- From: Willy Tarreau <w@1wt.eu>
- Date: Sat, 10 Aug 2013 12:01:45 +0200
- To: Poul-Henning Kamp <phk@phk.freebsd.dk>
- Cc: "Roy T. Fielding" <fielding@gbiv.com>, IETF HTTP WG <ietf-http-wg@w3.org>
On Sat, Aug 10, 2013 at 09:37:53AM +0000, Poul-Henning Kamp wrote:
> In message <20130810092804.GO4079@1wt.eu>, Willy Tarreau writes:
>
> >I disagree with deprecating it, it's the only way to send non-idempotent
> >request over a reused connection.
>
> Not really.
>
> Send the headers, wait a RTT to see if you get an ACK, if not abandon.
>
> It's much faster than using Expect and at least as realiable.
No, for two reasons :
- this is not really doable at the application level (can only be
OS dependant, as the application manipulates connected sockets,
not packets).
- the fact that the TCP stack in front of you has ACKed your data
only means they were received in socket buffers, not that they
were delivered to the application. And when the keep-alive
timeout strikes and the server closes, there is no way to ensure
that pending data are always taken care of (since they can come
in between the moment you check and the close).
Willy
Received on Saturday, 10 August 2013 10:02:10 UTC