Describe the issue:

I have an array with dtype datetime64. The dtype.metadata is None. After serializing the array with pickle dump/load, the dtype.metadata is some object "mappingproxy({})".

This is a problem because trying to np.save the array with dtype.metadata produces a UserWarning.

dtype.metadata appears to be some kind of undocumented feature(???) I've never heard of before now.

Reproduce the code example:

import numpy as np
import pickle
import io

a = np.asarray([np.datetime64('2000-01-01')])
print(f'{a = }')
print(f'{a.dtype = }')
print(f'{a.dtype.metadata = }')

b = pickle.loads(pickle.dumps(a))
print(f'{b = }')
print(f'{b.dtype = }')
print(f'{b.dtype.metadata = }')

print('Trying to save a')
f = io.BytesIO()
np.save(f, a)

print('Trying to save b')
f = io.BytesIO()
np.save(f, b)

Error message:

a = array(['2000-01-01'], dtype='datetime64[D]')
a.dtype = dtype('<M8[D]')
a.dtype.metadata = None
b = array(['2000-01-01'], dtype='datetime64[D]')
b.dtype = dtype('<M8[D]')
b.dtype.metadata = mappingproxy({})
Trying to save a
Trying to save b
<...>\site-packages\numpy\lib\_format_impl.py:393: UserWarning: metadata on a dtype is not saved to an npy/npz. Use another format (such as pickle) to store it.
d['descr'] = dtype_to_descr(array.dtype)

Python and NumPy Versions:

Numpy version
2.3.1

Python version
3.13.5 | packaged by conda-forge | (main, Jun 16 2025, 08:20:19)

Runtime Environment:

No response

Context for the issue:

I'm trying to do some multiprocessing with numpy arrays. The multiprocessing.Pool.map() uses pickle for serialization, which led to this issue.

It's pretty surprising that a visible UserWarning is produced because of some internal magic undocumented metadata silliness.