- From: Erik Nygren <erik@nygren.org>
- Date: Sat, 23 Aug 2014 21:23:23 -0400
- To: Martin Thomson <martin.thomson@gmail.com>
- Cc: "ietf-http-wg@w3.org Group" <ietf-http-wg@w3.org>
- Message-ID: <CAKC-DJhpbcG_xDTaJ9LfEESHA-cogzdCDUnjQ2ZXY-exHcs6Ow@mail.gmail.com>
On Fri, Aug 22, 2014 at 8:06 PM, Martin Thomson <martin.thomson@gmail.com> wrote: > I can potentially see why it might be interesting to learn about the > difference between shifting load and upgrades of various sorts. But > if that is all that you need, then we probably want to limit the > information to just that. At least one primary use-case that I know about is for load feedback as part of a load balancing control system. Some types of control systems need to split up load by how it got to the destination, especially if there are many potential ways this could have happened. Since load balancing is one of the primary use-cases described in the document, this is where it matters. (This isn't a marketing use-case at all.) Another use-case is an input into whether to return an Alt-Svc header or frame. Whether the client is already using one and which one was used by the client seems to play into most of the more obvious algorithms here. The initial proposal was just to include what was used to make the connection. For example: Host: www.example.com Alt-Svc-Used: h2="server-west.example.com:443" Using a few bits for Alt-Svc-Used enables some degree of inference. It won't tell you if the client used server-east or server-west, but if there are separate bits for changed-proto from changes-host, you can at least figure out if it was server-*.example.com or www.example.com. But just a single bit for Alt-Svc-Used doesn't even allow a server to distinguish that if both modes might be in-use. Incidentally, based on some other discussions that I've had, I've > basically concluded that the privacy leak isn't much alleviated by the > reduction in entropy on this field, so I'm open to re-opening that > discussion if you feel that would make this more straightforward. But > I think that I'd still be pushing for a reduction to zero, which would > remove what is a non-trivial disincentive to use Alt-Svc in some use > cases. > I do think that using something like the above with the alternate service that was ysed would be the most straight-forward. Erik
Received on Sunday, 24 August 2014 01:23:50 UTC