Many of these exceptions are due to resource unavailable, so the
existing code should be able to handle many more spurious errors.

It's system dependent.  This might fix the problem with test_rfind failing.

nor UNIX is defined.  This may have caused problems on cygwin.

Raymond's advice.

which isn't as bad as it sounds.  The close() *should* raise an exception,
so if it didn't we should give more time to sync and really raise it.

Will backport.

This should make the test less flaky.  It also speeds up the test
by about 75% on my box (20+ seconds -> ~4 seconds).

doesn't cause the exception, but the close() does.

Will backport.

and not trying to connect before.

this might fix the problem.

the original code object backing the generator. Patch by Collin Winter.

as the documentation says.

Saves the unnecessary creation of a tuple as intermediate container.

Make PySet_Add() work with frozensets.  Works like PyTuple_SetItem() to build-up values in a brand new frozenset.

that shouldn't cause the test to fail.  Just like it setUp.

global to determine when the server is ready to be used.  This slows the test
down, but should make it correct.  There was a race condition before where the
server could have assigned a port, yet it wasn't ready to serve requests.  If
the client sent a request before the server was completely ready, it would get
an exception.  There was machinery to try to handle this condition.  All of
that should be unnecessary and removed if this change works.  A NOTE was
added as a comment about what needs to be fixed.

The buildbots will tell us if there are more errors or
if this test is now stable.