[SPARK-4384] [PySpark] improve sort spilling

If there some big broadcasts (or other object) in Python worker, the free memory could be used for sorting will be too small, then it will keep spilling small files into disks, finally failed with too many open files.

This PR try to delay the spilling until the used memory goes over limit and start to increase since last spilling, it will increase the size of spilling files, improve the stability and performance in this cases. (We also do this in ExternalAggregator).

Author: Davies Liu <[email protected]>

Closes apache#3252 from davies/sort and squashes the following commits:

711fb6c [Davies Liu] improve sort spilling

python/pyspark/shuffle.py

@@ -478,13 +478,21 @@ def _get_path(self, n):
             os.makedirs(d)
         return os.path.join(d, str(n))
 
+    def _next_limit(self):
+        """
+        Return the next memory limit. If the memory is not released
+        after spilling, it will dump the data only when the used memory
+        starts to increase.
+        """
+        return max(self.memory_limit, get_used_memory() * 1.05)
+
     def sorted(self, iterator, key=None, reverse=False):
         """
         Sort the elements in iterator, do external sort when the memory
         goes above the limit.
         """
         global MemoryBytesSpilled, DiskBytesSpilled
-        batch = 100
+        batch, limit = 100, self._next_limit()
         chunks, current_chunk = [], []
         iterator = iter(iterator)
         while True:
@@ -504,6 +512,7 @@ def sorted(self, iterator, key=None, reverse=False):
                 chunks.append(self.serializer.load_stream(open(path)))
                 current_chunk = []
                 gc.collect()
+                limit = self._next_limit()
                 MemoryBytesSpilled += (used_memory - get_used_memory()) << 20
                 DiskBytesSpilled += os.path.getsize(path)