[e2e] tcp connection timeout

[Jim Gettys]

And being an application protocol designer who lived through the era
when Berkeley UNIX had keep alives on by default, all they did was cause
my applications to fail during transient failures, causing many more
failures than would otherwise occur (most of my applications are idle
most of the time).  Keep alives should be renamed "Keep killing".

>From my point of view, good riddance to keep-alives... I always
considered it one of Berkeley's poorest ideas.  Jon was right: it really
is an application problem, and not something I want the system to second
guess me on.
				- Jim Gettys


On Wed, 2006-03-01 at 12:29 -0800, Bob Braden wrote:
>  
>   *> 
>   *> Hi,
>   *> 
>   *> I recently got a TCP connection timeout error.
>   *> 
>   *> If both side don't send packets for a long time while connection is open, is 
>   *> there a default time out to abort the connection?
>   *> 
>   *> Thanks a lot!
>   *> 
>   *> - Kun 
>   *> 
> 
> Sigh.  This is periodic question #13, which recurs about every 2.3
> years.  Those who have been around the Internet more than 2.3 years can
> stop reading now.
> 
> The inventors of TCP, and in particular Jon Postel, were religiously
> opposed to a TCP idle timer ("keep alive").  The knowledge of when to
> time out an idle connection belongs logically to the application layer;
> TCP itself does not know.  But the OS guys (beginning with the Berkeley
> Unix folks) disagreed on pragmatic grounds, so BSD TCP had a
> keep-alive.
> 
> When the Host Requirements Working Group came to this issue, there was
> nearly unanimous agreement with Postel, that TCP should not have a keep
> alive.  The dissenter was the programmer responsible for the BSD code,
> which probably powered the majority of hosts at the time!  So, the Host
> Requirements group set requirements on Keep-Alives that were
> designed to come as close to outlawing them as possible without
> actually outlawing them.  Thus, keep-alives must be optional, they must
> default to "off", the default timeout interval must be very large,
> etc.
> 
> Bob Braden
>