Series.sum has inconsistent return type

[remiremi]

Series.sum returns a numpy type, except when it's empty, in which case it returns a python int of value "0":

In [2]: type(pd.Series([0]).sum())
Out[2]: numpy.int64

In [3]: type(pd.Series().sum())
Out[3]: int

This poses a problem when I do 1 / myserie.sum() because I expect to obtain np.inf rather than a divison by 0 exception.

I think the return type of Series.sum() for empty series should be inferred from the Series's dtype This way, Series([], dtype='str').sum() would return an empty string, and Series([]).sum() would return np.float64(0) since an empty series' default type seems to be float64.

Tested with Pandas 0.16.0

[shoyer]

We don't support numeric methods on strings, but otherwise this seasons reasonable to me.

Want to give putting together a patch a try?

[jreback]

@remiremi yeah I don't think we coerce scalars to anything. This could be made more consistent. We should in general NOT be returning python scalars.

[remiremi]

@shoyer @jreback I created a pull request which adds a fix and a relevant test: #9829
Let me know if I'm missing anything

[jorisvandenbossche]

This now returns correctly a numpy type:

In [22]: type(pd.Series().sum())
Out[22]: float

So repurpose this issue to add a test to confirm this.

[jorisvandenbossche]

Closing this as we have an issue to convert the Series dtype to object for empty creation, and #19813 is there for the strange return value.