conversion of pandas Timestamps to numpy datetime64[ns] looses nanoseconds

There seems to be a problem when converting pandas Timestamps at ‘ns’ resolution into numpy datetime64[ns] arrays…

The following lines show that the nanosecond values are lost in the conversion! I’m not sure if the problem is with pandas or numpy but it results in some annoying problems like the fact that you can not search for a converted value in case the nanoseconds are !=0

Reproducing code example:

import numpy as np
import pandas as pd

x = pd.Timestamp('2017-01-01 09:58:50.624025600', unit='ns')
y = np.array(x, dtype=np.dtype('datetime64[ns]'))
y2 = np.array(x.value, dtype=np.dtype('datetime64[ns]'))

print(x) # the original pandas Timestamp
>>> '2017-01-01 09:58:50.624025600'
print(y)
>>> '2017-01-01T09:58:50.624025000'
print(y2)
>>> '2017-01-01T09:58:50.624025600'

print(y - y2):
>>> 'numpy.timedelta64(-600,'ns')'

Numpy/Python version information:

1.18.1 3.6.7 (default, Dec 6 2019, 07:03:06) [MSC v.1900 64 bit (AMD64)]

[jbrockmendel]

Note that the pandas Timestamp has a resolution attribute which does not extend to ns, even when the stored value contains nanoseconds. From the above example:

I think the resolution was (and is) correct, but __repr__ has recently been fixed to correctly show nanoseconds on that case.

[jbrockmendel]

Perhaps the fix here is for Timestamp to implement array

Agreed. A PR to do this would be welcome in pandas.