Support for dask.graph_manipulation

doc/whats-new.rst

@@ -22,7 +22,9 @@ v0.17.1 (unreleased)
 
 New Features
 ~~~~~~~~~~~~
-
+- Support for `dask.graph_manipulation
+  <https://docs.dask.org/en/latest/graph_manipulation.html>`_ (requires dask >=2021.3)
+  By `Guido Imperiale <https://github.com/crusaderky>`_
 
 Breaking changes
 ~~~~~~~~~~~~~~~~

xarray/core/dataarray.py

@@ -839,15 +839,15 @@ def __dask_scheduler__(self):
 
     def __dask_postcompute__(self):
         func, args = self._to_temp_dataset().__dask_postcompute__()
-        return self._dask_finalize, (func, args, self.name)
+        return self._dask_finalize, (self.name, func) + args
 
     def __dask_postpersist__(self):
         func, args = self._to_temp_dataset().__dask_postpersist__()
-        return self._dask_finalize, (func, args, self.name)
+        return self._dask_finalize, (self.name, func) + args
 
     @staticmethod
-    def _dask_finalize(results, func, args, name):
-        ds = func(results, *args)
+    def _dask_finalize(results, name, func, *args, **kwargs):
+        ds = func(results, *args, **kwargs)
         variable = ds._variables.pop(_THIS_ARRAY)
         coords = ds._variables
         return DataArray(variable, coords, name=name, fastpath=True)

xarray/core/dataset.py

@@ -863,72 +863,83 @@ def __dask_scheduler__(self):
         return da.Array.__dask_scheduler__
 
     def __dask_postcompute__(self):
+        return self._dask_postcompute, ()
+
+    def __dask_postpersist__(self):
+        return self._dask_postpersist, ()
+
+    def _dask_postcompute(self, results: "Iterable[Variable]") -> "Dataset":
         import dask
 
-        info = [
-            (k, None) + v.__dask_postcompute__()
-            if dask.is_dask_collection(v)
-            else (k, v, None, None)
-            for k, v in self._variables.items()
-        ]
-        construct_direct_args = (
+        variables = {}
+        results_iter = iter(results)
+
+        for k, v in self._variables.items():
+            if dask.is_dask_collection(v):
+                rebuild, args = v.__dask_postcompute__()
+                v = rebuild(next(results_iter), *args)
+            variables[k] = v
+
+        return Dataset._construct_direct(
+            variables,
             self._coord_names,
             self._dims,
             self._attrs,
             self._indexes,
             self._encoding,
             self._close,
         )
-        return self._dask_postcompute, (info, construct_direct_args)
 
-    def __dask_postpersist__(self):
-        import dask
+    def _dask_postpersist(
+        self, dsk: Mapping, *, rename: Mapping[str, str] = None
+    ) -> "Dataset":
+        from dask import is_dask_collection
+        from dask.highlevelgraph import HighLevelGraph
+        from dask.optimization import cull
 
-        info = [
-            (k, None, v.__dask_keys__()) + v.__dask_postpersist__()
-            if dask.is_dask_collection(v)
-            else (k, v, None, None, None)
-            for k, v in self._variables.items()
-        ]
-        construct_direct_args = (
+        variables = {}
+
+        for k, v in self._variables.items():
+            if not is_dask_collection(v):
+                variables[k] = v
+                continue
+
+            if isinstance(dsk, HighLevelGraph):
+                # dask >= 2021.3
+                # __dask_postpersist__() was called by dask.highlevelgraph.
+                # Don't use dsk.cull(), as we need to prevent partial layers:
+                # https://github.com/dask/dask/issues/7137
+                layers = v.__dask_layers__()
+                if rename:
+                    layers = [rename.get(k, k) for k in layers]
+                dsk2 = dsk.cull_layers(layers)
+            elif rename:  # pragma: nocover
+                # At the moment of writing, this is only for forward compatibility.
+                # replace_name_in_key requires dask >= 2021.3.
+                from dask.base import flatten, replace_name_in_key
+
+                keys = [
+                    replace_name_in_key(k, rename) for k in flatten(v.__dask_keys__())
+                ]
+                dsk2, _ = cull(dsk, keys)
+            else:
+                # __dask_postpersist__() was called by dask.optimize or dask.persist
+                dsk2, _ = cull(dsk, v.__dask_keys__())
+
+            rebuild, args = v.__dask_postpersist__()
+            # rename was added in dask 2021.3
+            kwargs = {"rename": rename} if rename else {}
+            variables[k] = rebuild(dsk2, *args, **kwargs)
+
+        return Dataset._construct_direct(
+            variables,
             self._coord_names,
             self._dims,
             self._attrs,
             self._indexes,
             self._encoding,
             self._close,
         )
-        return self._dask_postpersist, (info, construct_direct_args)
-
-    @staticmethod
-    def _dask_postcompute(results, info, construct_direct_args):
-        variables = {}
-        results_iter = iter(results)
-        for k, v, rebuild, rebuild_args in info:
-            if v is None:
-                variables[k] = rebuild(next(results_iter), *rebuild_args)
-            else:
-                variables[k] = v
-
-        final = Dataset._construct_direct(variables, *construct_direct_args)
-        return final
-
-    @staticmethod
-    def _dask_postpersist(dsk, info, construct_direct_args):
-        from dask.optimization import cull
-
-        variables = {}
-        # postpersist is called in both dask.optimize and dask.persist
-        # When persisting, we want to filter out unrelated keys for
-        # each Variable's task graph.
-        for k, v, dask_keys, rebuild, rebuild_args in info:
-            if v is None:
-                dsk2, _ = cull(dsk, dask_keys)
-                variables[k] = rebuild(dsk2, *rebuild_args)
-            else:
-                variables[k] = v
-
-        return Dataset._construct_direct(variables, *construct_direct_args)
 
     def compute(self, **kwargs) -> "Dataset":
         """Manually trigger loading and/or computation of this dataset's data

xarray/core/variable.py

@@ -531,22 +531,15 @@ def __dask_scheduler__(self):
 
     def __dask_postcompute__(self):
         array_func, array_args = self._data.__dask_postcompute__()
-        return (
-            self._dask_finalize,
-            (array_func, array_args, self._dims, self._attrs, self._encoding),
-        )
+        return self._dask_finalize, (array_func,) + array_args
 
     def __dask_postpersist__(self):
         array_func, array_args = self._data.__dask_postpersist__()
-        return (
-            self._dask_finalize,
-            (array_func, array_args, self._dims, self._attrs, self._encoding),
-        )
+        return self._dask_finalize, (array_func,) + array_args
 
-    @staticmethod
-    def _dask_finalize(results, array_func, array_args, dims, attrs, encoding):
-        data = array_func(results, *array_args)
-        return Variable(dims, data, attrs=attrs, encoding=encoding)
+    def _dask_finalize(self, results, array_func, *args, **kwargs):
+        data = array_func(results, *args, **kwargs)
+        return Variable(self._dims, data, attrs=self._attrs, encoding=self._encoding)
 
     @property
     def values(self):

xarray/tests/test_dask.py

@@ -1599,3 +1599,38 @@ def test_optimize():
     arr = xr.DataArray(a).chunk(5)
     (arr2,) = dask.optimize(arr)
     arr2.compute()
+
+
+# The graph_manipulation module is in dask since 2021.2 but it became usable with
+# xarray only since 2021.3
+@pytest.mark.skipif(LooseVersion(dask.__version__) <= "2021.02.0", reason="new module")
+def test_graph_manipulation():
+    """dask.graph_manipulation passes an optional parameter, "rename", to the rebuilder
+    function returned by __dask_postperist__; also, the dsk passed to the rebuilder is
+    a HighLevelGraph whereas with dask.persist() and dask.optimize() it's a plain dict.
+    """
+    import dask.graph_manipulation as gm
+
+    v = Variable(["x"], [1, 2]).chunk(-1).chunk(1) * 2
+    da = DataArray(v)
+    ds = Dataset({"d1": v[0], "d2": v[1], "d3": ("x", [3, 4])})
+
+    v2, da2, ds2 = gm.clone(v, da, ds)
+
+    assert_equal(v2, v)
+    assert_equal(da2, da)
+    assert_equal(ds2, ds)
+
+    for a, b in ((v, v2), (da, da2), (ds, ds2)):
+        assert a.__dask_layers__() != b.__dask_layers__()
+        assert len(a.__dask_layers__()) == len(b.__dask_layers__())
+        assert a.__dask_graph__().keys() != b.__dask_graph__().keys()
+        assert len(a.__dask_graph__()) == len(b.__dask_graph__())
+        assert a.__dask_graph__().layers.keys() != b.__dask_graph__().layers.keys()
+        assert len(a.__dask_graph__().layers) == len(b.__dask_graph__().layers)
+
+    # Above we performed a slice operation; adding the two slices back together creates
+    # a diamond-shaped dependency graph, which in turn will trigger a collision in layer
+    # names if we were to use HighLevelGraph.cull() instead of
+    # HighLevelGraph.cull_layers() in Dataset.__dask_postpersist__().
+    assert_equal(ds2.d1 + ds2.d2, ds.d1 + ds.d2)