Re: HTTP -> Messages -> Transport factoring

On Thu, Apr 5, 2012 at 12:24 PM, Poul-Henning Kamp <phk@phk.freebsd.dk>wrote:
>
> On 05/04/2012, at 11:55 AM, Mike Belshe wrote:
>
> >> In other words, the least-common-denominator sucks....
>
> You don't have to resort to l-c-d in order to make your protocol
> non-hostile to other transports.
>
> In many ways TCP, certainly compared to many other contenders at
> the time of distinction, _is_ l-c-d: Just a byte stream.
>
> I think we can take it as a given that HTTP over UDP will exist pretty
> soon: Between a surrogate like Varnish or HA-Proxy and the http-server
> in the same rack, it would cut a fair bit time of transactions and
> a lot of objects fit perfectly well inside a 64Kb UDP packet.
>
I would propose that HTTP only concern itself with reliable transports. A
reliable transport provides both in-order delivery and guaranteed delivery.
 UDP does not provide guaranteed delivery (even if you use the 64K hack you
mention to avoid the in-order issues), and therefore is not a transport
worthy of discussion for HTTP/2.0 by itself.
Mike
> >> Mark Nottingham writes:
>
> >Right. I can see accommodating the potential for these things by putting
> >a bit of thought into how our specs are factored, but am *not*
> >suggesting that we require an existence proof, or spend significant time
> >assuring that they're possible.
>
> seconded.
>
>
> --
> Poul-Henning Kamp | UNIX since Zilog Zeus 3.20
> phk@FreeBSD.ORG | TCP/IP since RFC 956
> FreeBSD committer | BSD since 4.3-tahoe
> Never attribute to malice what can adequately be explained by incompetence.
>