[BACKPORT] Fix incorrect index_value in df.drop() (#1466) (#1488)

Author: Xuye (Chris) Qin

Diff for: mars/dataframe/base/drop.py

@@ -117,7 +117,7 @@ def tile(cls, op: 'DataFrameDrop'):
                     col_to_args[c.index[1]] = (new_dtypes, new_col_id)
 
                 params.update(dict(dtypes=new_dtypes, index=(c.index[0], new_col_id),
-                                   index_value=parse_index(None, (c.key, c.index_value.key))))
+                                   index_value=c.index_value))
                 if op.index is not None:
                     params.update(dict(shape=(np.nan, len(new_dtypes)),
                                        index_value=parse_index(None, (c.key, c.index_value.key))))

Diff for: mars/dataframe/base/tests/test_base.py

@@ -785,13 +785,25 @@ def testDrop(self):
         raw = pd.DataFrame(rs.randint(1000, size=(20, 8)),
                            columns=['c' + str(i + 1) for i in range(8)])
 
-        df = from_pandas_df(raw, chunk_size=3)
+        df = from_pandas_df(raw, chunk_size=8)
 
         with self.assertRaises(KeyError):
             df.drop(columns=['c9'])
         with self.assertRaises(NotImplementedError):
             df.drop(columns=from_pandas_series(pd.Series(['c9'])))
 
+        r = df.drop(columns=['c1'])
+        pd.testing.assert_index_equal(r.index_value.to_pandas(), raw.index)
+
+        tiled = r.tiles()
+        start = 0
+        for c in tiled.chunks:
+            raw_index = raw.index[start: start + c.shape[0]]
+            start += c.shape[0]
+            pd.testing.assert_index_equal(raw_index, c.index_value.to_pandas())
+
+        df = from_pandas_df(raw, chunk_size=3)
+
         columns = ['c2', 'c4', 'c5', 'c6']
         index = [3, 6, 7]
         r = df.drop(columns=columns, index=index)