override units for datetime64/timedelta64 variables to preserve integer dtype

doc/whats-new.rst

@@ -87,6 +87,8 @@ Bug fixes
 - ``.rolling_exp`` functions no longer mistakenly lose non-dimensioned coords
   (:issue:`6528`, :pull:`8114`)
   By `Maximilian Roos <https://github.com/max-sixty>`_.
+- Override the user-provided encoding dtype for datetime64/timedelta64 values if it is an integer dtype, but the units encoding would require floating point values for most faithful serialization to disk (:issue:`1064`, :pull:`8201`).
+  By `Kai Mühlbauer <https://github.com/kmuehlbauer>`_.
 
 Documentation
 ~~~~~~~~~~~~~

xarray/coding/times.py

@@ -777,6 +777,14 @@ def encode(self, variable: Variable, name: T_Name = None) -> Variable:
             safe_setitem(attrs, "units", units, name=name)
             safe_setitem(attrs, "calendar", calendar, name=name)
 
+            # drop dtype in encoding if encoded as float64
+            # to prevent unnecessary casts, see GH #1064
+            encoding_dtype = encoding.get("dtype", data.dtype)
+            if np.issubdtype(data.dtype, np.float64) and np.issubdtype(
+                encoding_dtype, np.integer
+            ):
+                encoding.pop("dtype", None)
+
             return Variable(dims, data, attrs, encoding, fastpath=True)
         else:
             return variable
@@ -810,6 +818,11 @@ def encode(self, variable: Variable, name: T_Name = None) -> Variable:
             data, units = encode_cf_timedelta(data, encoding.pop("units", None))
             safe_setitem(attrs, "units", units, name=name)
 
+            # drop dtype in encoding if encoded as float64
+            # to prevent unnecessary casts, see GH #1064
+            if np.issubdtype(data.dtype, np.float64):
+                encoding.pop("dtype", None)
+
             return Variable(dims, data, attrs, encoding, fastpath=True)
         else:
             return variable

xarray/tests/test_coding_times.py

@@ -30,7 +30,7 @@
 from xarray.coding.variables import SerializationWarning
 from xarray.conventions import _update_bounds_attributes, cf_encoder
 from xarray.core.common import contains_cftime_datetimes
-from xarray.testing import assert_allclose, assert_equal, assert_identical
+from xarray.testing import assert_equal, assert_identical
 from xarray.tests import (
     FirstElementAccessibleArray,
     arm_xfail,
@@ -1212,6 +1212,7 @@ def test_contains_cftime_lazy() -> None:
         ("1677-09-21T00:12:43.145224193", "ns", np.int64, None, False),
         ("1677-09-21T00:12:43.145225", "us", np.int64, None, False),
         ("1970-01-01T00:00:01.000001", "us", np.int64, None, False),
+        ("1677-09-21T00:21:52.901038080", "ns", np.float32, 20.0, True),
     ],
 )
 def test_roundtrip_datetime64_nanosecond_precision(
@@ -1261,14 +1262,17 @@ def test_roundtrip_datetime64_nanosecond_precision_warning() -> None:
     ]
     units = "days since 1970-01-10T01:01:00"
     needed_units = "hours"
-    encoding = dict(_FillValue=20, units=units)
+    encoding = dict(dtype=np.int64, _FillValue=20, units=units)
     var = Variable(["time"], times, encoding=encoding)
     wmsg = (
         f"Times can't be serialized faithfully with requested units {units!r}. "
         f"Resolution of {needed_units!r} needed. "
     )
     with pytest.warns(UserWarning, match=wmsg):
         encoded_var = conventions.encode_cf_variable(var)
+    assert encoded_var.dtype == np.float64
+    assert encoded_var.attrs["units"] == units
+    assert encoded_var.attrs["_FillValue"] == 20.0
 
     decoded_var = conventions.decode_cf_variable("foo", encoded_var)
     assert_identical(var, decoded_var)
@@ -1311,12 +1315,16 @@ def test_roundtrip_timedelta64_nanosecond_precision_warning() -> None:
         f"Timedeltas can't be serialized faithfully with requested units {units!r}. "
         f"Resolution of {needed_units!r} needed. "
     )
-    encoding = dict(_FillValue=20, units=units)
+    encoding = dict(dtype=np.int64, _FillValue=20, units=units)
     var = Variable(["time"], timedelta_values, encoding=encoding)
     with pytest.warns(UserWarning, match=wmsg):
         encoded_var = conventions.encode_cf_variable(var)
+    assert encoded_var.dtype == np.float64
+    assert encoded_var.attrs["units"] == units
+    assert encoded_var.attrs["_FillValue"] == 20.0
     decoded_var = conventions.decode_cf_variable("foo", encoded_var)
-    assert_allclose(var, decoded_var)
+    assert_identical(var, decoded_var)
+    assert decoded_var.encoding["dtype"] == np.float64
 
 
 def test_roundtrip_float_times() -> None: