Skip to content
Bootstrapping Websockets from HTTP/2
HTML
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
6455.html
README.md
draft-mcmanus-httpbis-h2-websockets-00.mkd
draft-mcmanus-httpbis-h2-websockets-00.xml
draft-mcmanus-httpbis-h2-websockets-01.mkd
draft-mcmanus-httpbis-h2-websockets-01.txt
draft-mcmanus-httpbis-h2-websockets-01.xml
draft-mcmanus-httpbis-h2-websockets-02.mkd
draft-mcmanus-httpbis-h2-websockets-02.txt
draft-mcmanus-httpbis-h2-websockets-02.xml
draft-mcmanus-httpbis-h2-websockets-03.mkd
draft.mkd

README.md

What Is This?

It is a mechanism for boostrapping RFC 6455 Websockets over an HTTP/2 stream. Other non-websockets streams can continue to use the other streams of the same connection.

What Problem Is Being Solved?

Connection behavior is not part of the semantic HTTP contract that remains available between versions of the protocol. 6455 relies on HTTP/1.1 connection behaviors that are not available in HTTP/2.

As a practical matter this means that servers that wish to transition to HTTP/2 and also offer WebSockets need to run extensive legacy stacks along side HTTP/2. This is an administrative burden and some server architecures may not even allow port sharing - creating problems with migration of existing markup. The situation dis-incentives the transition to HTTP/2.

Further, HTTP/2 brings more efficient network usage through its use of fewer TCP connections. This draft lets WebSockets into that arrangement - no longer needing a separate TCP connection for WebSockets than HTTP.

Why do we still need RFC 6455 - isn't HTTP/2 already bidirectional?

Chosing a shim on top of 6455 instead of doing a full native integration into HTTP/2 is indeed a design choice. It is informed largely by

  • Previous efforts in this space to boil the ocean have shown there is insufficient interest to bring a solution to market. The marginal wins do not justify the complexity.

  • Belief that the core problems with the current situation are driven by the lack of a unified serving architecture and can be satisfied at low risk (with low complexity) using this scheme. The other gains a full integration would provide, such as prioritization and multiplexing between websockets messages, are only considered minor nice-to-haves by the community.

Instead of overloading CONNECT, why not a new TUNNEL method?

Methods are generally end to end and, more importantly, HTTP version independent. CONNECT is already a special snowflake in this regard. Note that the only method 7540 defines is CONNECT - because all the others are inherited from the semantic layer of 723x. Extending it is fairly natural as its defined specifically for HTTP/2 while a new method would not be well known as a version specific mechanism.

If not a new method, what about a new Frame Type? That's definitely h2 specific

HTTP/2 has a well defined stream state machine that this proposal wants to reuse and introducing a new frame type would add a significant amount of complexity to the mechanism in order to extend those states. CONNECT is already special - this confines the specialness to that case only.

You can’t perform that action at this time.