Harmonize FillValue and missing_value during encoding and decodin…

…g steps (#3502)

* Replace `equivalent()` with `allclose_or_equiv()`

* Ensure _FillValue & missing_value are cast to same dtype as data's

* Use Numpy scalar during type casting

* Update ValueError message

* Formatting only

* Update whats-new.rst

doc/whats-new.rst

@@ -79,6 +79,8 @@ New Features
 
 Bug fixes
 ~~~~~~~~~
+- Harmonize `_FillValue`, `missing_value` during encoding and decoding steps. (:pull:`3502`)
+  By `Anderson Banihirwe <https://github.com/andersy005>`_. 
 - Fix regression introduced in v0.14.0 that would cause a crash if dask is installed
   but cloudpickle isn't (:issue:`3401`) by `Rhys Doyle <https://github.com/rdoyle45>`_
 - Fix grouping over variables with NaNs. (:issue:`2383`, :pull:`3406`).

xarray/coding/variables.py

@@ -8,7 +8,6 @@
 
 from ..core import dtypes, duck_array_ops, indexing
 from ..core.pycompat import dask_array_type
-from ..core.utils import equivalent
 from ..core.variable import Variable
 
 
@@ -152,18 +151,25 @@ def encode(self, variable, name=None):
         fv = encoding.get("_FillValue")
         mv = encoding.get("missing_value")
 
-        if fv is not None and mv is not None and not equivalent(fv, mv):
+        if (
+            fv is not None
+            and mv is not None
+            and not duck_array_ops.allclose_or_equiv(fv, mv)
+        ):
             raise ValueError(
-                "Variable {!r} has multiple fill values {}. "
-                "Cannot encode data. ".format(name, [fv, mv])
+                f"Variable {name!r} has conflicting _FillValue ({fv}) and missing_value ({mv}). Cannot encode data."
             )
 
         if fv is not None:
+            # Ensure _FillValue is cast to same dtype as data's
+            encoding["_FillValue"] = data.dtype.type(fv)
             fill_value = pop_to(encoding, attrs, "_FillValue", name=name)
             if not pd.isnull(fill_value):
                 data = duck_array_ops.fillna(data, fill_value)
 
         if mv is not None:
+            # Ensure missing_value is cast to same dtype as data's
+            encoding["missing_value"] = data.dtype.type(mv)
             fill_value = pop_to(encoding, attrs, "missing_value", name=name)
             if not pd.isnull(fill_value) and fv is None:
                 data = duck_array_ops.fillna(data, fill_value)

xarray/tests/test_coding.py

@@ -20,6 +20,23 @@ def test_CFMaskCoder_decode():
     assert_identical(expected, encoded)
 
 
+def test_CFMaskCoder_encode_missing_fill_values_conflict():
+    original = xr.Variable(
+        ("x",),
+        [0.0, -1.0, 1.0],
+        encoding={"_FillValue": np.float32(1e20), "missing_value": np.float64(1e20)},
+    )
+    coder = variables.CFMaskCoder()
+    encoded = coder.encode(original)
+
+    assert encoded.dtype == encoded.attrs["missing_value"].dtype
+    assert encoded.dtype == encoded.attrs["_FillValue"].dtype
+
+    with pytest.warns(variables.SerializationWarning):
+        roundtripped = coder.decode(coder.encode(original))
+        assert_identical(roundtripped, original)
+
+
 def test_CFMaskCoder_missing_value():
     expected = xr.DataArray(
         np.array([[26915, 27755, -9999, 27705], [25595, -9999, 28315, -9999]]),