diff --git a/.travis.yml b/.travis.yml
index 6f63f305ed2..efa903f5083 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -16,10 +16,10 @@ matrix:
   - env:
     - CONDA_ENV=py36
     - EXTRA_FLAGS="--run-flaky --run-network-tests"
+  - env: CONDA_ENV=py36-dask-dev
   - env: CONDA_ENV=py36-pandas-dev
   - env: CONDA_ENV=py36-rasterio
   - env: CONDA_ENV=py36-zarr-dev
-  - env: CONDA_ENV=py36-dask-dev
   - env: CONDA_ENV=docs
   - env: CONDA_ENV=lint
   - env: CONDA_ENV=typing
diff --git a/xarray/core/indexing.py b/xarray/core/indexing.py
index 1effb9347dd..1ba3175dc2f 100644
--- a/xarray/core/indexing.py
+++ b/xarray/core/indexing.py
@@ -453,7 +453,13 @@ def __array__(self, dtype=None):
 
     def __getitem__(self, key):
         key = expanded_indexer(key, self.ndim)
-        return self.array[self.indexer_cls(key)]
+        result = self.array[self.indexer_cls(key)]
+        if isinstance(result, ExplicitlyIndexed):
+            return type(self)(result, self.indexer_cls)
+        else:
+            # Sometimes explicitly indexed arrays return NumPy arrays or
+            # scalars.
+            return result
 
 
 class LazilyOuterIndexedArray(ExplicitlyIndexedNDArrayMixin):
diff --git a/xarray/core/variable.py b/xarray/core/variable.py
index ab1be181e31..cccb9663ad5 100644
--- a/xarray/core/variable.py
+++ b/xarray/core/variable.py
@@ -3,6 +3,7 @@
 import typing
 from collections import OrderedDict, defaultdict
 from datetime import timedelta
+from distutils.version import LooseVersion
 
 import numpy as np
 import pandas as pd
@@ -870,6 +871,7 @@ def chunk(self, chunks=None, name=None, lock=False):
         -------
         chunked : xarray.Variable
         """
+        import dask
         import dask.array as da
 
         if utils.is_dict_like(chunks):
@@ -892,7 +894,17 @@ def chunk(self, chunks=None, name=None, lock=False):
             # https://github.com/dask/dask/issues/2883
             data = indexing.ImplicitToExplicitIndexingAdapter(
                 data, indexing.OuterIndexer)
-            data = da.from_array(data, chunks, name=name, lock=lock)
+
+            # For now, assume that all arrays that we wrap with dask (including
+            # our lazily loaded backend array classes) should use NumPy array
+            # operations.
+            if LooseVersion(dask.__version__) > '1.2.2':
+                kwargs = dict(meta=np.ndarray)
+            else:
+                kwargs = dict()
+
+            data = da.from_array(
+                data, chunks, name=name, lock=lock, **kwargs)
 
         return type(self)(self.dims, data, self._attrs, self._encoding,
                           fastpath=True)
diff --git a/xarray/tests/test_indexing.py b/xarray/tests/test_indexing.py
index 9301abb5e32..59435fea88b 100644
--- a/xarray/tests/test_indexing.py
+++ b/xarray/tests/test_indexing.py
@@ -505,13 +505,20 @@ def test_decompose_indexers(shape, indexer_mode, indexing_support):
 
 
 def test_implicit_indexing_adapter():
-    array = np.arange(10)
+    array = np.arange(10, dtype=np.int64)
     implicit = indexing.ImplicitToExplicitIndexingAdapter(
         indexing.NumpyIndexingAdapter(array), indexing.BasicIndexer)
     np.testing.assert_array_equal(array, np.asarray(implicit))
     np.testing.assert_array_equal(array, implicit[:])
 
 
+def test_implicit_indexing_adapter_copy_on_write():
+    array = np.arange(10, dtype=np.int64)
+    implicit = indexing.ImplicitToExplicitIndexingAdapter(
+        indexing.CopyOnWriteArray(array))
+    assert isinstance(implicit[:], indexing.ImplicitToExplicitIndexingAdapter)
+
+
 def test_outer_indexer_consistency_with_broadcast_indexes_vectorized():
     def nonzero(x):
         if isinstance(x, np.ndarray) and x.dtype.kind == 'b':