Make HTTP/1 transports a behaviour

[whatyouhide]

@ericmj this is how a behaviour for transports would look like I think. IMO it looks really good, and the changes are quite contained. The ugliest change is in the Conn.stream/2 function which now has to find the message tags for the given transport and dispatch based on that as well, but I think keeping the pattern matching and everything it's still very readable.

I also had to go with XHTTP1.Transport instead of a transport shared by XHTTP1 and XHTTP2 because XHTTP2 requires stuff like :ssl.negotiated_protocol/2. We could go ahead and make those calls optional so that we can share transports. We can see later on though. :)

[whatyouhide]

\cc @fishcakez @josevalim, would love your take as well :)

[josevalim]

• if all we do is delegate (except for the message tags)
• we don't plan to have any new transport for the foreseeable future

👎 on a behaviour. We can always add it later.

[michalmuskala]

One example where a custom transport is useful in practice is when dealing with the PROXY protocol http://www.haproxy.org/download/1.8/doc/proxy-protocol.txt

The ranch architecture allows defining a separate transport wrapping gen_tcp or ssl and correctly dealing with it: https://github.com/heroku/ranch_proxy_protocol

[whatyouhide]

@josevalim those are good points, but they don't really take some kind of mocking for testing into consideration. I think if we want to mock properly in tests, this might be the way to go but I am not sure.

[fishcakez]

I am 👎 on adding a behavior too, with the added reason that I think the behavior becomes outdated at OTP 21 and we will want to use a different function calls for tcp (:prim_inet) and ssl (:ssl).

I don't think the ability to mock a socket in tests is an issue because it is trivial to create a pair of connected sockets and send/recv data from the peer - easier than setting up a custom transport.

[ericmj]

A possible problem with using sockets for tests is that we don't control what kind of buffering happens in the TCP stack. We have tests where we specifically want to send specific chunks of bytes and I'm not sure if we can get those guarantees when go through the whole stack.

https://github.com/ericmj/xhttp/blob/master/test/xhttp1/conn_properties_test.exs#L33-L54

[fishcakez]

We could use packet: 4 to force chunking

[whatyouhide]

Yeah that can work as well. @ericmj if you're on board, I'll kill this PR then and move tests to use a socket instead of a mock. I hope there won't be a performance hit for tests.

[ericmj]

Go ahead! 👍