BUG: invalid coercion raises in groupby uniques

[jreback]

df = pd.DataFrame([[1, 2], [3, 4], [5, 6]], dtype=object) + 9223372036854775807
df.groupby(0).sum()

...
C:\Anaconda\envs\bmc3\lib\site-packages\pandas\core\groupby.py in labels(self)
   2252     def labels(self):
   2253         if self._labels is None:
-> 2254             self._make_labels()
   2255         return self._labels
   2256

C:\Anaconda\envs\bmc3\lib\site-packages\pandas\core\groupby.py in _make_labels(self)
   2264         if self._labels is None or self._group_index is None:
   2265             labels, uniques = algos.factorize(self.grouper, sort=self.sort)
-> 2266             uniques = Index(uniques, name=self.name)
   2267             self._labels = labels
   2268             self._group_index = uniques

C:\Anaconda\envs\bmc3\lib\site-packages\pandas\indexes\base.py in __new__(cls, data, dtype, copy, name, fastpath, tupleize_cols, **kwargs)
    234                 if inferred == 'integer':
    235                     from .numeric import Int64Index
--> 236                     return Int64Index(subarr.astype('i8'), copy=copy,
    237                                       name=name)
    238                 elif inferred in ['floating', 'mixed-integer-float']:

OverflowError: int too big to convert

We are inferring int when the this should be 'object' (or just not integer):

In [36]: pd.lib.infer_dtype(np.array([9223372036854775808, 9223372036854775810, 9223372036854775812], dtype=object))
Out[36]: 'integer'

[jreback]

this works on master if someone wants to do a test.

[jreback]

cc @gfyoung

[gfyoung]

We are inferring int when the this should be 'object' (or just not integer):

Why should we inferring object? It's perfectly acceptable as uint64- this is yet another example of how we don't have great support for uint64.

[gfyoung]

I believe the correct return dtype should be uint64, as illustrated by the fact that we get an int when adding a smaller number:

>>> df = pd.DataFrame([[1, 2], [3, 4], [5, 6]], dtype=object) + 9
>>> df.groupby(0).sum().dtypes
1    int64
dtype: object

[jreback]

this was before uint64 support that's why it works now

just need a test to verify

i bet this will break though with a number out of range for uint64 though (but maybe separate issue)

[gfyoung]

this was before uint64 support that's why it works now

Ah, yes, confirmed. I'll add this as a test.

i bet this will break though with a number out of range for uint64 though (but maybe separate issue)

That's probably a separate issue. Handling numbers beyond what we can capably hold is going to very difficult in any case.