BUG: Handle Datetimelike data in DataFrame.combine

[TomAugspurger]

Closes #23079

[pep8speaks]

Hello @TomAugspurger! Thanks for submitting the PR.

• There are no PEP8 issues in the file pandas/core/frame.py !

• There are no PEP8 issues in the file pandas/tests/frame/test_combine_concat.py !

[TomAugspurger]

After this, we'll need an additional change on the PeriodArray PR, so that the integers are converted back to periods:

commit f73f83e2878e04d266d8f86c2675d6c1fae6d90e
Author: Tom Augspurger <tom.w.augspurger@gmail.com>
Date:   Wed Oct 24 11:28:59 2018 -0500

    TST: Unxfail combine

diff --git a/pandas/core/dtypes/cast.py b/pandas/core/dtypes/cast.py
index 56bf39472..46c8126f6 100644
--- a/pandas/core/dtypes/cast.py
+++ b/pandas/core/dtypes/cast.py
@@ -6,7 +6,7 @@ import numpy as np
 import warnings
 
 from pandas._libs import tslib, lib, tslibs
-from pandas._libs.tslibs import iNaT, OutOfBoundsDatetime
+from pandas._libs.tslibs import iNaT, OutOfBoundsDatetime, Period
 from pandas.compat import string_types, text_type, PY3
 from .common import (ensure_object, is_bool, is_integer, is_float,
                      is_complex, is_datetimetz, is_categorical_dtype,
@@ -164,6 +164,12 @@ def maybe_downcast_to_dtype(result, dtype):
                     result = to_datetime(result).tz_localize('utc')
                     result = result.tz_convert(dtype.tz)
 
+        elif dtype.type == Period:
+            # TODO(DatetimeArray): merge with previous elif
+            from pandas.core.arrays import PeriodArray
+
+            return PeriodArray(result, freq=dtype.freq)
+
     except Exception:
         pass
 
diff --git a/pandas/tests/frame/test_combine_concat.py b/pandas/tests/frame/test_combine_concat.py
index 2803db4f4..3b8d6e6c5 100644
--- a/pandas/tests/frame/test_combine_concat.py
+++ b/pandas/tests/frame/test_combine_concat.py
@@ -759,7 +759,6 @@ class TestDataFrameCombineFirst(TestData):
         tm.assert_frame_equal(res, exp)
         assert res['TD'].dtype == 'timedelta64[ns]'
 
-    @pytest.mark.xfail(reason="GH-23079", strict=True)
     def test_combine_first_period(self):
         data1 = pd.PeriodIndex(['2011-01', 'NaT', '2011-03',
                                 '2011-04'], freq='M')

[codecov]

Codecov Report

Merging #23317 into master will decrease coverage by <.01%.
The diff coverage is 93.33%.

@@            Coverage Diff             @@
##           master   #23317      +/-   ##
==========================================
- Coverage   92.22%   92.22%   -0.01%     
==========================================
  Files         169      169              
  Lines       50911    50914       +3     
==========================================
+ Hits        46954    46956       +2     
- Misses       3957     3958       +1

Flag	Coverage Δ
#multiple	90.65% <93.33%> (-0.01%)	⬇️
#single	42.29% <0%> (-0.01%)	⬇️

Impacted Files	Coverage Δ
pandas/core/frame.py	96.98% <93.33%> (-0.05%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1c26375...b659913. Read the comment docs.

[TomAugspurger]

@jreback do you have a chance to look at this? It's one of the blockers for PeriodArray.

Originally, this was done as part of
#3595. I think the conversion of i8 needs to happen inside the combiner function, not outside of it.

[TomAugspurger]

Actually... one sec, this may not be a blocker.

[TomAugspurger]

Oh, yes, it is kinda a blocker, but I mis-diagnosed the real issue. The real thing we need on the PeriodArray PR is the change in maybe_downcast_to_dtype turning the i8 array back into a PeriodArray. But we'll still need this so that we handle the unpacking from series correctly.

So, I think this is still a valuable change on its own.

[jreback]

might be some more issues about this (which this may fix)

[TomAugspurger]

Didn't see any other issues that this fixes. #7509 is still broken with this PR.

[jreback]

not super duper happy that we have to have coercions like this. @jbrockmendel I view combine like update, in that its a bit orphaned. We do most numerics right now by dispatching thru ops.py but a lot of things are in internals, and others are in generic. So we need a coherent strategy for this. but ok for now.

[jreback]

thanks @TomAugspurger