <[e-mail]> wrote:  
>> >>  
>> >>  
>> >>  
>> >> China blocks as a rule.  
>> >> Russia is running an experiment to block the rest of the internet.  
>> >> USA blocks net neutrality.  
>> >> EU blocks cookies.  
>> >> GB blocks itself.  
>> >>  
>> >>  
>> >> So blocking is not limited to what an operator considers best for business.  
>> >>  
>> >>  
>> >>  
>> >> On 10 April 2019 at 18.59.15, Border, John ([e-mail]) wrote:  
>> >>  
>> >>  
>> >> I understand why people want to come down on the side of preventing ossification. But, things have changed. There are a lot of more negative consequences now when people unnecessarily block things. I think operators would put a lot of pressure on vendors to  
>> >> not do it now and to fix it if they did "by accident". Of course, I am only one operator. It would be nice to hear from others...  
>> >>  
>> >>  
>> >>  
>> >> John  
>> >>  
>> >>  
>> >> -----Original Message-----  
>> >> From: QUIC <[e-mail]> On Behalf Of Roberto Peon  
>> >> Sent: Wednesday, April 10, 2019 12:52 PM  
>> >> To: G Fairhurst <[e-mail]>;  
>> >> [e-mail]  
>> >> Subject: Re: Is "Version Greasing" a new benfit or a new obstacle?  
>> >>  
>> >> WARNING: The sender of this email could not be validated and may not match the person in the "From" field.  
>> >>  
>> >> CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender and know the content is safe.  
>> >>  
>> >>  
>> >> You're kinda between a rock-and-a-hard-place either way:  
>> >>  
>> >> - We've seen how much fun ossification is in TCP and HTTP. If the thing is observable, it will be ossified seems to be the lesson. A lot of the reason why QUIC was started in the first place was because of the inability to improve TCP due to this ossification.  
>> >> - OTOH, there is the fear of unknown/unobservable which might cause operators to block things, whether predictably or not.  
>> >>  
>> >> My opinion is that it is better to start with preventing ossification, and then if that results in too large a percentage of operators blocking things, to re-evaluate.  
>> >>  
>> >> My guesses:  
>> >> IP+port tuples and traffic patterns are still observable (for better and worse), which implies operators will still have significant tools for managing traffic. I believe that these are acted on/matched (ML or not) regardless of any other data presented. In  
>> >> other words, I have a doubt that stating the version in an observable way will prevent the use of such tools.  
>> >>  
>> >> Most problems I've seen associated with implementations rather than protocol versions (though when the latter happens it is pretty severe).. If you believe this assertion, then acting on protocol version is less interesting than attempting to act based on implementation  
>> >> fingerprints.  
>> >> -=R  
>> >>  
>> >>  
>> >> On 4/10/19, 1:48 AM, "QUIC on behalf of G Fairhurst" <[e-mail] on behalf of  
>> >> [e-mail]> wrote:  
>> >>  
>> >> Obscuring the version of a protocol seems like a major design design  
>> >> decision for wider use cases. So, I'm trying to understand the  
>> >> motivation for version greasing.  
>> >>  
>> >> (1) I know there were instances where some early versions of QUIC were  
>> >> blocked due to an uninitentional matching of the header. Is there  
>> >> evidence of intentional attempts to block updates to protocols?  
>> >>  
>> >> (2) Thinking about operating a network that cares about user support and  
>> >> protection from unwanted traffic, I would expect that there would be  
>> >> cases where traffic pattern anomolies are found and the appropriate  
>> >> thing would be to try to determine if a new protocol had been deployed  
>> >> and monitor it, if not, then the next most obvious thing could be to  
>> >> block all unexpected traffic, that seems like a decision to hide the  
>> >> version could increase ossification for new versions in these cases..  
>> >>  
>> >> (3) Similarly, if a threat is known to impact only a specific (older)  
>> >> version, it would seem to motivate a drop of that traffic that seeks to  
>> >> use that version, while still permitting other traffic. Forcing version  
>> >> detection to use pattern matching/ML will lead to less predictable  
>> >> outcomes, or blocking based on address, etc.  
>> >>  
>> >> (4) This obfusticates the most basic piece of reporting information used  
>> >> for support. It hides the extent of deployment of the current protocol  
>> >> version and prevlance of old implementations.  
>> >>  
>> >> (5) On the support, if a problem only emerges when a particular version  
>> >> is used with a particular address, then this helps pinpoint the issues.  
>> >> Matching client versions to servers is much more of an issue if the user  
>> >> community uses a wide range of servers (less so, I expect for major  
>> >> providers: google, facebook, etc, etc), but significant when there is a  
>> >> use of a diverse set of external sites and sites with their own load  
>> >> balancers, etc and a need to manage interactions with L2 services.  
>> >>  
>> >> I am intersted in knowing if this is likely to benefit or be a new obstacle?  
>> >>  
>> >> Gorry