W3C home > Mailing lists > Public > ietf-http-wg@w3.org > July to September 2006

Re: HTTP/1.1 pconns to 1.0 servers

From: Roy T. Fielding <fielding@gbiv.com>
Date: Wed, 20 Sep 2006 13:29:23 -0700
Message-Id: <63CD1E95-59D8-4B0F-B65B-3677212E7820@gbiv.com>
Cc: ietf-http-wg@w3.org
To: Mark Nottingham <mnot@yahoo-inc.com>

On Sep 20, 2006, at 10:08 AM, Mark Nottingham wrote:

> HTTP/1.0 persistent connections are documented in RFC2068. One case  
> that's not explicitly covered is when a HTTP/1.1 client sends a  
> request to a HTTP/1.0 server without a Connection: keep-alive header.
>
> My reading of 2068, 2616 and 2145 is that the fact that the client  
> indicates HTTP/1.1 in the request advertises their support for  
> persistent connections, and a HTTP/1.0 server (whether origin or  
> proxy) may safely use a Content-Length delimited persistent response.

What is an HTTP/1.0 server?

> In my testing, two popular clients that advertise themselves as  
> HTTP/1.1 devices without sending a Connection token -- curl and  
> Apple's Safari -- handle persistent connections from HTTP/1.0  
> servers very well.
>
> However, the widely deployed Squid proxy cache <http://www.squid- 
> cache.org/> does not behave in this manner; it requires a  
> Connection: keep-alive header (or Proxy-Connection, but that's  
> another discussion) in requests in order for them to persist.
>
> In discussion on their mailing list*, it seems they believe that  
> some HTTP/1.1 clients may not be able to handle a persistent  
> connection to a HTTP/1.0 server that they didn't explicitly "ask  
> for" with a Connection token.

Maybe, though it is more likely that the client will adapt its behavior
based on the contents of the response regardless of the request.  A  
better
question to ask is what happens to the connection if the client does not
understand keep-alive? If the content-length is accurate, the HTTP/1.1
client should close the connection and everyone will be happy. However,
the HTTP/1.1 client might be programmed or configured to assume that all
HTTP/1.0 servers do a lousy job of setting content-length (which was  
true
before keep-alive was deployed), so the client might just hold the
connection open waiting for more content until the server times-out.

I don't know of any clients that would do such a thing, just as I don't
know of any HTTP/1.1 clients that don't implement HTTP/1.0 keep-alive.
However, no clients should accept HTTP/1.0 keep-alive from a proxy
because of the well-known problem with dumb-forwarding proxies.

> What was the intent here? I can't imagine that it was the intent of  
> the WG to require HTTP/1.1 clients to send Connection tokens in  
> perpetuity for backwards compatibility.

It was the intent of the WG that folks who claim to be using HTTP on
the Internet should be doing so with HTTP/1.1, instead of making
silly excuses not to.  Anyone still claiming to use HTTP/1.0 will have
to fend for themselves and test against every deployed system, since
the only standard is for HTTP/1.1.

....Roy
Received on Wednesday, 20 September 2006 20:29:45 GMT

This archive was generated by hypermail 2.2.0+W3C-0.50 : Friday, 27 April 2012 06:49:46 GMT