CI: numpydev failure

[TomAugspurger]

https://travis-ci.org/pandas-dev/pandas/jobs/436434493#L2766

____________ TestSelection.test_groupby_duplicated_column_errormsg _____________
[gw1] linux -- Python 3.7.0 /home/travis/miniconda3/envs/pandas/bin/python
self = <pandas.tests.groupby.test_grouping.TestSelection object at 0x7fb0ebdafeb8>
    def test_groupby_duplicated_column_errormsg(self):
        # GH7511
        df = DataFrame(columns=['A', 'B', 'A', 'C'],
                       data=[range(4), range(2, 6), range(0, 8, 2)])
    
>       pytest.raises(ValueError, df.groupby, 'A')
pandas/tests/groupby/test_grouping.py:42: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
pandas/core/generic.py:7152: in groupby
    observed=observed, **kwargs)
pandas/core/groupby/groupby.py:1966: in groupby
    return klass(obj, by, **kwds)
pandas/core/groupby/groupby.py:365: in __init__
    mutated=self.mutated)
pandas/core/groupby/grouper.py:588: in _get_grouper
    if is_categorical_dtype(gpr) and len(gpr) != obj.shape[axis]:
pandas/core/dtypes/common.py:574: in is_categorical_dtype
    return CategoricalDtype.is_dtype(arr_or_dtype)
pandas/core/dtypes/base.py:93: in is_dtype
    return cls.construct_from_string(dtype) is not None
pandas/core/dtypes/dtypes.py:359: in construct_from_string
    if string == 'category':
pandas/core/ops.py:1923: in f
    try_cast=False)
pandas/core/frame.py:4930: in _combine_const
    return ops.dispatch_to_series(self, other, func)
pandas/core/ops.py:1714: in dispatch_to_series
    new_data = expressions.evaluate(column_op, str_rep, left, right)
pandas/core/computation/expressions.py:205: in evaluate
    return _evaluate(op, op_str, a, b, **eval_kwargs)
pandas/core/computation/expressions.py:65: in _evaluate_standard
    return op(a, b)
pandas/core/ops.py:1694: in column_op
    for i in range(len(a.columns))}
pandas/core/ops.py:1694: in <dictcomp>
    for i in range(len(a.columns))}
pandas/core/ops.py:1546: in wrapper
    res = na_op(values, other)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
x = array([0, 2, 0]), y = 'category'
    def na_op(x, y):
        # TODO:
        # should have guarantess on what x, y can be type-wise
        # Extension Dtypes are not called here
    
        # Checking that cases that were once handled here are no longer
        # reachable.
        assert not (is_categorical_dtype(y) and not is_scalar(y))
    
        if is_object_dtype(x.dtype):
            result = _comp_method_OBJECT_ARRAY(op, x, y)
    
        elif is_datetimelike_v_numeric(x, y):
            return invalid_comparison(x, y, op)
    
        else:
    
            # we want to compare like types
            # we only want to convert to integer like if
            # we are not NotImplemented, otherwise
            # we would allow datetime64 (but viewed as i8) against
            # integer comparisons
    
            # we have a datetime/timedelta and may need to convert
            assert not needs_i8_conversion(x)
            mask = None
            if not is_scalar(y) and needs_i8_conversion(y):
                mask = isna(x) | isna(y)
                y = y.view('i8')
                x = x.view('i8')
    
            method = getattr(x, op_name, None)
            if method is not None:
                with np.errstate(all='ignore'):
>                   result = method(y)
E                   FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison
pandas/core/ops.py:1429: FutureWarning

I haven't been able to reproduce locally yet.

[TomAugspurger]

I can reproduce locally now. Not sure what was going on earlier.

#22901 was the commit that introduced the failure, but we didn't catch it there since numpydev is in the "allowed failure" section. I think we should move it up to the non-allowed section.

I think the root cause is calling CategoricalDtype.construct_from_string(dtype) on a DataFrame. We used to do that in a bare except:, but now it's except TypeError. Ideally we wouldn't be calling that on data in the first place, since the results are unpredictable.

ipdb> pp string
   A  A
0  0  2
1  2  4
2  0  4

We do

ipdb> string == 'category'
*** FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison

That FutureWarning is elevated to an error. I didn't realize this, but you can catch warnings, so the very simple fix is to also catch FutureWarning here, but that feels fragile. I'll investigate a better fix...