Re: GOAWAY and proxies

On 04/10/2014 03:09 PM, Martin Thomson wrote:
> Can't the proxy just issue RST_STREAM toward its clients for all the
> affected requests if it can't tolerate the memory hit? REFUSED_STREAM
> seems like the right code there.
>
> RST_STREAM will cause an end-to-end retry, if anything, which is
> slower.  But retrying requests is a feature that the proxy optionally
> provides, and if it can't tolerate the costs of providing the feature,
> then maybe it can do something to avoid taking on that cost.
>
I was not aware that REFUSED_STREAM caused clients to retry. That seems 
to generally solve this issue! I am worried that some implementations 
may *not* reissue the request (the language only says the client MAY 
retry), leading to a request that could have succeeded not being serviced.

Should the language read that the client SHOULD retry, subject to some 
rate-limiting?