Fix CFTimeIndex-related errors stemming from updates in pandas

doc/whats-new.rst

@@ -93,6 +93,12 @@ Internal Changes
 - Changed test_open_mfdataset_list_attr to only run with dask installed
   (:issue:`3777`, :pull:`3780`).
   By `Bruno Pagani <https://github.com/ArchangeGabriel>`_.
+- Preserved the ability to index with ``method="nearest"`` with a
+  :py:class:`CFTimeIndex` with pandas versions greater than 1.0.1
+  (:issue:`3751`). By `Spencer Clark <https://github.com/spencerkclark>`_.
+- Greater flexibility and improved test coverage of subtracting various types
+  of objects from a :py:class:`CFTimeIndex`. By `Spencer Clark
+  <https://github.com/spencerkclark>`_.
 - Updated Azure CI MacOS image, given pending removal.
   By `Maximilian Roos <https://github.com/max-sixty>`_
 - Removed xfails for scipy 1.0.1 for tests that append to netCDF files (:pull:`3805`).

xarray/coding/cftimeindex.py

@@ -49,6 +49,7 @@
 
 from xarray.core.utils import is_scalar
 
+from ..core.common import _contains_cftime_datetimes
 from .times import _STANDARD_CALENDARS, cftime_to_nptime, infer_calendar_name
 
 
@@ -326,6 +327,32 @@ def _get_string_slice(self, key):
             raise KeyError(key)
         return loc
 
+    def _get_nearest_indexer(self, target, limit, tolerance):
+        """Adapted from pandas.Index._get_nearest_indexer"""
+        left_indexer = self.get_indexer(target, "pad", limit=limit)
+        right_indexer = self.get_indexer(target, "backfill", limit=limit)
+        left_distances = abs(self.values[left_indexer] - target.values)
+        right_distances = abs(self.values[right_indexer] - target.values)
+
+        if self.is_monotonic_increasing:
+            condition = (left_distances < right_distances) | (right_indexer == -1)
+        else:
+            condition = (left_distances <= right_distances) | (right_indexer == -1)
+        indexer = np.where(condition, left_indexer, right_indexer)
+
+        if tolerance is not None:
+            indexer = self._filter_indexer_tolerance(target, indexer, tolerance)
+        return indexer
+
+    def _filter_indexer_tolerance(self, target, indexer, tolerance):
+        """Adapted from pandas.Index._filter_indexer_tolerance"""
+        if isinstance(target, pd.Index):
+            distance = abs(self.values[indexer] - target.values)
+        else:
+            distance = abs(self.values[indexer] - target)
+        indexer = np.where(distance <= tolerance, indexer, -1)
+        return indexer
+
     def get_loc(self, key, method=None, tolerance=None):
         """Adapted from pandas.tseries.index.DatetimeIndex.get_loc"""
         if isinstance(key, str):
@@ -427,24 +454,29 @@ def __radd__(self, other):
         return CFTimeIndex(other + np.array(self))
 
     def __sub__(self, other):
-        import cftime
-
-        if isinstance(other, (CFTimeIndex, cftime.datetime)):
+        if _contains_datetime_timedeltas(other):
+            return CFTimeIndex(np.array(self) - other)
+        elif isinstance(other, pd.TimedeltaIndex):
+            return CFTimeIndex(np.array(self) - other.to_pytimedelta())
+        elif _contains_cftime_datetimes(np.array(other)):
             try:
                 return pd.TimedeltaIndex(np.array(self) - np.array(other))
             except OverflowError:
                 raise ValueError(
                     "The time difference exceeds the range of values "
                     "that can be expressed at the nanosecond resolution."
                 )
-
-        elif isinstance(other, pd.TimedeltaIndex):
-            return CFTimeIndex(np.array(self) - other.to_pytimedelta())
         else:
-            return CFTimeIndex(np.array(self) - other)
+            return NotImplemented
 
     def __rsub__(self, other):
-        return pd.TimedeltaIndex(other - np.array(self))
+        try:
+            return pd.TimedeltaIndex(other - np.array(self))
+        except OverflowError:
+            raise ValueError(
+                "The time difference exceeds the range of values "
+                "that can be expressed at the nanosecond resolution."
+            )
 
     def to_datetimeindex(self, unsafe=False):
         """If possible, convert this index to a pandas.DatetimeIndex.
@@ -633,6 +665,12 @@ def _parse_array_of_cftime_strings(strings, date_type):
     ).reshape(strings.shape)
 
 
+def _contains_datetime_timedeltas(array):
+    """Check if an input array contains datetime.timedelta objects."""
+    array = np.atleast_1d(array)
+    return isinstance(array[0], timedelta)
+
+
 def _cftimeindex_from_i8(values, date_type, name):
     """Construct a CFTimeIndex from an array of integers.
 

xarray/tests/test_cftimeindex.py

@@ -451,10 +451,21 @@ def test_sel_date_scalar(da, date_type, index):
 
 
 @pytest.mark.xfail(reason="https://github.com/pydata/xarray/issues/3751")
+@requires_cftime
+def test_sel_date_distant_date(da, date_type, index):
+    expected = xr.DataArray(4).assign_coords(time=index[3])
+    result = da.sel(time=date_type(2000, 1, 1), method="nearest")
+    assert_identical(result, expected)
+
+
 @requires_cftime
 @pytest.mark.parametrize(
     "sel_kwargs",
-    [{"method": "nearest"}, {"method": "nearest", "tolerance": timedelta(days=70)}],
+    [
+        {"method": "nearest"},
+        {"method": "nearest", "tolerance": timedelta(days=70)},
+        {"method": "nearest", "tolerance": timedelta(days=1800000)},
+    ],
 )
 def test_sel_date_scalar_nearest(da, date_type, index, sel_kwargs):
     expected = xr.DataArray(2).assign_coords(time=index[1])
@@ -738,7 +749,7 @@ def test_timedeltaindex_add_cftimeindex(calendar):
 
 
 @requires_cftime
-def test_cftimeindex_sub(index):
+def test_cftimeindex_sub_timedelta(index):
     date_type = index.date_type
     expected_dates = [
         date_type(1, 1, 2),
@@ -753,6 +764,27 @@ def test_cftimeindex_sub(index):
     assert isinstance(result, CFTimeIndex)
 
 
+@requires_cftime
+@pytest.mark.parametrize(
+    "other",
+    [np.array(4 * [timedelta(days=1)]), np.array(timedelta(days=1))],
+    ids=["1d-array", "scalar-array"],
+)
+def test_cftimeindex_sub_timedelta_array(index, other):
+    date_type = index.date_type
+    expected_dates = [
+        date_type(1, 1, 2),
+        date_type(1, 2, 2),
+        date_type(2, 1, 2),
+        date_type(2, 2, 2),
+    ]
+    expected = CFTimeIndex(expected_dates)
+    result = index + timedelta(days=2)
+    result = result - other
+    assert result.equals(expected)
+    assert isinstance(result, CFTimeIndex)
+
+
 @requires_cftime
 @pytest.mark.parametrize("calendar", _CFTIME_CALENDARS)
 def test_cftimeindex_sub_cftimeindex(calendar):
@@ -784,6 +816,14 @@ def test_cftime_datetime_sub_cftimeindex(calendar):
     assert isinstance(result, pd.TimedeltaIndex)
 
 
+@requires_cftime
+@pytest.mark.parametrize("calendar", _CFTIME_CALENDARS)
+def test_distant_cftime_datetime_sub_cftimeindex(calendar):
+    a = xr.cftime_range("2000", periods=5, calendar=calendar)
+    with pytest.raises(ValueError, match="difference exceeds"):
+        a.date_type(1, 1, 1) - a
+
+
 @requires_cftime
 @pytest.mark.parametrize("calendar", _CFTIME_CALENDARS)
 def test_cftimeindex_sub_timedeltaindex(calendar):
@@ -795,6 +835,25 @@ def test_cftimeindex_sub_timedeltaindex(calendar):
     assert isinstance(result, CFTimeIndex)
 
 
+@requires_cftime
+@pytest.mark.parametrize("calendar", _CFTIME_CALENDARS)
+def test_cftimeindex_sub_index_of_cftime_datetimes(calendar):
+    a = xr.cftime_range("2000", periods=5, calendar=calendar)
+    b = pd.Index(a.values)
+    expected = a - a
+    result = a - b
+    assert result.equals(expected)
+    assert isinstance(result, pd.TimedeltaIndex)
+
+
+@requires_cftime
+@pytest.mark.parametrize("calendar", _CFTIME_CALENDARS)
+def test_cftimeindex_sub_not_implemented(calendar):
+    a = xr.cftime_range("2000", periods=5, calendar=calendar)
+    with pytest.raises(TypeError, match="unsupported operand"):
+        a - 1
+
+
 @requires_cftime
 def test_cftimeindex_rsub(index):
     with pytest.raises(TypeError):