On Tue, Jun 25, 2024 at 09:42:59AM +1000, Martin Thomson wrote:
> CF says:
> HTTP/2 58.5%
> HTTP/3 29.5%
> HTTP/1.X 12.0%
>
> Our numbers are a per-request metric for our Beta and Nightly channels
> respectively.
> HTTP/2 45.46% 56.30%
> HTTP/3 25.09% 26.13%
> HTTP/1.1 29.25% 17.44%
> HTTP/1.0 0.19% 0.12%
>
> We don't collect these metrics for most people, so these are skewed by virtue
> of the skew in prerelease channels. They also fluctuate a fair bit. There
> is a weekly cycle on a lot of metrics, but these are not obviously affected
> by that cycle.
>
> These are clearly not at the same level as Meta properties, but there is a
> lot of Web out there. Some of it upgrades, a lot doesn't.
>
> These numbers seem fairly stable over a longer timescale (I have easy access
> to metrics only for about 6 weeks, but the charts are basically level over
> that period).
>
> The question I'd ask is "so?" Should we care that QUIC isn't racing to the
> moon?
That's probably the most interesting question. In other words, "should
those who don't use it expect any gains by switching to it?". If so, it
means that a lack of aggressiveness prevents them from having a better
experience. But if being more aggressive means degrading experience for
a significant part of the users, maybe we're already close to the optimal
situation.
One thing I'm waiting for is DNS records. On some sites with few pages and
objects, there's extremely little QUIC. On haproxy.org, we're having only
1.6% of connections in QUIC and 1% of requests. The reasons for this are
multiple:
- the majority of the traffic is "git clone" and mirroring of sources.
When it comes to browsers, these ones mostly visit the home page only.
Analysis from logs over the last two days while eliminating most obvious
bots shows:
51.62% HTTP/2
40.12% HTTP/1
8.26% HTTP/3
- the site doesn't force a redirection to HTTPS is linked to from many
places in HTTP. If I limit the analysis to HTTPS traffic, the numbers
are a bit better:
72.07% HTTP/2
16.40% HTTP/1
11.53% HTTP/3
- the site doesn't provide tons of browsable contents (most of what
users are looking for is on the first page). Refining the search
above to clients having made more than a few requests further
changes the ratios:
58.90% HTTP/2
21.26% HTTP/1
19.83% HTTP/3
So in short, bots and tools don't use H3 and sites seeing them a lot
will keep a low rate on these. Sites made of mostly one landing page
containing the info users are looking for are unlikely to see H3 being
used that much due to alt-svc being retrieved with the page itself. Of
course, DNS could help but for now, support for the records is not
widespread among DNS providers. I do think this is what we should count
on to significantly increase ratios on small sites like above, where an
apparent 1% of requests finally is around 20% of eligible requests.
Regards,
Willy
