The QUIC FaceBook jumps over to HTTP3

[iamlost]

Note: I had been looking at, testing, QUIC/HTTP3 as a down the road Google ‘thang’ that while useful was not transformative. My tests showed a ~4% improvement benefit far less than the ~67% I got switching to HTTP2 from HTTP1.1.

However, this linked FaceBook engineering post tosses that assumption out the window; QUIC is not just coming, it is here.

>75% of FaceBook internet traffic uses QUIC and HTTP/3 [engineering.fb.com] by Matt Joras, Yang Chi, 21-October-2020.

(we refer to QUIC and HTTP/3 together as QUIC)

...

Our tests have shown that QUIC offers improvements on several metrics. People on Facebook experienced a 6 percent reduction in request errors, a 20 percent tail latency reduction, and a 5 percent reduction in response header size relative to HTTP/2. This had cascading effects on other metrics as well, indicating that peoples’ experience was greatly enhanced by QUIC.

However, there were regressions. What was most puzzling was that, despite QUIC being enabled only for dynamic requests, we observed increased error rates for static content downloaded with TCP. The root cause of this would be a common theme we’d run into when transitioning traffic to QUIC: App logic was changing the type and quantity of requests for certain types of content based on the speed and reliability of requests for other types of content. So improving one type of request may have had detrimental side effects for others.

...

The experiments showed that QUIC had a transformative effect on video metrics in the Facebook app. Mean time between rebuffering (MTBR), a measure of the time between buffering events, improved in aggregate by up to 22 percent, depending on the platform. The overall error count on video requests was reduced by 8 percent. The rate of video stalls was reduced by 20 percent. Several other metrics, including meta-metrics, considering a variety of factors and specifically outlier conditions, were significantly improved as well. QUIC improved the video viewing experience, with an outsized impact on networks with relatively poorer conditions, especially those in emerging markets.
But the path to these results came with roadblocks of its own. Similar to our experience with dynamic content, we encountered heuristics in the app that had been tuned to TCP’s behavior.

...

Today, QUIC is deployed on Instagram for iOS and Instagram for Android. Both versions of Instagram have seen metrics that are comparable to or better than those of the Facebook app. Facebook and Instagram on the web also have QUIC enabled...

...

The IETF is on track to finalize the QUIC protocol as a request for comments (RFC) document sometime in 2021.

My tests over the past year have consistently shown improvements in the -4% to +4% range. Generally, as with FB above, the negative results were due to specific TLS optimisation tuning.
Note: much as HTTP1.1 optimisations had to be removed for HTTP2 to shine.

Given the relatively low improvement benefit I won’t be switching until after the standard is firmed; may wait a bit longer, though testing, just to let others ‘get the bugs out’. :)

While mid single digit improvements are ginormous benefits at G and FB scale they umm are somewhat less so for moi. :)

[JorgeV]

Hello,

75% of FaceBook internet traffic uses QUIC and HTTP/3

Considering that only MacOS11 Safari 's users have HTTP/3 enabled by default, I doubt that 75% of FB traffic is over QUIC/HTTP/3.

Other browsers supports it, but requires a manual activation of it.

[caniuse.com...]

edit: or, may be they count traffic from their own mobile app, which is certainly using HTTP/3

[phranque]

may be they count traffic from their own mobile app

Today, QUIC is deployed on Instagram for iOS and Instagram for Android. Both versions of Instagram have seen metrics that are comparable to or better than those of the Facebook app. Facebook and Instagram on the web also have QUIC enabled...