Merge pull request pandas-dev#8172 from behzadnouri/lj-cnt-sort

jreback · jreback · commit 7800290a1e9a · 2014-09-06T16:45:00.000-04:00
counting sort instead of np.argsort in left outer join
diff --git a/pandas/src/join.pyx b/pandas/src/join.pyx
@@ -106,15 +106,15 @@ def left_outer_join(ndarray[int64_t] left, ndarray[int64_t] right,
     if not sort:  # if not asked to sort, revert to original order
         if len(left) == len(left_indexer):
             # no multiple matches for any row on the left
-            # this is a short-cut to avoid np.argsort;
+            # this is a short-cut to avoid groupsort_indexer
             # otherwise, the `else` path also works in this case
             if left_sorter.dtype != np.int_:
                 left_sorter = left_sorter.astype(np.int_)
 
             rev = np.empty(len(left), dtype=np.int_)
             rev.put(left_sorter, np.arange(len(left)))
         else:
-            rev = np.argsort(left_indexer)
+            rev, _ = groupsort_indexer(left_indexer, len(left))
 
         right_indexer = right_indexer.take(rev)
         left_indexer = left_indexer.take(rev)
diff --git a/vb_suite/join_merge.py b/vb_suite/join_merge.py
@@ -237,3 +237,16 @@ def sample(values, k):
 join_non_unique_equal = Benchmark('fracofday * temp[fracofday.index]', setup,
                                    start_date=datetime(2013, 1, 1))
 
+
+setup = common_setup + '''
+np.random.seed(2718281)
+n = 50000
+
+left = DataFrame(np.random.randint(1, n/500, (n, 2)),
+        columns=['jim', 'joe'])
+
+right = DataFrame(np.random.randint(1, n/500, (n, 2)),
+        columns=['jolie', 'jolia']).set_index('jolie')
+'''
+
+left_outer_join_index = Benchmark("left.join(right, on='jim')", setup)
