Surprising deepcopy semantics with dtype='object'

[darikg]

from copy import deepcopy
import numpy as np
import xarray as xr

class Dummy: pass

a0 = np.array([Dummy()])
a1 = deepcopy(a0)
print(a0[0] is a1[0])  # False, as expected

x0 = xr.DataArray(a0, dims='dummy')
x1 = deepcopy(x0)
print(x0.values[0] is x1.values[0])  # unexpectedly True

I think this is a bug, and would be fixed with an extra deepcopy around here

[keewis]

the issue is that we use numpy.ndarray.copy to copy the data:

xarray/xarray/core/variable.py

Line 939 in 43a2a4b

data = data.copy()

(numpy.ndarray defines __array_function__).

You can reproduce the issue using just numpy:

In [14]: a0.copy()[0] is a0[0]                                                                                  
Out[14]: True

We might want to use deepcopy instead because duck arrays are not required to implement .copy.

[shoyer]

I agree, marking this as a bug. We should use deepcopy() recursively on NumPy arrays.

[mathause]

@darikg are you interested to send in a PR?

[darikg]

Sure! I'll be back in a couple days when I get a chance