_reduce promotes numerics to strings when axis=0 and mixed type.

[dalejung]

import pandas as pd
import numpy as np

df = pd.DataFrame({'num':[1,2,3], 'num2': np.random.randn(3), 'strings':list('abc')})
df.sum().values # array(['6', '-1.47818918058', 'abc'], dtype=object)

The issue is that axis=0 tries to coerce the dtypes and np.array will promote the numerics to strings.

Note that:

df.T.sum(axis=1).values # array([6, -1.4781891805780178, 'abc'], dtype=object)

Maybe com._coerce_to_dtypes should just return the list?

[jreback]

I think could just make that returned array object dtype

hate doing that though so maybe just returning a list is better as I think series contruction will handle better

[dalejung]

So thinking about this more.

In [24]: df = pd.DataFrame({
   ....:     'string_data': ['a', 'b', 'c', 'd', 'e'],
   ....:     'bool_data': [True, True, False, False, False],
   ....:     'int_data': [10, 20, 30, 40, 50],
   ....: })

In [25]: df.sum()
Out[25]:
bool_data       True
int_data         150
string_data    abcde
dtype: object

In [26]: df.T.sum(axis=1)
Out[26]:
bool_data          2
int_data         150
string_data    abcde
dtype: object

In [27]: np.sum([True,True])
Out[27]: 2

Note that the bool_data is converted back to its original dtype for axis=0. Should df.op() be equivalent to df.T.op(axis=1)?

[jreback]

I think the coercion for bool data should only convert to bool if it's 1/0 else leave it as int
(that's why it's True)

it is evaluating fine -

it then tries to coerce to float64 ( kind of a legacy thing) if it fails then it will item by item coerce (which is relatively new - I added because timedelta arithmetic was not coercing properly)