Most _domains_ may not support HTTP/3, but I fully expect that within a few years most _traffic_ will be HTTP/3.

Qualys' "SSL Pulse" says 47.1% of surveyed sites offered HTTP/2 and about 30% of surveyed sites offer TLS 1.3

Increasingly "corporations" out-source this problem to specialists who are only too pleased to use newer technologies with better performance and collect the same money.

Because 47% of sites run on Cloudflare or similar CDN that started enabling HTTP/2 for non-paying customers.

The application servers running the site do not accept HTTP/2 and most likely can't support it at all (we're a python shop and none of the web frameworks we use could do HTTP/2 when we looked into it).

But exactly that's the point. It can be handled transparently in a much easier way then the Ipv6 switch. (Because IPv6 is so much more different then just IPv4 + more addresses, and worse, many people don't realize it and treat it as IPv4 with more addresses which resulted in many problems).

For HTTP/2 at least, I think the main benefit in terms of performance applies to the "last hop", so you still get a more reliable experience even if the connection between the CDN/proxy and app server is http/1.1

True, through many web-sites don't have to care about the performance difference.

But for companies like CloudFlare or Googl HTTP/2 means less traffic overhead (multiplexing+header compression) and can save them a lot of bandwidth (aka. money) with that.

Exactly. AWS and GCP load balancers accept HTTP/2 but allow for those requests to be forward to backend instances as HTTP/1 because of this.

In fact AWS ALB does not even support HTTP/2 on the backends which is really annoying.

The fact that UDP involves much smaller PCBs than TCP alone will drive adoption of HTTP/3 because it will free up a fair bit of memory.

More availability of HTTP version gateways in load balancers and other reverse proxies is all that's needed, and that's coming along.

Protocol Control Block for anyone else wondering https://www.oreilly.com/library/view/tcpip-illustrated/02016...

You still need to keep your unacknowledged data buffered somewhere. If the kernel isn't holding it then it's in userspace.

More specifically no PCB, for UDP itself.

It's not nil. For "connected" UDP sockets, it's smaller than TCP's, but not nil because, well, buffers. And for non-connected UDP sockets there's still buffers. The main thing is that you can have much less buffer space because you might always be willing to drop packets. Ultimately you can have much lower memory pressure from those buffers and the smaller PCBs.