BUG: Fix read of py3 PeriodIndex DataFrame HDF made in py2 (#16781)

In Python3, reading a DataFrame with a PeriodIndex from an HDF file
created in Python2 would incorrectly return a DataFrame with an
Int64Index.

Author: forbdonut

pandas/io/pytables.py

@@ -2591,8 +2591,8 @@ def read_index_node(self, node, start=None, stop=None):
         if 'name' in node._v_attrs:
             name = _ensure_str(node._v_attrs.name)
 
-        index_class = self._alias_to_class(getattr(node._v_attrs,
-                                                   'index_class', ''))
+        index_class = self._alias_to_class(_ensure_decoded(
+            getattr(node._v_attrs, 'index_class', '')))
         factory = self._get_index_factory(index_class)
 
         kwargs = {}

pandas/tests/io/data/testpi.h5

@@ -5264,6 +5264,19 @@ def test_fspath(self):
             with pd.HDFStore(path) as store:
                 assert os.fspath(store) == str(path)
 
+    def test_read_py2_hdf_file_in_py3(self):
+        # GH 16781
+
+        # tests reading a PeriodIndex DataFrame written in Python2 in Python3
+        expected = pd.DataFrame([1., 2, 3], index=pd.PeriodIndex(
+            ['2015-01-01', '2015-01-02', '2015-01-05'], freq='B'))
+
+        with ensure_clean_store(
+                tm.get_data_path('testpi.h5'),
+                mode='r') as store:
+            result = store['p']
+            assert_frame_equal(result, expected)
+
 
 class TestHDFComplexValues(Base):
     # GH10447