BUG: `np.datetime64("nat", "ms")` does not use unit in `repr` (unlike timedelta)

[mroeschke]

Describe the issue:

When astypeing a unit-less datetime64 NaT to the same type with specified unit, the unit is not preserved unlike the analgous operation for timedelta64

Reproduce the code example:

>>> import numpy as np
>>> np.__version__
'2.3.0.dev0+git20250301.dfb054e'
>>> np.timedelta64("nat")
np.timedelta64('NaT')
>>> np.timedelta64("nat").astype(np.dtype("timedelta64[ns]"))
np.timedelta64('NaT','ns')
>>> np.datetime64("nat")
np.datetime64('NaT')
>>> np.datetime64("nat").astype(np.dtype("datetime64[ns]"))
np.datetime64('NaT')

Error message:

Python and NumPy Versions:

'2.3.0.dev0+git20250301.dfb054e'

Runtime Environment:

No response

Context for the issue:

No response

[seberg]

Thanks for reporting. The difference is actually that the repr for datetime64 NaT doesn't include the unit (while it does for timedelta64, which I agree is better):

np.datetime64("nat", "ms")
# np.datetime64('NaT')
np.datetime64("nat", "ms").dtype
# dtype('<M8[ms]')
np.datetime64("nat").astype(np.dtype("datetime64[ns]")).dtype
# dtype('<M8[ns]')

So, the unit is correctly preserved/stored, but not correctly shown.

The (ancient) reason is presumably, that datetime prints dates as e.g. "2014-12-12T12:12:12" which usually implicitly indicates the unit by how much is printed (e.g. in that case it stops with seconds).

I would agree we should show it for NaT and honestly wouldn't showing it generally (although we should maybe discuss it and/or include that in the print_options legacy option).