BUG: Columns used in sum on 1 row dataframe depend on values instead of dtype

[jcrist]

It seems that somehow the columns used in sum when applied to a 1 row dataframe depend on the values in the row instead of just the dtypes. Observe:

import pandas as pd
import numpy as np

# Frame with some non-numeric dtypes
df = pd.DataFrame({'a': [1], 'b': [1.1], 'c': ['foo'], 'd': [pd.Timestamp('2000-01-01')]})
# Only change here is that `d` is `NaT`
df2 = pd.DataFrame({'a': [1], 'b': [1.1], 'c': ['foo'], 'd': [pd.NaT]})
# This is just the first one twice
df3 = pd.concat([df, df])

# I'd expect all 3 to use the same columns in the reduction
df_sum = df.sum()
df2_sum = df2.sum()
df3_sum = df3.sum()

Loading that in an ipython session:

In [1]: df_sum
Out[1]:
a                      1
b                    1.1
c                    foo
d    2000-01-01 00:00:00
dtype: object

In [2]: df2_sum
Out[2]:
a    1.0
b    1.1
dtype: float64

In [3]: df3_sum
Out[3]:
a    2.0
b    2.2
dtype: float64

In [4]: pd.__version__
Out[4]: u'0.18.1'

In [5]: np.__version__
Out[5]: '1.11.1'

I'd expect all 3 to only use the columns ['a', 'b'], as these are the only numeric columns. Strangely, _get_numeric_data does return just ['a', 'b'] in all cases, so it's not that.

[jreback]

xref #13416

[jreback]

This is a bug here: https://github.com/pydata/pandas/blob/master/pandas/core/nanops.py#L637

need to handle object correctly rather than trying to coerce to float (which fails as there are strings embeded); this path doesn't happen for the 1st one (and succeeds), while the 2nd raises a TypeError.

So we need to raise in the first case on a mixed type on axis=0.