Fix DataArray.stack() with non-unique coordinates on pandas 0.23 (#2168)

doc/whats-new.rst

@@ -59,6 +59,10 @@ Bug fixes
   dimension were improperly skipped.
   By `Stephan Hoyer <https://github.com/shoyer>`_
 
+- Fix :meth:`~DataArray.stack` with non-unique coordinates on pandas 0.23
+  (:issue:`2160`).
+  By `Stephan Hoyer <https://github.com/shoyer>`_
+
 - Selecting data indexed by a length-1 ``CFTimeIndex`` with a slice of strings
   now behaves as it does when using a length-1 ``DatetimeIndex`` (i.e. it no
   longer falsely returns an empty array when the slice includes the value in

xarray/core/utils.py

@@ -76,22 +76,24 @@ def safe_cast_to_index(array):
 def multiindex_from_product_levels(levels, names=None):
     """Creating a MultiIndex from a product without refactorizing levels.
 
-    Keeping levels the same is faster, and also gives back the original labels
-    when we unstack.
+    Keeping levels the same gives back the original labels when we unstack.
 
     Parameters
     ----------
-    levels : sequence of arrays
-        Unique labels for each level.
+    levels : sequence of pd.Index
+        Values for each MultiIndex level.
     names : optional sequence of objects
         Names for each level.
 
     Returns
     -------
     pandas.MultiIndex
     """
-    labels_mesh = np.meshgrid(*[np.arange(len(lev)) for lev in levels],
-                              indexing='ij')
+    if any(not isinstance(lev, pd.Index) for lev in levels):
+        raise TypeError('levels must be a list of pd.Index objects')
+
+    split_labels, levels = zip(*[lev.factorize() for lev in levels])
+    labels_mesh = np.meshgrid(*split_labels, indexing='ij')
     labels = [x.ravel() for x in labels_mesh]
     return pd.MultiIndex(levels, labels, sortorder=0, names=names)
 

xarray/tests/test_dataarray.py

@@ -1673,6 +1673,13 @@ def test_unstack_pandas_consistency(self):
         actual = DataArray(s, dims='z').unstack('z')
         assert_identical(expected, actual)
 
+    def test_stack_nonunique_consistency(self):
+        orig = DataArray([[0, 1], [2, 3]], dims=['x', 'y'],
+                         coords={'x': [0, 1], 'y': [0, 0]})
+        actual = orig.stack(z=['x', 'y'])
+        expected = DataArray(orig.to_pandas().stack(), dims='z')
+        assert_identical(expected, actual)
+
     def test_transpose(self):
         assert_equal(self.dv.variable.transpose(),
                      self.dv.transpose().variable)

xarray/tests/test_utils.py

@@ -72,7 +72,8 @@ def test_safe_cast_to_index_datetime_datetime(enable_cftimeindex):
 
 
 def test_multiindex_from_product_levels():
-    result = utils.multiindex_from_product_levels([['b', 'a'], [1, 3, 2]])
+    result = utils.multiindex_from_product_levels(
+        [pd.Index(['b', 'a']), pd.Index([1, 3, 2])])
     np.testing.assert_array_equal(
         result.labels, [[0, 0, 0, 1, 1, 1], [0, 1, 2, 0, 1, 2]])
     np.testing.assert_array_equal(result.levels[0], ['b', 'a'])
@@ -82,6 +83,15 @@ def test_multiindex_from_product_levels():
     np.testing.assert_array_equal(result.values, other.values)
 
 
+def test_multiindex_from_product_levels_non_unique():
+    result = utils.multiindex_from_product_levels(
+        [pd.Index(['b', 'a']), pd.Index([1, 1, 2])])
+    np.testing.assert_array_equal(
+        result.labels, [[0, 0, 0, 1, 1, 1], [0, 0, 1, 0, 0, 1]])
+    np.testing.assert_array_equal(result.levels[0], ['b', 'a'])
+    np.testing.assert_array_equal(result.levels[1], [1, 2])
+
+
 class TestArrayEquiv(TestCase):
     def test_0d(self):
         # verify our work around for pd.isnull not working for 0-dimensional