Re: Fragmentation for headers: why jumbo != continuation.


On Sat, Jul 12, 2014 at 07:59:59AM +0000, Poul-Henning Kamp wrote:
> I know of hackish loadbalancers which operate at IP packet level, and
> which expects that the information they require for load-balancing fit
> inside a single ethernet frame, but given that no HTTP standard ever
> talked about ethernet frames, I think we can agree that they are not
> anywhere near being compliant.

[ One thing I like with HTTP/2 is that the crappy devices above will
  be forced to disappear and suddenly the web will start to work much
  better ]

> So please tell us more about these mythological "header-streaming"
> proxies you keep talking about, nobody else seems to have ever seen
> or heard about them.

Roberto already explained his specific use, and I understood that the
LB didn't need to inspect the response since it trusted the servers, and
that he wanted to stream it from the servers to improve latency. That was
the reason why we diverged on the subject of latency because while I full
respect this usage model (HTTP is for everyone). In this specific case, I
expected the extra latency ought to be extremely small (let even measurable)
thanks to the high speed links between the LB and the servers, and probably
not worth maintaining the extra complexity in the protocol.


Received on Saturday, 12 July 2014 08:23:46 UTC