- From: Mark Nottingham <mnot@mnot.net>
- Date: Fri, 9 Jul 2010 11:52:45 +1000
- To: "Thomson, Martin" <Martin.Thomson@andrew.com>
- Cc: Salvatore Loreto <salvatore.loreto@ericsson.com>, Greg Wilkins <gregw@webtide.com>, HTTP Working Group <ietf-http-wg@w3.org>
On 02/07/2010, at 3:07 PM, Thomson, Martin wrote: > That's OK. Put what you want in there. It's an upper bound thing. If the local implementation has a timeout, then it (as an intermediary that is aware of the header) can tweak the value downward. > > The value can be tuned. Encounter a gateway timeout and you can adjust the time downwards. > > The problem we are attempting to address is the fact that the server is blind. The server knows that you are long polling - because that's the way that most long-polling resources are setup. What they really want to know is how long to wait before sending their 304 if nothing changes. We want to remove that guesswork. > > Obviously, you can't just remove it straight away, because intermediaries don't help out. You have to rely on the client tuning the timeout. In the end-game, the client doesn't have to. That's the aim. > >> All of this leads me to believe that there's little value in setting a >> timeout; what we really need is just a flag that says "I'm long-polling >> here." It's a lot easier to implement support for this (e.g., in Squid, >> Apache Traffic Server, elsewhere), and more difficult for it to cause >> problems. > > If you don't accept that intermediaries can change, then I'd be inclined to agree. If that one bit is all that intermediaries get from the header, then we probably haven't helped a lot. That said They can change, but I don't see the incentive for them to change here. From an intermediary's perspective, the most useful thing here is to have a separate policy for requests that are long polling. Beyond that, rewriting headers is a pain; remember that intermediaries are very performance-sensitive. Also, the rate of change for intermediaries is a *lot* slower than that of clients or servers; many are in hardware, and many are no longer supported by vendors (unfortunately). I suspect clients will have to keep an eye on the timeout for a *long* time -- on the sort of timescale where just using websockets/spdy/waka/[insert proposal here] becomes more attractive. Cheers, -- Mark Nottingham http://www.mnot.net/
Received on Friday, 9 July 2010 01:53:18 UTC