Dask-friendly nan check in xr.corr() and xr.cov()

doc/whats-new.rst

@@ -21,11 +21,14 @@ v0.18.3 (unreleased)
 
 New Features
 ~~~~~~~~~~~~
-
 - Allow assigning values to a subset of a dataset using positional or label-based
   indexing (:issue:`3015`, :pull:`5362`). By `Matthias Göbel <https://github.com/matzegoebel>`_.
 - Attempting to reduce a weighted object over missing dimensions now raises an error (:pull:`5362`).
   By `Mattia Almansi <https://github.com/malmans2>`_.
+- :py:func:`xarray.cov` and :py:func:`xarray.corr` now lazily check for missing
+  values if inputs are dask arrays (:issue:`4804`, :pull:`5284`).
+  By `Andrew Williams <https://github.com/AndrewWilliams3142>`_.
+
 
 Breaking changes
 ~~~~~~~~~~~~~~~~

xarray/core/computation.py

@@ -1352,12 +1352,23 @@ def _cov_corr(da_a, da_b, dim=None, ddof=0, method=None):
 
     # 2. Ignore the nans
     valid_values = da_a.notnull() & da_b.notnull()
+    valid_count = valid_values.sum(dim) - ddof
 
-    if not valid_values.all():
-        da_a = da_a.where(valid_values)
-        da_b = da_b.where(valid_values)
+    def _get_valid_values(da, other):
+        """
+        Function to lazily mask da_a and da_b
+        following a similar approach to
+        https://github.com/pydata/xarray/pull/4559
+        """
+        missing_vals = np.logical_or(da.isnull(), other.isnull())
+        if missing_vals.any():
+            da = da.where(~missing_vals)
+            return da
+        else:
+            return da
 
-    valid_count = valid_values.sum(dim) - ddof
+    da_a = da_a.map_blocks(_get_valid_values, args=[da_b])
+    da_b = da_b.map_blocks(_get_valid_values, args=[da_a])
 
     # 3. Detrend along the given dim
     demeaned_da_a = da_a - da_a.mean(dim=dim)

xarray/tests/test_computation.py

@@ -22,7 +22,7 @@
     unified_dim_sizes,
 )
 
-from . import has_dask, requires_dask
+from . import has_dask, raise_if_dask_computes, requires_dask
 
 dask = pytest.importorskip("dask")
 
@@ -1386,6 +1386,7 @@ def arrays_w_tuples():
         da.isel(time=range(2, 20)).rolling(time=3, center=True).mean(),
         xr.DataArray([[1, 2], [1, np.nan]], dims=["x", "time"]),
         xr.DataArray([[1, 2], [np.nan, np.nan]], dims=["x", "time"]),
+        xr.DataArray([[1, 2], [2, 1]], dims=["x", "time"]),
     ]
 
     array_tuples = [
@@ -1394,12 +1395,40 @@ def arrays_w_tuples():
         (arrays[1], arrays[1]),
         (arrays[2], arrays[2]),
         (arrays[2], arrays[3]),
+        (arrays[2], arrays[4]),
+        (arrays[4], arrays[2]),
         (arrays[3], arrays[3]),
+        (arrays[4], arrays[4]),
     ]
 
     return arrays, array_tuples
 
 
+@pytest.mark.parametrize("ddof", [0, 1])
+@pytest.mark.parametrize(
+    "da_a, da_b",
+    [
+        arrays_w_tuples()[1][3],
+        arrays_w_tuples()[1][4],
+        arrays_w_tuples()[1][5],
+        arrays_w_tuples()[1][6],
+        arrays_w_tuples()[1][7],
+        arrays_w_tuples()[1][8],
+    ],
+)
+@pytest.mark.parametrize("dim", [None, "x", "time"])
+def test_lazy_corrcov(da_a, da_b, dim, ddof):
+    # GH 5284
+    from dask import is_dask_collection
+
+    with raise_if_dask_computes():
+        cov = xr.cov(da_a.chunk(), da_b.chunk(), dim=dim, ddof=ddof)
+        assert is_dask_collection(cov)
+
+        corr = xr.corr(da_a.chunk(), da_b.chunk(), dim=dim)
+        assert is_dask_collection(corr)
+
+
 @pytest.mark.parametrize("ddof", [0, 1])
 @pytest.mark.parametrize(
     "da_a, da_b",