DataArray.sum does not respect dtype keyword

[gerritholl]

Code Sample, a copy-pastable example if possible

# Your code here
da = xarray.DataArray(arange(5, dtype="i2"))
print(da.sum(dtype="i4").dtype)

Problem description

The result is int64. This is a problem because I asked for int32.

Expected Output

Expected output int32.

Output of xr.show_versions()

# Paste the output here xr.show_versions() here INSTALLED VERSIONS ------------------ commit: None python: 3.6.1.final.0 python-bits: 64 OS: Linux OS-release: 2.6.32-696.6.3.el6.x86_64 machine: x86_64 processor: x86_64 byteorder: little LC_ALL: None LANG: en_GB.UTF-8 LOCALE: en_GB.UTF-8

xarray: 0.10.0+dev12.gf882a58
pandas: 0.22.0
numpy: 1.14.0
scipy: 1.0.0
netCDF4: 1.3.1
h5netcdf: None
Nio: None
bottleneck: 1.2.1
cyordereddict: None
dask: 0.16.1
matplotlib: 2.1.1
cartopy: 0.15.1
seaborn: 0.8.1
setuptools: 38.4.0
pip: 9.0.1
conda: 4.3.16
pytest: 3.1.2
IPython: 6.1.0
sphinx: 1.6.2

[fujiisoup]

I notice here

xarray/xarray/core/duck_array_ops.py

Lines 177 to 181 in f3deb2f

	def f(values, axis=None, skipna=None, **kwargs):
	# ignore keyword args inserted by np.mean and other numpy aggregators
	# automatically:
	kwargs.pop('dtype', None)
	kwargs.pop('out', None)

we removed dtype arguments from reduce methods, but I'm not sure why.
This might be a bug.

As a temporal solution,

da.reduce(np.sum, dtype="i4")

preserves dtype.

[shoyer]

We do this for two reasons:

• bottleneck's aggregation functions like bottleneck.nansum() don't have a dtype argument, so passing on a dtype argument causes an error to be raised.
• If you call a numpy function like numpy.sum() on an xarray object, it calls the appropriate method with all keyword arguments, e.g., numpy.sum(xarray_obj) -> xarray_obj.sum(axis=None, dtype=None, out=None).

What we should probably do here instead of ignoring dtype and out entirely is to look at their values:

• If dtype is not None, use numpy's aggregation function instead of bottleneck's.
• If out is not None, raise an error.