On Tue, Jun 25, 2024, at 13:19, Chris Box wrote:
> On Mon, 24 Jun 2024 at 22:08, Matt Joras <[email protected]> wrote:
>> There's a lot to unpack here. Setting some context, as of today our (Meta) 
>> Internet egress to end users hovers at around 89% QUIC. The reason that is 
>> not closer to 100% mostly stems from certain products not completely 
>> utilizing QUIC yet, though every one of our platforms uses it to some degree.
> 
> Impressive results!
>  
>> Instances of UDP blocking are mostly isolated to smaller networks and are 
>> not widespread.
> 
> That's good to know. So it's clear blocking is not the cause of it being a 
> minority.
>  
>> For third party clients who use more cautious policies, the numbers are not 
>> as high. The percentage of HTTP/3 by browser as seen by our servers 
>> currently is roughly:
>> Microsoft Edge: ~83%
>> Chrome Desktop: ~75%
>> Chrome Mobile: ~70%
>> Firefox: ~60%
>> Safari/Mobile Safari: ~40%
>> Firefox Mobile: 30%
>> 
>> We use the same alt-svc and HTTPS record strategy for all our major domains. 
>> This indicates that there are likely improvements to be made in both 
>> advertising QUIC's availability but more importantly browsers using it more 
>> proactively than they currently are.
> 
> This appears significant. IIUC, those numbers could potentially reach ~89% as 
> that's what you achieve using first party apps. Safari routinely performs 
> HTTPS lookups and yet only reaches 40%. The use of old operating systems is 
> probably only part of the reason.
> 
>> This is where many small impediments can start to accrue and stall out 
>> adoption.
> 
> Yes of course for many reasons the long tail will be very slow to adopt 
> HTTP/3. I was mainly thinking about CDNs, whose business offering is to 
> deliver web content as quickly as possible. It's still surprising to me that 
> H3 is a small percentage on many CDNs.
> 
> 
> On Tue, 25 Jun 2024 at 00:43, Martin Thomson <[email protected]> wrote:
>> HTTP/3 25.09% 26.13%
> 
> Thanks for the data, which is pretty similar to Cloudflare's. If that was 
> filtered to Meta destinations, perhaps you'd see the 60% figure that Matt 
> sees.
> 
>> The question I'd ask is "so?"  Should we care that QUIC isn't racing to the 
>> moon?
> 
> I care about making the internet work better and be more responsive to users. 
> QUIC is a major improvement in that sense and it's sad to see it being 
> limited to a subset. I'd like to deliver these benefits to more people, more 
> of the time. As a community we can take actions that result in H3 overtaking 
> H2 on the internet.

Picking up on the performance angle only:

QUIC protocol features offer advantages for sure. Some of those are targetted 
at particular network types, or stream usage patterns. For instance, HoL 
blocking avoidance helps when the network is lossy and HTTP requests are 
independent; that isn't always the case though. Each web page has its own 
special flavour of critical path.

It can be difficult to monitor and prove with actual data that end-users see 
meaningful performance improvements due to leveraging the features of QUIC. The 
web platform only provides so much information. Advanced operators may have a 
bunch of client side analytics that can give deep introspection to every angle 
of their web app. Some folks might rely on RUM providers to capture 
industry-standard performance metrics such as Core Web Vitals. There are 
difficulties getting at information about the local network config (cell, wifi) 
or other access network details that would be interesting to understand if QUIC 
is addressing the needs of the network path. 

As a platform or service operator, its likely that you'll have no idea what the 
KPIs of the web applications that pass through you are. Focusing on low-level 
observables, such as connection or request times doesn't give the full, actual, 
picture. There is a disjoint, that can be hard to fill. In contrast, things 
like YouTube or Meta have purpose built clients that can do cool things. That's 
*in addition to* a culture of performance monitoring, tuning and ever 
incremental improvements. There can be layer 8 challenges to solve to enable or 
shift to any tech, especially if an organisations culture is already geared for 
it. This is not a new problem, web performance experts have been trying to deal 
with this for as long as I can remember.

There are also many other opportunities for improving performance of web sites 
that are less deeply entwined with transport details, and work across all HTTP 
versions. QUIC deployment isn't just competing with TCP, it's competing with 
every other technology that can boost performance. For example, Speculation 
Rules [1]. There's only so much engineering hours to go around, and companies 
make a value judgement about where to invest that for peak ROI.

As others have noted, some folks are risk averse and are happy enough with the 
status quo. If there are performance wins to be had, they'll likely be wanting 
to see case studies before the CTO signs off on turning something on that might 
even have the slightest hint of beung able to cause some regression in 
performance or stability.

It also behoves me to mention we have QUIC powering the web in different ways 
than the OP question presents it. Apple Private Relay, for example, uses MASQUE 
for tunneling traffic - leverage QUIC's unique DATAGRAM feature. Similarly, 
Cloudflare is in the process of migrating its WARP traffic from Wireguard to 
MASQUE [2].

We also have WebTransport and MoQ, which are intended for use on the Web. These 
deliver possibilities that HTTP can't, such as other forms of interaction, or 
addressing the particular needs of low latency video distribution. I don't 
think we'll see deployments of these shift HTTP version share numbers 
dramatically. I don't think that matters. If more people watch more video due 
to MoQ providing an excellent QoE, that doesn't necessarily change their 
existing browsing habits.

Cheers
Lucas

[1] - https://developer.chrome.com/docs/web-platform/prerender-pages
[2] - https://blog.cloudflare.com/zero-trust-warp-with-a-masque
> 
> Chris
>  

Reply via email to