min and max methods changing dtype to float64

[ifariasg]

I'm working with datasets that have coordinates with lots of decimal places. As I don't need this I reassign cords and transform the dtype to float32.
For calculating the extent of the dataset a standard procedure is to use the min() and max() methods but using them over the float32 coordinates yields the float64 values which leads to misaligned coordinates.

Minimum reproducible example:

import xarray as xr
import numpy as np
 
da = xr.DataArray(
    data=np.array([1.2, 2.2, 3.3], dtype=np.float32),
    coords={"x": [1.345, 2.345, 3.345]},
    dims=("x",),
)
 
print(da[0].dtype)  # dtype('float32')
print(da.min().dtype)  # dtype('float64')
print(type(np.min(da.values)))  # dtype('float32')

Is there a reason for this?

[keewis]

I'm not sure about this, but I think the reason is that we're using bottleneck to speed up the computation, which returns a python float object for scalars. To be able to wrap it, xarray will convert it to a 0d array, which by default has a dtype of float64.

np.{min,max} returns np.float32 instances, so the dtype information is not lost with that.

To work around that, you can disable bottleneck:

xr.set_options(use_bottleneck=False)