Repository navigation
BUG: timedelta64 with ns units converts to python int rather than a python timedelta #28596
Description
Describe the issue:
When converting a timedelta64 with nanosecond units to a python timedelta object, the result unexpectedly has the int type. With microsecond (or coarser) units the results is a python timedelta object, as expected.
Instead of silently generating an object with an int type, I would have expected:
- A
timedeltaobject with the nanosecond resolution dropped (sincetimedeltaonly supports microsecond resolution), perhaps with a warning emitted. - An error.
Reproduce the code example:
>>> import numpy as np
>>> obj = np.timedelta64(1, "ns")
>>> from datetime import timedelta
>>> type(obj.astype("timedelta64[us]").astype(timedelta))
<class 'datetime.timedelta'>
>>> type(obj.astype("timedelta64[ns]").astype(timedelta))
<class 'int'>
>>> type(obj.astype(timedelta))
<class 'int'>Error message:
Python and NumPy Versions:
1.23.5
3.11.11 |Enthought, Inc. (x86_64)| (default, Mar 26 2025, 11:29:43) [Clang 15.0.0 (clang-1500.1.0.2.5)]
Runtime Environment:
No response
Context for the issue:
No response
Activity
[-]BUG: `timedelta64` with ns units converts to python int rather than a python datetime[/-][+]BUG: `timedelta64` with ns units converts to python `int` rather than a python `datetime`[/+]shoyer commentedon Mar 27, 2025
This is a long-term usability issue for NumPy's datetime/timedelta types.
I think a reasonable choice would be to convert to a
datetime.timedeltaonly if the nano-second part of the delta is zero. Otherwise we should raise error.Always raising an error and requiring users to switch to another precision would also be OK, although nano-second precision timedeltas that don't actually require nanosecond precision are common because until recently,
timedelta64[ns]was all that were supported by pandas.mdsmarte commentedon Mar 28, 2025
Note that the same problem also exists for
datetime64.....astype(datetime)converts to anintinstead of adatetimewhen it has has nanosecond units.[-]BUG: `timedelta64` with ns units converts to python `int` rather than a python `datetime`[/-][+]BUG: `timedelta64` with ns units converts to python `int` rather than a python `timdelta`[/+][-]BUG: `timedelta64` with ns units converts to python `int` rather than a python `timdelta`[/-][+]BUG: `timedelta64` with ns units converts to python `int` rather than a python `timedelta`[/+]Mohit-Kundu commentedon Apr 1, 2025
Hi @mdsmarte @shoyer, I'm willing to work on this!
jorenham commentedon Apr 9, 2025
I'm not a big fan of making the
timedelta64.item()type dependent on its value. It doesn't really feel very NEP-50 to me. But most importantly, that would make it impossible to annotate in the stubs.JeisonPham commentedon Jun 16, 2025
@Mohit-Kundu Are you still working on this? Can I take it over?