Re: [hybi] Is there a traffic jam?

[Ian Hickson <[e-mail]>]

On Tue, 14 Apr 2009, Greg Wilkins wrote:
> 
> But we can already tunnel arbitrary protocols over HTTP, so if the 
> ability to transport arbitrary protocols is sufficient, then we are 
> already there and nothing new is needed.

Personally I think that having a single underlying TCP connection and the 
ability to do this across domains are pretty important new aspects.


> To be frank, I'd actually rather use HTTP has a basis than websocket.

As I've said before, I have nothing against people using HTTP, either 
straight up as is done today, or in some other protocol like rHTTP.


> The pain points I'm having deploying bidirectional webapplications are 
> mostly about sharing connections:
> 
>  1) How to stop a browser sharing a connection by queuing
>     normal requests behind long polled requests.

WebSocket addresses this by not sharing TCP connections with regular HTTP 
traffic.


>  2) How to share limited connections between
>     widgets/frames/tabs/windows.

WebSocket and SharedWorkers can be used together to get the number of 
long-lived connections per origin (domain) down to one.


>  3) How to limit the growth of connections on the server.

I don't know how to reduce it to less than one TCP connection per origin 
per user. I'm certainly interested in proposals that achieve this without 
sacrificing the other requirements.


> Websocket potentially helps with #1, but only if upgraded connections 
> are excluded from connection accounting (obvious to do, but not sure if 
> that is specified anywhere?).

Since WebSocket isn't HTTP, the HTTP limit doesn't apply, as far as I can 
tell. However, I'm happy to edit the spec to make this clearer, or to 
introduce a limit if one should be imposed.


> Websocket doesn't help with #2 and #3 at all.  In fact it probably makes 
> it worse.  While individual connections are bidirectional, each widgets 
> will be able to each initiate their own websocket connection to their 
> own backend handling.  So the end result is likely to be a lot more 
> unshared connection.

I don't see any reason why this should be the case. Why would it be 
possible to use one pair of connections with HTTP today, but not use a 
single WebSocket connection in the future? Especially considering new 
mechanisms like SharedWorker, which allow resources from one origin to be 
shared across multiple tabs in a browser.

-- 
Ian Hickson               U+1047E                )\._.,--....,'``.    fL
http://ln.hixie.ch/       U+263A                /,   _.. \   _\  ;`._ ,.
Things that are impossible just take longer.   `._.-(,_..'--(,_..'`-.;.'