BUG: fix memory error in sortlevel when many multiindex levels. close #…

…2684

RELEASE.rst

@@ -104,6 +104,8 @@ pandas 0.10.1
   - Raise a more helpful error message in merge operations when one DataFrame
     has duplicate columns (GH2649_)
   - Fix partial date parsing issue occuring only when code is run at EOM  (GH2618_)
+  - Prevent MemoryError when using counting sort in sortlevel with
+    high-cardinality MultiIndex objects (GH2684_)
 
 **API Changes**
 
@@ -133,6 +135,7 @@ pandas 0.10.1
 .. _GH2643: https://github.com/pydata/pandas/issues/2643
 .. _GH2649: https://github.com/pydata/pandas/issues/2649
 .. _GH2668: https://github.com/pydata/pandas/issues/2668
+.. _GH2684: https://github.com/pydata/pandas/issues/2684
 .. _GH2689: https://github.com/pydata/pandas/issues/2689
 .. _GH2690: https://github.com/pydata/pandas/issues/2690
 .. _GH2692: https://github.com/pydata/pandas/issues/2692

pandas/core/groupby.py

@@ -2244,7 +2244,12 @@ def _indexer_from_factorized(labels, shape, compress=True):
         comp_ids = group_index
         max_group = np.prod(shape)
 
-    indexer, _ = _algos.groupsort_indexer(comp_ids.astype(np.int64), max_group)
+    if max_group > 1e6:
+        # Use mergesort to avoid memory errors in counting sort
+        indexer = comp_ids.argsort(kind='mergesort')
+    else:
+        indexer, _ = _algos.groupsort_indexer(comp_ids.astype(np.int64),
+                                              max_group)
 
     return indexer
 

pandas/tests/test_multilevel.py

@@ -612,6 +612,15 @@ def test_sortlevel(self):
         rs.sortlevel(0, inplace=True)
         assert_frame_equal(rs, self.frame.sortlevel(0))
 
+    def test_sortlevel_large_cardinality(self):
+        # #2684
+        index = MultiIndex.from_arrays([np.arange(4000)]*3)
+        df = DataFrame(np.random.randn(4000), index=index)
+
+        # it works!
+        result = df.sortlevel(0)
+        self.assertTrue(result.index.lexsort_depth == 3)
+
     def test_delevel_infer_dtype(self):
         tuples = [tuple for tuple in cart_product(['foo', 'bar'],
                                                   [10, 20], [1.0, 1.1])]