[SPARK-8202] [PYSPARK] fix infinite loop during external sort in PySpark

The batch size during external sort will grow up to max 10000, then shrink down to zero, causing infinite loop.
Given the assumption that the items usually have similar size, so we don't need to adjust the batch size after first spill.

cc JoshRosen rxin angelini

Author: Davies Liu <[email protected]>

Closes #6714 from davies/batch_size and squashes the following commits:

b170dfb [Davies Liu] update test
b9be832 [Davies Liu] Merge branch 'batch_size' of github.com:davies/spark into batch_size
6ade745 [Davies Liu] update test
5c21777 [Davies Liu] Update shuffle.py
e746aec [Davies Liu] fix batch size during sort

Author: Davies Liu

python/pyspark/shuffle.py

@@ -486,7 +486,7 @@ def sorted(self, iterator, key=None, reverse=False):
         goes above the limit.
         """
         global MemoryBytesSpilled, DiskBytesSpilled
-        batch, limit = 100, self.memory_limit
+        batch, limit = 100, self._next_limit()
         chunks, current_chunk = [], []
         iterator = iter(iterator)
         while True:
@@ -512,9 +512,6 @@ def load(f):
                     f.close()
                 chunks.append(load(open(path, 'rb')))
                 current_chunk = []
-                gc.collect()
-                batch //= 2
-                limit = self._next_limit()
                 MemoryBytesSpilled += max(used_memory - get_used_memory(), 0) << 20
                 DiskBytesSpilled += os.path.getsize(path)
                 os.unlink(path)  # data will be deleted after close

python/pyspark/tests.py

@@ -179,9 +179,12 @@ def test_in_memory_sort(self):
                          list(sorter.sorted(l, key=lambda x: -x, reverse=True)))
 
     def test_external_sort(self):
+        class CustomizedSorter(ExternalSorter):
+            def _next_limit(self):
+                return self.memory_limit
         l = list(range(1024))
         random.shuffle(l)
-        sorter = ExternalSorter(1)
+        sorter = CustomizedSorter(1)
         self.assertEqual(sorted(l), list(sorter.sorted(l)))
         self.assertGreater(shuffle.DiskBytesSpilled, 0)
         last = shuffle.DiskBytesSpilled