Actually, I take it back. The default state of logging appears to do the right thing with no special handler installed - both the .critical() and .exception() messages are written out to stderr by default, so the futures tests pass even after Brian's patch is applied.

So I think removing the handler installation code is the right thing to do, even if (as it turns out) it doesn't fix the test failure.

I also thought of a more minimal way to reproduce the test failure that makes it clear pydoc isn't really involved:

  ./python -m test test_concurrent_futures test_logging test_concurrent_futures

The first execution of the test will pass, the second will fail (both with and without Brian's patch to remove the handler installation).

Adding Vinay to the nosy list - I suspect Antoine is right that the logging tests are leaving existing loggers in a slightly unhealthy state. A better save/restore in regrtest.py might be a place to start, but I don't know the internals of the logging package well enough to improve on what I already added.