[BUG] Can not assign a datetime index into a dataframe

[mrocklin]

Things seem to be unhappy when we assign a datetime index into a dataframe.

This comes up when I try to do indexed joins

import pandas as pd
df = pd.date_range('2000', '2001', freq='1M').to_frame()

import cudf
gdf = cudf.from_pandas(df)
gdf['foo'] = gdf.index

---------------------------------------------------------------------------
GDFError                                  Traceback (most recent call last)
<ipython-input-6-d57fa9fc0872> in <module>
----> 1 gdf['foo'] = gdf.index

~/cudf/python/cudf/dataframe/dataframe.py in __setitem__(self, name, col)
    268             self._cols[name] = self._prepare_series_for_add(col)
    269         else:
--> 270             self.add_column(name, col)
    271
    272     def __delitem__(self, name):

~/cudf/python/cudf/dataframe/dataframe.py in add_column(self, name, data, forceindex)
    794         if isinstance(data, GeneratorType):
    795             data = Series(data)
--> 796         series = self._prepare_series_for_add(data, forceindex=forceindex)
    797         series.name = name
    798         self._cols[name] = series

~/cudf/python/cudf/dataframe/dataframe.py in _prepare_series_for_add(self, col, forceindex)
    770         empty_index = len(self._index) == 0
    771         series = Series(col)
--> 772         if forceindex or empty_index or self._index.equals(series.index):
    773             if empty_index:
    774                 self._index = series.index

~/cudf/python/cudf/dataframe/index.py in equals(self, other)
    184         if len(self) != len(other):
    185             return False
--> 186         return (self == other)._values.all()
    187
    188     def join(self, other, method, how='left', return_indexers=False):

~/cudf/python/cudf/dataframe/index.py in __eq__(self, other)
    164
    165     def __eq__(self, other):
--> 166         return self._apply_op('__eq__', other)
    167
    168     def __ne__(self, other):

~/cudf/python/cudf/dataframe/index.py in _apply_op(self, fn, other)
    124         op = getattr(idx_series, fn)
    125         if other is not None:
--> 126             return as_index(op(other))
    127         else:
    128             return as_index(op())

~/cudf/python/cudf/dataframe/series.py in __eq__(self, other)
    459
    460     def __eq__(self, other):
--> 461         return self._unordered_compare(other, 'eq')
    462
    463     def __ne__(self, other):

~/cudf/python/cudf/dataframe/series.py in _unordered_compare(self, other, cmpops)
    443         nvtx_range_push("CUDF_UNORDERED_COMP", "orange")
    444         other = self._normalize_binop_value(other)
--> 445         outcol = self._column.unordered_compare(cmpops, other._column)
    446         result = self._copy_construct(data=outcol)
    447         result.name = None

~/cudf/python/cudf/dataframe/datetime.py in unordered_compare(self, cmpop, rhs)
    148             lhs, rhs,
    149             op=_unordered_impl[cmpop],
--> 150             out_dtype=np.bool
    151         )
    152

~/cudf/python/cudf/dataframe/datetime.py in binop(lhs, rhs, op, out_dtype)
    214     masked = lhs.has_null_mask or rhs.has_null_mask
    215     out = columnops.column_empty_like(lhs, dtype=out_dtype, masked=masked)
--> 216     null_count = _gdf.apply_binaryop(op, lhs, rhs, out)
    217     out = out.replace(null_count=null_count)
    218     nvtx_range_pop()

~/cudf/python/cudf/_gdf.py in apply_binaryop(binop, lhs, rhs, out)
    100     args = (lhs.cffi_view, rhs.cffi_view, out.cffi_view)
    101     # apply binary operator
--> 102     binop(*args)
    103     # validity mask
    104     if out.has_null_mask:

~/miniconda/envs/cudf/lib/python3.7/site-packages/libgdf_cffi/wrapper.py in wrap(*args)
     25                     if errcode != self._api.GDF_SUCCESS:
     26                         errname, msg = self._get_error_msg(errcode)
---> 27                         raise GDFError(errname, msg)
     28
     29                 wrap.__name__ = fn.__name__

GDFError: GDF_UNSUPPORTED_DTYPE

[kkraus14]

I believe this should be isolated to DatetimeIndex types and I'm hesitant to try to get this working for 0.6. This may be fixed by #892, but if not we should push to 0.7.

[thomcom]

I'm pretty sure this is fixed. @mrocklin's reproducer now works.

>>> import pandas as pd
... df = pd.date_range('2000', '2001', freq='1M').to_frame()
...
... import cudf
... gdf = cudf.from_pandas(df)
... gdf['foo'] = gdf.index

>>> print(gdf['foo'])
2000-01-31T00:00:00.000   2000-01-31T00:00:00.000
2000-02-29T00:00:00.000   2000-02-29T00:00:00.000
2000-03-31T00:00:00.000   2000-03-31T00:00:00.000
2000-04-30T00:00:00.000   2000-04-30T00:00:00.000
2000-05-31T00:00:00.000   2000-05-31T00:00:00.000
2000-06-30T00:00:00.000   2000-06-30T00:00:00.000
2000-07-31T00:00:00.000   2000-07-31T00:00:00.000
2000-08-31T00:00:00.000   2000-08-31T00:00:00.000
2000-09-30T00:00:00.000   2000-09-30T00:00:00.000
2000-10-31T00:00:00.000   2000-10-31T00:00:00.000
[2 more rows]
Name: foo, dtype: datetime64[ms]