sparse option to reindex and unstack

doc/whats-new.rst

@@ -38,6 +38,10 @@ Breaking changes
 
 New Features
 ~~~~~~~~~~~~
+- Added the ``sparse`` option to :py:meth:`~xarray.DataArray.unstack`, 
+  :py:meth:`~xarray.Dataset.unstack`, :py:meth:`~xarray.DataArray.reindex`, 
+  :py:meth:`~xarray.Dataset.reindex` (:issue:`3518`).
+  By `Keisuke Fujii <https://github.com/fujiisoup>`_.
 
 - Added the ``fill_value`` option to :py:meth:`~xarray.DataArray.unstack` and
   :py:meth:`~xarray.Dataset.unstack` (:issue:`3518`).

xarray/core/alignment.py

@@ -466,6 +466,7 @@ def reindex_variables(
     tolerance: Any = None,
     copy: bool = True,
     fill_value: Optional[Any] = dtypes.NA,
+    sparse: bool = False,
 ) -> Tuple[Dict[Hashable, Variable], Dict[Hashable, pd.Index]]:
     """Conform a dictionary of aligned variables onto a new set of variables,
     filling in missing values with NaN.
@@ -503,6 +504,8 @@ def reindex_variables(
         the input. In either case, new xarray objects are always returned.
     fill_value : scalar, optional
         Value to use for newly missing values
+    sparse: bool, optional
+        Use an sparse-array
 
     Returns
     -------
@@ -571,6 +574,8 @@ def reindex_variables(
 
     for name, var in variables.items():
         if name not in indexers:
+            if sparse:
+                var = var._as_sparse(fill_value=fill_value)
             key = tuple(
                 slice(None) if d in unchanged_dims else int_indexers.get(d, slice(None))
                 for d in var.dims

xarray/core/dataarray.py

@@ -1729,6 +1729,7 @@ def unstack(
         self,
         dim: Union[Hashable, Sequence[Hashable], None] = None,
         fill_value: Any = dtypes.NA,
+        sparse: bool = False,
     ) -> "DataArray":
         """
         Unstack existing dimensions corresponding to MultiIndexes into
@@ -1742,6 +1743,7 @@ def unstack(
             Dimension(s) over which to unstack. By default unstacks all
             MultiIndexes.
         fill_value: value to be filled. By default, np.nan
+        sparse: use sparse-array if True
 
         Returns
         -------
@@ -1773,7 +1775,7 @@ def unstack(
         --------
         DataArray.stack
         """
-        ds = self._to_temp_dataset().unstack(dim, fill_value)
+        ds = self._to_temp_dataset().unstack(dim, fill_value, sparse)
         return self._from_temp_dataset(ds)
 
     def to_unstacked_dataset(self, dim, level=0):

xarray/core/dataset.py

@@ -2254,6 +2254,7 @@ def reindex(
         tolerance: Number = None,
         copy: bool = True,
         fill_value: Any = dtypes.NA,
+        sparse: bool = False,
         **indexers_kwargs: Any,
     ) -> "Dataset":
         """Conform this object onto a new set of indexes, filling in
@@ -2286,6 +2287,7 @@ def reindex(
             the input. In either case, a new xarray object is always returned.
         fill_value : scalar, optional
             Value to use for newly missing values
+        sparse: use sparse-array. By default, False
         **indexers_kwarg : {dim: indexer, ...}, optional
             Keyword arguments in the same form as ``indexers``.
             One of indexers or indexers_kwargs must be provided.
@@ -2444,6 +2446,7 @@ def reindex(
             tolerance,
             copy=copy,
             fill_value=fill_value,
+            sparse=sparse,
         )
         coord_names = set(self._coord_names)
         coord_names.update(indexers)
@@ -3333,7 +3336,7 @@ def ensure_stackable(val):
 
         return data_array
 
-    def _unstack_once(self, dim: Hashable, fill_value) -> "Dataset":
+    def _unstack_once(self, dim: Hashable, fill_value, sparse) -> "Dataset":
         index = self.get_index(dim)
         index = index.remove_unused_levels()
         full_idx = pd.MultiIndex.from_product(index.levels, names=index.names)
@@ -3342,7 +3345,9 @@ def _unstack_once(self, dim: Hashable, fill_value) -> "Dataset":
         if index.equals(full_idx):
             obj = self
         else:
-            obj = self.reindex({dim: full_idx}, copy=False, fill_value=fill_value)
+            obj = self.reindex(
+                {dim: full_idx}, copy=False, fill_value=fill_value, sparse=sparse
+            )
 
         new_dim_names = index.names
         new_dim_sizes = [lev.size for lev in index.levels]
@@ -3372,6 +3377,7 @@ def unstack(
         self,
         dim: Union[Hashable, Iterable[Hashable]] = None,
         fill_value: Any = dtypes.NA,
+        sparse: bool = False,
     ) -> "Dataset":
         """
         Unstack existing dimensions corresponding to MultiIndexes into
@@ -3385,6 +3391,7 @@ def unstack(
             Dimension(s) over which to unstack. By default unstacks all
             MultiIndexes.
         fill_value: value to be filled. By default, np.nan
+        sparse: use sparse-array if True
 
         Returns
         -------
@@ -3422,7 +3429,7 @@ def unstack(
 
         result = self.copy(deep=False)
         for dim in dims:
-            result = result._unstack_once(dim, fill_value)
+            result = result._unstack_once(dim, fill_value, sparse)
         return result
 
     def update(self, other: "CoercibleMapping", inplace: bool = None) -> "Dataset":

xarray/core/duck_array_ops.py

@@ -13,12 +13,14 @@
 
 from . import dask_array_ops, dtypes, npcompat, nputils
 from .nputils import nanfirst, nanlast
-from .pycompat import dask_array_type
+from .pycompat import dask_array_type, sparse_array_type
 
 try:
     import dask.array as dask_array
+    import sparse
 except ImportError:
     dask_array = None  # type: ignore
+    sparse = None  # type: ignore
 
 
 def _dask_or_eager_func(
@@ -251,6 +253,9 @@ def count(data, axis=None):
 
 def where(condition, x, y):
     """Three argument where() with better dtype promotion rules."""
+    # sparse support
+    if isinstance(x, sparse_array_type) or isinstance(y, sparse_array_type):
+        return sparse.where(condition, x, y)
     return _where(condition, *as_shared_dtype([x, y]))
 
 

xarray/core/variable.py

@@ -993,6 +993,32 @@ def chunk(self, chunks=None, name=None, lock=False):
 
         return type(self)(self.dims, data, self._attrs, self._encoding, fastpath=True)
 
+    def _as_sparse(self, sparse_format=_default, fill_value=dtypes.NA):
+        """
+        use sparse-array as backend.
+        """
+        import sparse
+
+        # TODO  what to do if dask-backended?
+        if fill_value is dtypes.NA:
+            dtype, fill_value = dtypes.maybe_promote(self.dtype)
+        else:
+            dtype = self.dtype
+
+        if sparse_format is _default:
+            sparse_format = "coo"
+        as_sparse = getattr(sparse, "as_{}".format(sparse_format.lower()))
+        data = as_sparse(self.data.astype(dtype), fill_value=fill_value)
+        return self._replace(data=data)
+
+    def _to_dense(self):
+        """
+        Change backend from sparse to np.array
+        """
+        if hasattr(self._data, "todense"):
+            return self._replace(data=self._data.todense())
+        return self.copy(deep=False)
+
     def isel(
         self: VariableType,
         indexers: Mapping[Hashable, Any] = None,
@@ -2021,6 +2047,14 @@ def chunk(self, chunks=None, name=None, lock=False):
         # Dummy - do not chunk. This method is invoked e.g. by Dataset.chunk()
         return self.copy(deep=False)
 
+    def _as_sparse(self, sparse_format=_default, fill_value=_default):
+        # Dummy
+        return self.copy(deep=False)
+
+    def _to_dense(self):
+        # Dummy
+        return self.copy(deep=False)
+
     def _finalize_indexing_result(self, dims, data):
         if getattr(data, "ndim", 0) != 1:
             # returns Variable rather than IndexVariable if multi-dimensional

xarray/tests/test_dataset.py

@@ -1748,6 +1748,21 @@ def test_reindex(self):
         actual = ds.reindex(x=[0, 1, 3], y=[0, 1])
         assert_identical(expected, actual)
 
+    @requires_sparse
+    def test_reindex_sparse(self):
+        data = create_test_data()
+        dim3 = list("abdeghijk")
+        actual = data.reindex(dim3=dim3, sparse=True)
+        expected = data.reindex(dim3=dim3, sparse=False)
+        for k, v in data.data_vars.items():
+            np.testing.assert_equal(actual[k].data.todense(), expected[k].data)
+
+        data["var3"] = data["var3"].astype(int)
+        actual = data.reindex(dim3=dim3, sparse=True, fill_value=-10)
+        expected = data.reindex(dim3=dim3, sparse=False, fill_value=-10)
+        for k, v in data.data_vars.items():
+            np.testing.assert_equal(actual[k].data.todense(), expected[k].data)
+
     def test_reindex_warning(self):
         data = create_test_data()
 
@@ -2811,6 +2826,23 @@ def test_unstack_fill_value(self):
         expected = ds["var"].unstack("index").fillna(-1).astype(np.int)
         assert actual.equals(expected)
 
+    @requires_sparse
+    def test_unstack_fill_value(self):
+        ds = xr.Dataset(
+            {"var": (("x",), np.arange(6))},
+            coords={"x": [0, 1, 2] * 2, "y": (("x",), ["a"] * 3 + ["b"] * 3)},
+        )
+        # make ds incomplete
+        ds = ds.isel(x=[0, 2, 3, 4]).set_index(index=["x", "y"])
+        # test fill_value
+        actual = ds.unstack("index", sparse=True)
+        expected = ds.unstack("index")
+        assert actual["var"].variable._to_dense().equals(expected["var"].variable)
+
+        actual = ds["var"].unstack("index", sparse=True)
+        expected = ds["var"].unstack("index")
+        assert actual.variable._to_dense().equals(expected.variable)
+
     def test_stack_unstack_fast(self):
         ds = Dataset(
             {

xarray/tests/test_variable.py

@@ -33,6 +33,7 @@
     assert_identical,
     raises_regex,
     requires_dask,
+    requires_sparse,
     source_ndarray,
 )
 
@@ -1862,6 +1863,17 @@ def test_getitem_with_mask_nd_indexer(self):
         )
 
 
+@requires_sparse
+class TestVariableWithSparse:
+    # TODO inherit VariableSubclassobjects to cover more tests
+
+    def test_as_sparse(self):
+        data = np.arange(12).reshape(3, 4)
+        var = Variable(("x", "y"), data)._as_sparse(fill_value=-1)
+        actual = var._to_dense()
+        assert_identical(var, actual)
+
+
 class TestIndexVariable(VariableSubclassobjects):
     cls = staticmethod(IndexVariable)