Add xarray-specific encoding convention for pd.IntervalArray

xarray/coding/times.py

@@ -1358,9 +1358,9 @@ def __init__(
         self.time_unit = time_unit
 
     def encode(self, variable: Variable, name: T_Name = None) -> Variable:
-        if np.issubdtype(
-            variable.data.dtype, np.datetime64
-        ) or contains_cftime_datetimes(variable):
+        if np.issubdtype(variable.dtype, np.datetime64) or contains_cftime_datetimes(
+            variable
+        ):
             dims, data, attrs, encoding = unpack_for_encoding(variable)
 
             units = encoding.pop("units", None)
@@ -1477,7 +1477,7 @@ def __init__(
         self._emit_decode_timedelta_future_warning = False
 
     def encode(self, variable: Variable, name: T_Name = None) -> Variable:
-        if np.issubdtype(variable.data.dtype, np.timedelta64):
+        if np.issubdtype(variable.dtype, np.timedelta64):
             dims, data, attrs, encoding = unpack_for_encoding(variable)
             dtype = encoding.get("dtype", None)
             units = encoding.pop("units", None)

xarray/coding/variables.py

@@ -245,7 +245,7 @@
             # otherwise numpy unsigned ints will silently cast to the signed counterpart
             fill_value = fill_value.item()
         # passes if provided fill value fits in encoded on-disk type
         new_fill = encoded_dtype.type(fill_value)
     except OverflowError:
         encoded_kind_str = "signed" if encoded_dtype.kind == "i" else "unsigned"
         warnings.warn(
@@ -698,3 +698,53 @@
 
     def decode(self, variable: Variable, name: T_Name = None) -> Variable:
         raise NotImplementedError()
+
+
+class IntervalCoder(VariableCoder):
+    """
+    Xarray-specific Interval Coder to roundtrip 1D pd.IntervalArray objects.
+    """
+
+    encoded_dtype = "pandas_interval"
+    encoded_bounds_dim = "__xarray_bounds__"
+
+    def encode(self, variable: Variable, name: T_Name = None) -> Variable:
+        if isinstance(dtype := variable.dtype, pd.IntervalDtype):
+            dims, data, attrs, encoding = unpack_for_encoding(variable)
+
+            new_data = np.stack([data.left, data.right], axis=0)
+            dims = (self.encoded_bounds_dim, *dims)
+            safe_setitem(attrs, "closed", dtype.closed, name=name)
+            safe_setitem(attrs, "dtype", self.encoded_dtype, name=name)
+            safe_setitem(attrs, "bounds_dim", self.encoded_bounds_dim, name=name)
+            return Variable(dims, new_data, attrs, encoding, fastpath=True)
+        else:
+            return variable
+
+    def decode(self, variable: Variable, name: T_Name = None) -> Variable:
+        if (
+            variable.attrs.get("dtype", None) == self.encoded_dtype
+            and self.encoded_bounds_dim in variable.dims
+        ):
+            if variable.ndim != 2:
+                raise ValueError(
+                    f"Cannot decode intervals for variable named {name!r} with more than two dimensions."
+                )
+
+            dims, data, attrs, encoding = unpack_for_decoding(variable)
+            pop_to(attrs, encoding, "dtype", name=name)
+            pop_to(attrs, encoding, "bounds_dim", name=name)
+            closed = pop_to(attrs, encoding, "closed", name=name)
+
+            _, new_dims = variable.dims
+            variable = variable.load()
+            new_data = pd.arrays.IntervalArray.from_arrays(
+                variable.isel({self.encoded_bounds_dim: 0}).data,
+                variable.isel({self.encoded_bounds_dim: 1}).data,
+                closed=closed,
+            )
+            return Variable(
+                dims=new_dims, data=new_data, attrs=attrs, encoding=encoding
+            )
+        else:
+            return variable

xarray/conventions.py

@@ -90,6 +90,9 @@ def encode_cf_variable(
     ensure_not_multiindex(var, name=name)
 
     for coder in [
+        # IntervalCoder must be before CFDatetimeCoder,
+        # so we can first encode the interval, then datetimes if necessary
+        variables.IntervalCoder(),
         CFDatetimeCoder(),
         CFTimedeltaCoder(),
         variables.CFScaleOffsetCoder(),
@@ -238,6 +241,8 @@ def decode_cf_variable(
             )
         var = decode_times.decode(var, name=name)
 
+    var = variables.IntervalCoder().decode(var)
+
     if decode_endianness and not var.dtype.isnative:
         var = variables.EndianCoder().decode(var)
         original_dtype = var.dtype

xarray/tests/test_coding.py

@@ -147,3 +147,31 @@ def test_decode_signed_from_unsigned(bits) -> None:
     decoded = coder.decode(encoded)
     assert decoded.dtype == signed_dtype
     assert decoded.values == original_values
+
+
+@pytest.mark.parametrize(
+    "data",
+    [
+        [1, 2, 3, 4],
+        np.array([1, 2, 3, 4], dtype=float),
+        pd.date_range("2001-01-01", "2002-01-01", freq="MS"),
+    ],
+)
+@pytest.mark.parametrize("closed", ["left", "right", "both", "neither"])
+def test_roundtrip_pandas_interval(data, closed) -> None:
+    v = xr.Variable("time", pd.IntervalIndex.from_breaks(data, closed=closed))
+    coder = variables.IntervalCoder()
+    encoded = coder.encode(v)
+    expected = xr.Variable(
+        dims=("__xarray_bounds__", "time"),
+        data=np.stack([data[:-1], data[1:]], axis=0),
+        attrs={
+            "dtype": "pandas_interval",
+            "bounds_dim": "__xarray_bounds__",
+            "closed": closed,
+        },
+    )
+    assert_identical(encoded, expected)
+
+    decoded = coder.decode(encoded)
+    assert_identical(decoded, v)

xarray/tests/test_conventions.py

@@ -675,3 +675,23 @@ def test_decode_cf_variables_decode_timedelta_warning() -> None:
 
     with pytest.warns(FutureWarning, match="decode_timedelta"):
         conventions.decode_cf_variables(variables, {})
+
+
+@pytest.mark.parametrize(
+    "data",
+    [
+        [1, 2, 3, 4],
+        np.array([1, 2, 3, 4], dtype=float),
+        pd.date_range("2001-01-01", "2002-01-01", freq="MS"),
+    ],
+)
+@pytest.mark.parametrize("closed", ["left", "right", "both", "neither"])
+def test_roundtrip_pandas_interval(data, closed) -> None:
+    v = Variable("time", pd.IntervalIndex.from_breaks(data, closed=closed))
+    encoded = conventions.encode_cf_variable(v)
+    if isinstance(data, pd.DatetimeIndex):
+        # make sure we've encoded datetimes.
+        assert "units" in encoded.attrs
+        assert "calendar" in encoded.attrs
+    roundtripped = conventions.decode_cf_variable("foo", encoded)
+    assert_identical(roundtripped, v)