Fix to_iris conversion issues (#2111)

* TST: assert lazy array maintained by to_iris (#2046)

* Add masked_invalid array op, resolves to_iris rechunking issue (#2046)

* Fix dask_module in duck_array_ops.masked_invalid

* Really fix it

* Resolving to_iris dask array issues

Author: AlexHilson

doc/whats-new.rst

@@ -86,6 +86,8 @@ Bug fixes
   By `Deepak Cherian <https://github.com/dcherian>`_.
 - Colorbar limits are now determined by excluding ±Infs too.
   By `Deepak Cherian <https://github.com/dcherian>`_.
+- Fixed ``to_iris`` to maintain lazy dask array after conversion (:issue:`2046`).
+  By `Alex Hilson <https://github.com/AlexHilson>`_ and `Stephan Hoyer <https://github.com/shoyer>`_.
 
 .. _whats-new.0.10.3:
 

xarray/convert.py

@@ -6,6 +6,7 @@
 
 from .coding.times import CFDatetimeCoder, CFTimedeltaCoder
 from .conventions import decode_cf
+from .core import duck_array_ops
 from .core.dataarray import DataArray
 from .core.dtypes import get_fill_value
 from .core.pycompat import OrderedDict, range
@@ -94,7 +95,6 @@ def to_iris(dataarray):
     # Iris not a hard dependency
     import iris
     from iris.fileformats.netcdf import parse_cell_methods
-    from xarray.core.pycompat import dask_array_type
 
     dim_coords = []
     aux_coords = []
@@ -121,13 +121,7 @@ def to_iris(dataarray):
         args['cell_methods'] = \
             parse_cell_methods(dataarray.attrs['cell_methods'])
 
-    # Create the right type of masked array (should be easier after #1769)
-    if isinstance(dataarray.data, dask_array_type):
-        from dask.array import ma as dask_ma
-        masked_data = dask_ma.masked_invalid(dataarray)
-    else:
-        masked_data = np.ma.masked_invalid(dataarray)
-
+    masked_data = duck_array_ops.masked_invalid(dataarray.data)
     cube = iris.cube.Cube(masked_data, **args)
 
     return cube

xarray/core/duck_array_ops.py

@@ -101,6 +101,10 @@ def isnull(data):
 einsum = _dask_or_eager_func('einsum', array_args=slice(1, None),
                              requires_dask='0.17.3')
 
+masked_invalid = _dask_or_eager_func(
+    'masked_invalid', eager_module=np.ma,
+    dask_module=getattr(dask_array, 'ma', None))
+
 
 def asarray(data):
     return data if isinstance(data, dask_array_type) else np.asarray(data)

xarray/tests/test_dataarray.py

@@ -3038,12 +3038,11 @@ def test_to_and_from_iris_dask(self):
         roundtripped = DataArray.from_iris(actual)
         assert_identical(original, roundtripped)
 
-        # If the Iris version supports it then we should get a dask array back
+        # If the Iris version supports it then we should have a dask array
+        # at each stage of the conversion
         if hasattr(actual, 'core_data'):
-            pass
-            # TODO This currently fails due to the decoding loading
-            # the data (#1372)
-            # self.assertEqual(type(original.data), type(roundtripped.data))
+            self.assertEqual(type(original.data), type(actual.core_data()))
+            self.assertEqual(type(original.data), type(roundtripped.data))
 
         actual.remove_coord('time')
         auto_time_dimension = DataArray.from_iris(actual)