diff --git a/plotnine/scales/scale.py b/plotnine/scales/scale.py
index eec55dd29..78459daa9 100644
--- a/plotnine/scales/scale.py
+++ b/plotnine/scales/scale.py
@@ -221,7 +221,7 @@ def train_df(self, df):
         """
         Train scale from a dataframe
         """
-        aesthetics = set(self.aesthetics) & set(df.columns)
+        aesthetics = sorted(set(self.aesthetics) & set(df.columns))
         for ae in aesthetics:
             self.train(df[ae])
 
diff --git a/plotnine/tests/baseline_images/test_scale_internals/multiple_aesthetics.png b/plotnine/tests/baseline_images/test_scale_internals/multiple_aesthetics.png
index 1676a879a..f26e0b1fd 100644
Binary files a/plotnine/tests/baseline_images/test_scale_internals/multiple_aesthetics.png and b/plotnine/tests/baseline_images/test_scale_internals/multiple_aesthetics.png differ
diff --git a/plotnine/tests/test_scale_internals.py b/plotnine/tests/test_scale_internals.py
index d264039a7..8a4375965 100644
--- a/plotnine/tests/test_scale_internals.py
+++ b/plotnine/tests/test_scale_internals.py
@@ -402,8 +402,6 @@ def scale_scale_discrete_mapping_nulls():
     assert res[2] == expected[2]
 
 
-@pytest.mark.skipif(
-    six.PY2, reason="I do not know why it fails on travis")
 def test_multiple_aesthetics():
 
     df = pd.DataFrame({