BUG: series construction from dict of Timedelta scalar doesn't work

[arw2019]

Construction of Series from a dictionary of scalars sometimes fails:

Code Sample, a copy-pastable example

In [31]: import pandas as pd
    ...: import pandas._testing as tm
    ...: 
    ...: td = pd.Timedelta(nanoseconds=500)
    ...: ser = pd.Series({"a": td})
    ...: expected = pd.Series(td, index=["a"], dtype="timedelta64[ns]")
    ...: 
    ...: tm.assert_series_equal(ser, expected)
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<ipython-input-31-a9c6a6312101> in <module>
      6 expected = pd.Series(td, index=["a"], dtype="timedelta64[ns]")
      7 
----> 8 tm.assert_series_equal(ser, expected)

    [... skipping hidden 1 frame]

~/repos/pandas/pandas/_testing.py in assert_extension_array_equal(left, right, check_dtype, index_values, check_less_precise, check_exact, rtol, atol)
   1243         # Avoid slow object-dtype comparisons
   1244         # np.asarray for case where we have a np.MaskedArray
-> 1245         assert_numpy_array_equal(
   1246             np.asarray(left.asi8), np.asarray(right.asi8), index_values=index_values
   1247         )

    [... skipping hidden 1 frame]

~/repos/pandas/pandas/_testing.py in _raise(left, right, err_msg)
   1155             diff = diff * 100.0 / left.size
   1156             msg = f"{obj} values are different ({np.round(diff, 5)} %)"
-> 1157             raise_assert_detail(obj, msg, left, right, index_values=index_values)
   1158 
   1159         raise AssertionError(err_msg)

~/repos/pandas/pandas/_testing.py in raise_assert_detail(obj, message, left, right, diff, index_values)
   1085         msg += f"\n[diff]: {diff}"
   1086 
-> 1087     raise AssertionError(msg)
   1088 
   1089 

AssertionError: numpy array are different

numpy array values are different (100.0 %)
[index]: [a]
[left]:  [500]
[right]: [0]

Ran on 1.2 master

[jorisvandenbossche]

It seems it is actually the "expected" construction that doesn't work, and not the construction from a dict:

In [34]: td
Out[34]: Timedelta('0 days 00:00:00.000000500')

In [35]: pd.Series({"a": td})
Out[35]: 
a   0 days 00:00:00.000000500
dtype: timedelta64[ns]

In [36]: pd.Series(td, index=["a"], dtype="timedelta64[ns]")
Out[36]: 
a   0 days
dtype: timedelta64[ns]

So this last method seems to lose the nanoseconds ..

[ma3da]

pd.Timedelta insertion into an ndarray behaves strangely for small values.

>>> a = np.empty(1, dtype="timedelta64[ns]")
>>> a.fill(pd.Timedelta(999))
>>> a
array([0], dtype='timedelta64[ns]')
>>> a.fill(pd.Timedelta(1000))
>>> a
array([1000], dtype='timedelta64[ns]')

For this issue, this happens here.

[jorisvandenbossche]

Ah, that's because numpy only knows datetime.timedelta, and thus handles a pandas.Timedelta that way, and datetme.timedelta don't have nanoseconds ..

So we will need to handle that a bit differently on the pandas side for datetime/timedelta dtypes (eg filling the array with the integer representation fo the datetime/timedelta, and casting to datetime64/timedelta64 afterwards).

[jorisvandenbossche]

@ma3da want to do a PR for this?

[ma3da]

Ah ok, so there's no hope to fix the np cast "directly". Ok, i'll look into it.