Re: HTTP/2 GREASE, Results, and Implications

[Willy Tarreau <[e-mail]>]

On Thu, Oct 31, 2019 at 06:10:33PM +0000, Lucas Pardue wrote:
> On Thu, Oct 31, 2019 at 5:49 PM Mike Bishop <[e-mail]> wrote:
> 
> > Upon further investigation, it appears that Akamai hosts most of the sites
> > where the connection gets closed, and Cloudflare hosts the sites which
> > close the stream but leave the connection open (I'm guessing, from what the
> > dev tools show).  Given that we're both represented here, it seems likely
> > that we should be able to drive some internal bugfixes and improve the
> > state of the ecosystem.
> >
> 
> It's possible the Cloudflare closures are related to the client-oriented
> half-closed issue that Willy and Bence have discussed.

Maybe, but regardless, if we have to accept any frame type, half-closed
makes no difference. For haproxy I still purposely decided not to accept
to create a new stream from a unknown frame type (i.e. a frame type
designating an idle stream must be either HEADERS or PRIORITY as stated
in the spec). Because if we start to let such things pass through, we
can't really count on the other end's behavior for subsequent frames :-/

> We'd be happy to
> coordinate with other members of the group on a rerun of the experiment in
> order to come to a more definitive answer.

It would be nice to retest at world scale and be a bit more prepared to
assist users who report breakage.

Willy