SocketServer.ForkingMixIn.collect_children() does:

       pid, status = os.waitpid(0, options)

By setting pid = 0, it will wait on any children, even
if the children were not started by the SocketServer. 
This causes sporadic test failures:

 
http://mail.python.org/pipermail/python-dev/2006-August/068216.html

It later assumes that if any child finished, that this
server owned the process, which is not a valid
assumption.  Any part of the code could have spawned
the child.

I will fix the test suite, so this problem shouldn't
occur (well at least reduce the likelihood by reaping
the children before each test).

I'm not certain how this problem should really be fixed
(or if it's even worth fixing).  Two possible solutions
are: 
 * only wait on the children we own
 * ignore returned pids that we don't own

Logged In: YES 
user_id=33168

I forgot to mention that this problem seems to have
sporadically hit many of the buildbots.  Though it seemed to
always occur on cygwin.

Logged In: YES 
user_id=6380

It should keep track of the pids of all its children and
only wait for those.  Shouldn't be hard.  Can be done in
2.5.1 since it's a bugfix.

Mostly fixed in r61106. This should make the buildbots happy, but there
is still a corner case in which we waitpid(0) and could confuse other
libraries.

Jeffrey, this very commit broke it. See also issue 5814. The same
happened for me last week. What makes it especially bad is that even
once enough children terminate, the SocketServer won't be able to handle
any more connections; it just throws this Exception again and again and
closes new connections immediately.

Yes, sorry. That was fixed in r69927.