Here's a fix for test_robotparser.py.  With this patch, the command 'python -m unittest discover Lib/test/ "test_*.py"' can actually be run--before the patch, test_robotparser's unique TestCase subclass causes unexpected errors for discovery.

I got to looking over my patch here again, and thought of another possible fix.  I'm not sure which is uglier, though :)

The problem is that most of the test cases are created using a custom subclass of unittest.TestCase which takes extra constructor arguments.  Discovery doesn't expect a TestCase subclass to want extra arguments, so it fails noisily trying to 'discover' RobotTestCase.

Patch version 2 just makes a couple small adjustments to version 1, which replaces the custom subclass with a base class to provide runTest, and a factory function to create the TestCase instances.

Version 3 is a different approach, which provides default arguments to the extra constructor parameters, and bails out of __init__ early if the index argument is not an int (checking for None doesn't work because at some point, discovery calls 'RobotTestCase("runTest")', which makes 'index' "runTest").

Both approaches look kind of ugly to me, but both get the job done.

Do you think doing something like:

class BaseRobotTest:
    def setUp(self):
        lines = io.StringIO(robots_txt).readlines()
        self.parser = urllib.robotparser.RobotFileParser()
        parser.parse(lines)

    def test_good(self):
         for url in good:
             self.assertTrue(self.parser.can_fetch(...))

    def test_bad(self):
         for url in bad:
             self.assertFalse(self.parser.can_fetch(...))

class RobotTestX(BaseRobotTest, unittest.TestCase):
    doc = "..."
    good = [...]
    bad = [...]

...

would be a better approach?

On one hand is a bit more verbose and doesn't create a separate test for each URL (I don't think that's important though), but on the other hand it gets rid of lot of magic and makes the test more understandable.

I agree that, long-term, test_robotparser should probably be rewritten with less magic.  Short-term, though, I think it would be good to get a temporary fix for discovery in place since this is the only test file that completely blows up discovery before it even starts testing, and open a new issue for the rewrite.  The v3 patch does the job with the least disruption in the file, though it does add a bit more magic.

New changeset fa051c6276a0 by Ezio Melotti in branch '3.3':
#17066: test_robotparser now works with unittest test discovery.  Patch by Zachary Ware.
http://hg.python.org/cpython/rev/fa051c6276a0

New changeset 8fdce849d0b3 by Ezio Melotti in branch 'default':
#17066: merge with 3.3.
http://hg.python.org/cpython/rev/8fdce849d0b3

Fixed, thanks for the patch!