Series constructor with missing values and int dtype fails or passes depending on data type

[TomAugspurger]

I'm not sure what the expected behavior is, but these should match

In [16]: pd.Series([1, 2, np.nan], dtype=int)

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-16-f0075752a7f2> in <module>()
----> 1 pd.Series([1, 2, np.nan], dtype=int)

~/sandbox/pandas/pandas/core/series.py in __init__(self, data, index, dtype, name, copy, fastpath)
    280             else:
    281                 data = _sanitize_array(data, index, dtype, copy,
--> 282                                        raise_cast_failure=True)
    283
    284                 data = SingleBlockManager(data, index, fastpath=True)

~/sandbox/pandas/pandas/core/series.py in _sanitize_array(data, index, dtype, copy, raise_cast_failure)
   4154         if dtype is not None:
   4155             try:
-> 4156                 subarr = _try_cast(data, False)
   4157             except Exception:
   4158                 if raise_cast_failure:  # pragma: no cover

~/sandbox/pandas/pandas/core/series.py in _try_cast(arr, take_fast_path)
   4087             # that we can convert the data to the requested dtype.
   4088             if is_float_dtype(dtype) or is_integer_dtype(dtype):
-> 4089                 subarr = maybe_cast_to_integer_array(arr, dtype)
   4090
   4091             subarr = maybe_cast_to_datetime(arr, dtype)

~/sandbox/pandas/pandas/core/dtypes/cast.py in maybe_cast_to_integer_array(arr, dtype, copy)
   1341     try:
   1342         if not hasattr(arr, "astype"):
-> 1343             casted = np.array(arr, dtype=dtype, copy=copy)
   1344         else:
   1345             casted = arr.astype(dtype, copy=copy)

ValueError: cannot convert float NaN to integer

and when data is an array, we end up with float dtype.

In [17]: pd.Series(np.array([1, 2, np.nan]), dtype=np.int64)
Out[17]:
0    1.0
1    2.0
2    NaN
dtype: float64

[alanbato]

My initial guess would be that since we're explicitly requesting an int dtype we should get that as the type of the resulting Series, and maybe make a special case for npn.nan so it doesn't fail when it can't be casted?

However this is not any of those two behaviors, so if we want to go with one of those two I would go to falling back to what happens when data is an array.

[arw2019]

They match on 1.2 master:

In [19]: pd.Series([1, 2, np.nan], dtype=np.int)                                                                         
...
/workspaces/pandas-arw2019/pandas/core/dtypes/cast.py in maybe_cast_to_integer_array(arr, dtype, copy)
   1743     try:
   1744         if not hasattr(arr, "astype"):
-> 1745             casted = np.array(arr, dtype=dtype, copy=copy)
   1746         else:
   1747             casted = arr.astype(dtype, copy=copy)

ValueError: cannot convert float NaN to integer

In [20]: pd.Series([1, 2, np.nan], dtype=int)                                                                            
(...)
/workspaces/pandas-arw2019/pandas/core/dtypes/cast.py in maybe_cast_to_integer_array(arr, dtype, copy)
   1743     try:
   1744         if not hasattr(arr, "astype"):
-> 1745             casted = np.array(arr, dtype=dtype, copy=copy)
   1746         else:
   1747             casted = arr.astype(dtype, copy=copy)

ValueError: cannot convert float NaN to integer