[Python-Dev] Long-time shy failure in test_socket_ssl

[Tim Peters]

[Tim]
>> Neal plugged another hole later, but-- alas --I have seen the same shy
>> failure since then on WinXP.  One of the most recent buildbot test
>> runs saw it too, on a non-Windows box:
>>
>> http://www.python.org/dev/buildbot/trunk/g5%20osx.3%20trunk/builds/204/step-test/0
>>
>> test_socket_ssl
>> test test_socket_ssl crashed --
>>     exceptions.TypeError: 'NoneType' object is not callable
>>
>> in the second test run there.

[Neal]
> For closure, I believe this problem was addressed by revs 42842 and
> 42844 to Lib/test/test_importhooks.py.

I agree the mystery-change there made the problem go away (yay!), but
would prefer to reserve "addressed" for some notion of "understood".

Who understands test_importhooks?  Without Neal's change to delete
"urllib" from sys.modules there, and when running tests in a specific
order, test_socket_ssl dies because the module-global `urlparse`
magically becomes None on line 147 of urlparse.py's urljoin()
function.  What sins does test_importhooks commit to provoke this?
What else does running test_importhooks distort?  That is, what else
should test_importhooks be doing to clean up after itself that it's
not currently doing (in the absence of understanding the original
problem, there's no apparent answer)?

> If anyone sees spurious failures with the buildbot (one time failures,
> crashes, etc), please report the problems to python-dev.  It would be
> great to see if you can reproduce the results with the same tests that
> failed.  We need to determine if it is architecture specific,
> test-order related, or something else.

Yes, it's loads of fun :-)