Re: GOAWAY and proxies

Ages ago in spdy-land we discussed a LAME_DUCK frame, which would indicate
that a server would going away in a certain amount of time, but... never
seemed important enough to bother.
-=R


On Thu, Apr 10, 2014 at 2:56 PM, Daniel Sommermann <dcsommer@fb.com> wrote:

> In our implementation experience, GOAWAY as specified today is not an
> ideal mechanism for graceful shutdown. When there is high RTT between a
> proxy and the server, many requests may be in flight when the server
> generates a GOAWAY. For clients, it seems the solution is to reissue the
> request. The same strategy doesn't work for proxies. A proxy would have to
> buffer the headers and body of all requests until it sees the beginning of
> a response just in case a GOAWAY is received from the server. This is
> obviously not ideal as it puts high memory requirements on the proxy for a
> rare case.
>
> The server could implement some workarounds like "pre-acking" some number
> of streams in its GOAWAY just in case there are some in flight requests,
> but this doesn't seem like a great solution.
>
> Is there a way we can improve GOAWAY so that proxies can issue zero errors
> during a server restart and stay safe from a memory usage perspective? I'd
> be interested to hear others' implementation experience on this topic.
>
>