- From: Roberto Peon <grmocg@gmail.com>
- Date: Fri, 2 May 2014 14:22:19 -0700
- To: Martin Thomson <martin.thomson@gmail.com>
- Cc: Daniel Sommermann <dcsommer@fb.com>, William Chan (ιζΊζ) <willchan@chromium.org>, HTTP Working Group <ietf-http-wg@w3.org>
- Message-ID: <CAP+FsNcC++VqBOXjpL6c+t9A25ktmg97BH5=NZ_+cehFbnVTOw@mail.gmail.com>
I'm not entirely happy with the GOAWAY change, honestly. As written now, the receiver of GOAWAY has zero certainty about which streams were successfully processed as the server gets to change its mind and isn't limited to simply changing the number in one direction. -=R On Fri, May 2, 2014 at 11:31 AM, Martin Thomson <martin.thomson@gmail.com>wrote: > On 2 May 2014 10:42, Daniel Sommermann <dcsommer@fb.com> wrote: > > Now that Martin has clarified the possibility of receiving multiple > GOAWAYs > > in the text > > ( > https://github.com/http2/http2-spec/commit/d00b4f82001305a5d01e810a7b45140223cb2d08 > ), > > I have submitted https://github.com/http2/http2-spec/pull/475/files to > make > > the graceful shutdown mechanism explicit in the spec. I'd be happy to get > > hear feedback on this proposed change. > > > So that I can discuss this: > > + There is an inherent race condition when proxying from a > HTTP/1.1 client to a HTTP/2 > + server and the server goes down for maintenance. Since some > in-flight client requests may > + be unacknowledged by the server GOAWAY and there is no way > to retry these requests, > + servers SHOULD implement a two phase shutdown for > maintenance. In the first phase, the > + server sends a GOAWAY with NO_ERROR and a last stream id > equal to 2^31 - 1. After at > + least 1/2 RTT elapses, the server sends a second GOAWAY > with NO_ERROR and the last stream > + id it actually accepted. > > I think that this is a problem with intermediation in general. And > there may even be some clients that are unable to retry for other > reasons (intermediaries aren't the only entities that attempt to > minimize their state commitment). > > Half an RTT is definitely not enough. From the server perspective it > needs to be >1RTT. > > Maybe: > > A client that is unable to retry requests loses all requests that are > in flight when a server closes > down a connection. This is especially true for intermediaries that > might not be serving clients > using HTTP/2. If the server is attempting to gracefully shut down a > connection, it can send an > initial GOAWAY frame with the last frame identifier set to > 2<x:sup>31</x:sup>-1 and a > <x:ref>NO_ERROR</x:ref> code. This signals to the client that a > shutdown is imminent and that no > further requests can be initiated. The server can then wait at least > a round trip time to allow for > requests that the client sent prior to receiving the GOAWAY frame > before sending another GOAWAY > frame with an updated last stream identifier. This ensures that a > connection can be cleanly shut > down without losing requests. > >
Received on Friday, 2 May 2014 21:22:46 UTC