Series[offsets] + Series[datetime64] fails

[jbrockmendel]

ser = pd.Series([pd.Timestamp('2000-1-1'), pd.Timestamp('2000-2-1')])
other = pd.Series([pd.offsets.DateOffset(years=1), pd.offsets.MonthEnd()])

>>> ser + other
0   2001-01-01
1   2000-02-29
dtype: datetime64[ns]

>>> other + ser
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "pandas/core/ops.py", line 698, in wrapper
    result = wrap_results(safe_na_op(lvalues, rvalues))
  File "pandas/core/ops.py", line 665, in safe_na_op
    lambda x: op(x, rvalues))
  File "pandas/_libs/algos_common_helper.pxi", line 1212, in pandas._libs.algos.arrmap_object
    result[i] = func(index[i])
  File "pandas/core/ops.py", line 665, in <lambda>
    lambda x: op(x, rvalues))
  File "pandas/tseries/offsets.py", line 352, in __add__
    return self.apply(other)
  File "pandas/tseries/offsets.py", line 77, in wrapper
    result = func(self, other)
  File "pandas/tseries/offsets.py", line 212, in apply
    other = other + self._offset
TypeError: ufunc add cannot use operands with types dtype('<M8[ns]') and dtype('O')

pd.Index(other) + ser is OK, as is other + pd.DatetimeIndex(ser)

[chris-b1]

More fundamental than the arithmetic is that Series/arrays of offsets aren't meaningfully defined as a type, compared to, e.g. Timedelta.

It wouldn't be an unreasonable addition, something that would have occasionally been useful to me, but a pretty large change. Probably less painful after #19174, et al.

[chris-b1]

As you noted, this does actually sort of work (surprise to me!). I suppose may not need to be that invasive, could just involve the dispatch for + on object Series.

In [220]: pd.Index(other) + ser
C:\Users\chris.bartak\AppData\Local\Continuum\Anaconda3\envs\py36\lib\site-packages\pandas\core\ops.py:556: PerformanceWarning: Adding/subtracting array of DateOffsets to Series not vectorized
  "Series not vectorized", PerformanceWarning)
Out[220]: DatetimeIndex(['2001-01-01', '2000-02-29'], dtype='datetime64[ns]', freq=None)

[jbrockmendel]

The direction we’re going with series arithmetic is to dispatch to the relevant Index subclasses (#19024, #18824) implementations, since they handle type checking (and timezones) much more carefully than the current Series implementations.

[mroeschke]

Looks like this works now on master Could use a test

In [32]: ser = pd.Series([pd.Timestamp('2000-1-1'), pd.Timestamp('2000-2-1')])
    ...: other = pd.Series([pd.offsets.DateOffset(years=1), pd.offsets.MonthEnd()])

In [34]: other + ser
/Users/matthewroeschke/pandas-mroeschke/pandas/core/arrays/datetimelike.py:1305: PerformanceWarning: Adding/subtracting array of DateOffsets to DatetimeArray not vectorized
  PerformanceWarning,
Out[34]:
0   2001-01-01
1   2000-02-29
dtype: datetime64[ns]

In [37]: pd.__version__
Out[37]: '1.1.0.dev0+1068.g49bc8d8c9'

[ghost]

take

[jbrockmendel]

This is tested by tests.arithmetic.test_datetime64.test_dt64arr_add_sub_offset_array, among others. Closing.