BUG: Exception raised for a duplicate MultiIndex level name

[seth-p]

The following two tests seem inconsistent. Why does the first test for a KeyError while the second tests for a ValueError? Also, the first error message, Level foo not found, doesn't seem correct.

test_index.py:

    def test_duplicate_names(self):
        self.index.names = ['foo', 'foo']
        assertRaisesRegexp(KeyError, 'Level foo not found',
                           self.index._get_level_number, 'foo')

test_frame.py:

    def test_unstack_non_unique_index_names(self):
        idx = MultiIndex.from_tuples([('a', 'b'), ('c', 'd')],
                                     names=['c1', 'c1'])
        df = DataFrame([1, 2], index=idx)
        with tm.assertRaises(ValueError):
            df.unstack('c1')

        with tm.assertRaises(ValueError):
            df.T.stack('c1')

[shoyer]

I agree -- that first error looks very strange. It's particularly strange that the test is for a private method. We probably should be checking for duplicate names when index.names is set, and should be raising ValueError.

[jreback]

private methods should be explicity tested just as public api methods.

These tests are for 2 different things. The first because of an invalid level name KeyError makes senses. The 2nd for an invalid reshape operation (see below).

The issue here is this. a MultiIndex technically can have duplicate names, but in point of fact they are pretty much useless and tend to break / degrade other ops. E.g. stack/unstack, and make your columns duplicate when reset. So I think we should actually raise on construction to solve this entire problem.

So names can be:

• None (where they get defaulted level numbers)
• unique integers / names
• None / mixed integers/names (usually arise from non-named reduction ops).

[seth-p]

Sounds good to me!

[mroeschke]

Seems like there hasn't been much interest in this feature over the years so closing