A valid concern.
> which currently, effectively don't work.
Citation needed.
> HTTP 2 is for example slower than HTTP 1.1 over modern TLS even on near perfect connections.
H2 attempts to solve a number of problems present in HTTP 1.1 when your connection _isn't_ perfect -- head-of-line (HOL) blocking comes to mind, although since it still uses a single TCP connection, it's not a complete solution. It also provides header compression which can yield significant performance wins (bandwidth usage, latency if it reduces RTTs, CPU usage although this is always a mixed bag with compression).
H3 solves the transport-layer HOL blocking issue, and has a number of other RTT-reducing features (e.g. removing the need for TCP 3-way handshake -> TLS handshake -> actual sending of requests by mandating the use of TLS, which admittedly can have performance impacts compared to, say, unencrypted HTTP/1.1).
It does have a number of aspects which could negatively impact performance depending on your use case (e.g. userspace congestion control, always-on encryption).
If you're dealing with small resources without HTTP pipelining where individual HTTP requests can and will complete in 1 RTT (and parallelization is limited because of dependencies between resources, e.g. a HTTP page contains an iframe which contains a js file which fetches an image), then your end-user latency will be dominated by the number of RTTs imposed by the underlying transport.
And even if you include pipelining (assuming a non-buggy implementation, which inherently limits concurrency in order to reap its benefits), then you're still subject to HOL blocking, both request- and connection-level, which is problematic in moderate loss scenarios (wifi, mobile).
by multiplexing a many pseudo connections down one TCP pipe. I still don't quite understand how that was meant to improve performance over a lossy link.
It seemed to me that Http2 was meant to be a file transfer protocol, but designed by people who don't really quite understand how TCP performs in the real world (especially on mobile)
It does!
First, consider the case that HTTP pipelining addresses: lots of small independent requests. By packing multiple requests into the same connection, you can avoid handshakes, TCP slow start, etc. Perhaps more importantly, if your connection has more bytes in it and therefore more packets to ACK, then you can actually trigger TCP's detection of packet loss via duplicate acknowledgments, as opposed to waiting for retransmission timeouts.
Further, browsers typically limit the number of parallel requests, so you potentially have another source of HOL blocking if you've exhausted your connection pool (as happens much more quickly without pipelining).
That said, pipelining is also subject to HOL blocking: let's say that the first external resource to be requested is some huge JS file (not all that uncommon if you're putting <script> tags in your <head>). That needs to finish loading before we can fetch resources that are actually needed for layout, e.g. inline images.
H2 provides a handful of innovations here: one, it allows for prioritization of resources (so you can say that the JS file isn't as important); two, multiplexing allows multiple underlying responses to progress even if that prioritization isn't explicitly in place.
Yes, it still runs into problems with lossy links (as that's a problem with TCP + traditional loss-based CC algs like New Reno / CUBIC). But it's also better than the status quo either with HTTP/1.1 + connection pooling or with pipelining. And it has the noted advantage over QUIC in that it looks the same as H1 to ISPs.
And not only mobile, same applies for server-server connections over near perfect, inter-dc connections.
Even gigabit, and above links drop packets... because it's how TCP is supposed to work!
The higher is the speed, the more aggressive curve will be used by the congestion control algorithm.
In other words it will hit drops fast, be in the drop zone for longer, and this will be recurring more frequently than on slower links.
In fact, faster congestion protocols for speeds above 500 mbps get their speed by having A LOT MORE drops, not less.
Any protocol designed for the TCP transport must have been designed to work in concert TCP mechanics, not fight, or try to workaround it.
This is a perfect example of how AB test vodoo lead people down the rabbit hole.
There is a reality of running code on other peoples hardware: they impose limitations on network traffic that you cannot change. Example: Home-grade routers do NAT, but have table limitations which make it hard to have multiple connections. The connections get dropped, and users complain, not realizing it was due to them having bought a cheap router. Other actors in the network path between the client and the server impose their own limitations, making it hard to have multiple connections. (corporate firewalls being another big one).
HTTP/2 is a compromise, admitting that it is not possible to go fix or replace all this hardware. I worked on a few HTTP/2 implementations, and the complexity it incurs is significantly less than the alternatives.
What "real world data" you have to contradict Google's and basically everyone's personal experience?
I have sites that are unbearable over HTTP/1.x and instant over HTTP/2. But maybe I was fooled by Google.
Devices are not static anymore and tolerating a switch from WiFi to mobile etc is a big plus
[citation needed]
As a practical matter if your page requires a large number of resources (say 50+), and you are measuring the time to complete loading all of them, HTTP/2 loads much faster than HTTP/1.1. There are some nice demo sites which show this, with pages full of small images.
This is because all requests can be issued in parallel over HTTP/2 as soon as the client knows what to request, and this reduces the number of round trip times before all resources complete loading.
Even without Push, the client usually knows all the requests after it receives the HTML or sometimes the CSS.
With HTTP/1.1, the client requests are bottlenecked in a queue limited by the number of parallel TCP+TLS connections. Some sites boost this limit by using extra subdomains. But running 50+ TCP connections bursting at the same time is not efficient, even if it does bring down the latency by parallelising. With that many, there will be a burst of congestion and possible packet loss at the low bandwidth receiving end. And the 3-way handshake and TLS handshake needed for each new TCP+TLS means the parallel request latencies are a multiple of HTTP/2 request latency.
HTTP/2 also has request header compression, which makes a difference when upstream bandwidth from the client is limited.
Why would anyone want to do it on desktop? Let alone server-server APIs?
HTTP 2 and 3 are spec monstrosities, which currently, effectively don't work.
HTTP 2 is for example slower than HTTP 1.1 over modern TLS even on near perfect connections.
HTTP 3 seems to be even slower in real world use.
I strongly believe Google fudged its "real world trial data" to push W3C to adopt it.