Adapt to change in behavior of pandas.Timestamp constructor (#10944)

* Adapt to change in behavior of pandas.Timestamp constructor

* Fix typing

* Use safer return logic

Co-authored-by: Stephan Hoyer <[email protected]>

---------

Co-authored-by: Stephan Hoyer <[email protected]>

Author: spencerkclark

xarray/coding/times.py

@@ -98,6 +98,8 @@
     "scale_factor",
 ]
 
+_ORDERED_PANDAS_TIME_RESOLUTIONS: list[PDDatetimeUnitOptions] = ["s", "ms", "us", "ns"]
+
 
 def _is_standard_calendar(calendar: str) -> bool:
     return calendar.lower() in _STANDARD_CALENDARS
@@ -293,6 +295,21 @@ def _maybe_strip_tz_from_timestamp(date: pd.Timestamp) -> pd.Timestamp:
     return date
 
 
+def _cast_timestamp_to_coarsest_resolution(timestamp: pd.Timestamp) -> pd.Timestamp:
+    # Cast timestamp to the coarsest resolution that can be used without
+    # changing its value. If provided a string, the pandas.Timestamp
+    # constructor used to automatically infer this from the resolution of the
+    # string, but this behavior was changed in pandas-dev/pandas#62801. This
+    # function allows us to approximately restore the old behavior in a way
+    # that is perhaps more consistent with how we infer the resolution of the
+    # data values themselves.
+    for unit in _ORDERED_PANDAS_TIME_RESOLUTIONS:
+        coarsest_timestamp = timestamp.as_unit(unit)
+        if coarsest_timestamp == timestamp:
+            return coarsest_timestamp
+    return timestamp
+
+
 def _unpack_time_unit_and_ref_date(
     units: str,
 ) -> tuple[NPDatetimeUnitOptions, pd.Timestamp]:
@@ -301,6 +318,7 @@ def _unpack_time_unit_and_ref_date(
     time_unit, _ref_date = _unpack_netcdf_time_units(units)
     time_unit = _netcdf_to_numpy_timeunit(time_unit)
     ref_date = pd.Timestamp(_ref_date)
+    ref_date = _cast_timestamp_to_coarsest_resolution(ref_date)
     ref_date = _maybe_strip_tz_from_timestamp(ref_date)
     return time_unit, ref_date
 
@@ -442,8 +460,8 @@ def _check_higher_resolution(
     time_unit: PDDatetimeUnitOptions,
 ) -> tuple[np.ndarray, PDDatetimeUnitOptions]:
     """Iterate until fitting resolution found."""
-    res: list[PDDatetimeUnitOptions] = ["s", "ms", "us", "ns"]
-    new_units = res[res.index(time_unit) :]
+    index = _ORDERED_PANDAS_TIME_RESOLUTIONS.index(time_unit)
+    new_units = _ORDERED_PANDAS_TIME_RESOLUTIONS[index:]
     for new_time_unit in new_units:
         if not ((np.unique(flat_num_dates % 1) > 0).any() and new_time_unit != "ns"):
             break

xarray/tests/test_variable.py

@@ -1086,7 +1086,7 @@ def test_numpy_same_methods(self):
         [
             (np.datetime64("2000-01-01"), "s"),
             (
-                pd.Timestamp("2000-01-01T00"),
+                pd.Timestamp("2000-01-01T00").as_unit("s"),
                 "s" if has_pandas_3 else "ns",
             ),
             (
@@ -1128,7 +1128,7 @@ def test_0d_str(self):
         assert v.values == "foo".encode("ascii")
 
     def test_0d_datetime(self):
-        v = Variable([], pd.Timestamp("2000-01-01"))
+        v = Variable([], pd.Timestamp("2000-01-01").as_unit("s"))
         expected_unit = "s" if has_pandas_3 else "ns"
         assert v.dtype == np.dtype(f"datetime64[{expected_unit}]")
         assert v.values == np.datetime64("2000-01-01", expected_unit)  # type: ignore[call-overload]