[SPARK-6194] [SPARK-677] [PySpark] fix memory leak in collect()

Because circular reference between JavaObject and JavaMember, an Java object can not be released until Python GC kick in, then it will cause memory leak in collect(), which may consume lots of memory in JVM.

This PR change the way we sending collected data back into Python from local file to socket, which could avoid any disk IO during collect, also avoid any referrers of Java object in Python.

cc JoshRosen

Author: Davies Liu <[email protected]>

Closes #4923 from davies/fix_collect and squashes the following commits:

d730286 [Davies Liu] address comments
24c92a4 [Davies Liu] fix style
ba54614 [Davies Liu] use socket to transfer data from JVM
9517c8f [Davies Liu] fix memory leak in collect()

Author: Davies Liu

core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala

@@ -19,26 +19,27 @@ package org.apache.spark.api.python
 
 import java.io._
 import java.net._
-import java.util.{List => JList, ArrayList => JArrayList, Map => JMap, UUID, Collections}
-
-import org.apache.spark.input.PortableDataStream
+import java.util.{Collections, ArrayList => JArrayList, List => JList, Map => JMap}
 
 import scala.collection.JavaConversions._
 import scala.collection.mutable
 import scala.language.existentials
 
 import com.google.common.base.Charsets.UTF_8
-
 import org.apache.hadoop.conf.Configuration
 import org.apache.hadoop.io.compress.CompressionCodec
-import org.apache.hadoop.mapred.{InputFormat, OutputFormat, JobConf}
+import org.apache.hadoop.mapred.{InputFormat, JobConf, OutputFormat}
 import org.apache.hadoop.mapreduce.{InputFormat => NewInputFormat, OutputFormat => NewOutputFormat}
+
 import org.apache.spark._
-import org.apache.spark.api.java.{JavaSparkContext, JavaPairRDD, JavaRDD}
+import org.apache.spark.api.java.{JavaPairRDD, JavaRDD, JavaSparkContext}
 import org.apache.spark.broadcast.Broadcast
+import org.apache.spark.input.PortableDataStream
 import org.apache.spark.rdd.RDD
 import org.apache.spark.util.Utils
 
+import scala.util.control.NonFatal
+
 private[spark] class PythonRDD(
     @transient parent: RDD[_],
     command: Array[Byte],
@@ -341,21 +342,33 @@ private[spark] object PythonRDD extends Logging {
   /**
    * Adapter for calling SparkContext#runJob from Python.
    *
-   * This method will return an iterator of an array that contains all elements in the RDD
+   * This method will serve an iterator of an array that contains all elements in the RDD
    * (effectively a collect()), but allows you to run on a certain subset of partitions,
    * or to enable local execution.
+   *
+   * @return the port number of a local socket which serves the data collected from this job.
    */
   def runJob(
       sc: SparkContext,
       rdd: JavaRDD[Array[Byte]],
       partitions: JArrayList[Int],
-      allowLocal: Boolean): Iterator[Array[Byte]] = {
+      allowLocal: Boolean): Int = {
     type ByteArray = Array[Byte]
     type UnrolledPartition = Array[ByteArray]
     val allPartitions: Array[UnrolledPartition] =
       sc.runJob(rdd, (x: Iterator[ByteArray]) => x.toArray, partitions, allowLocal)
     val flattenedPartition: UnrolledPartition = Array.concat(allPartitions: _*)
-    flattenedPartition.iterator
+    serveIterator(flattenedPartition.iterator,
+      s"serve RDD ${rdd.id} with partitions ${partitions.mkString(",")}")
+  }
+
+  /**
+   * A helper function to collect an RDD as an iterator, then serve it via socket.
+   *
+   * @return the port number of a local socket which serves the data collected from this job.
+   */
+  def collectAndServe[T](rdd: RDD[T]): Int = {
+    serveIterator(rdd.collect().iterator, s"serve RDD ${rdd.id}")
   }
 
   def readRDDFromFile(sc: JavaSparkContext, filename: String, parallelism: Int):
@@ -575,15 +588,44 @@ private[spark] object PythonRDD extends Logging {
     dataOut.write(bytes)
   }
 
-  def writeToFile[T](items: java.util.Iterator[T], filename: String) {
-    import scala.collection.JavaConverters._
-    writeToFile(items.asScala, filename)
-  }
+  /**
+   * Create a socket server and a background thread to serve the data in `items`,
+   *
+   * The socket server can only accept one connection, or close if no connection
+   * in 3 seconds.
+   *
+   * Once a connection comes in, it tries to serialize all the data in `items`
+   * and send them into this connection.
+   *
+   * The thread will terminate after all the data are sent or any exceptions happen.
+   */
+  private def serveIterator[T](items: Iterator[T], threadName: String): Int = {
+    val serverSocket = new ServerSocket(0, 1)
+    serverSocket.setReuseAddress(true)
+    // Close the socket if no connection in 3 seconds
+    serverSocket.setSoTimeout(3000)
+
+    new Thread(threadName) {
+      setDaemon(true)
+      override def run() {
+        try {
+          val sock = serverSocket.accept()
+          val out = new DataOutputStream(new BufferedOutputStream(sock.getOutputStream))
+          try {
+            writeIteratorToStream(items, out)
+          } finally {
+            out.close()
+          }
+        } catch {
+          case NonFatal(e) =>
+            logError(s"Error while sending iterator", e)
+        } finally {
+          serverSocket.close()
+        }
+      }
+    }.start()
 
-  def writeToFile[T](items: Iterator[T], filename: String) {
-    val file = new DataOutputStream(new FileOutputStream(filename))
-    writeIteratorToStream(items, file)
-    file.close()
+    serverSocket.getLocalPort
   }
 
   private def getMergedConf(confAsMap: java.util.HashMap[String, String],

python/pyspark/context.py

@@ -21,6 +21,8 @@
 from threading import Lock
 from tempfile import NamedTemporaryFile
 
+from py4j.java_collections import ListConverter
+
 from pyspark import accumulators
 from pyspark.accumulators import Accumulator
 from pyspark.broadcast import Broadcast
@@ -30,13 +32,11 @@
 from pyspark.serializers import PickleSerializer, BatchedSerializer, UTF8Deserializer, \
     PairDeserializer, AutoBatchedSerializer, NoOpSerializer
 from pyspark.storagelevel import StorageLevel
-from pyspark.rdd import RDD
+from pyspark.rdd import RDD, _load_from_socket
 from pyspark.traceback_utils import CallSite, first_spark_call
 from pyspark.status import StatusTracker
 from pyspark.profiler import ProfilerCollector, BasicProfiler
 
-from py4j.java_collections import ListConverter
-
 
 __all__ = ['SparkContext']
 
@@ -59,7 +59,6 @@ class SparkContext(object):
 
     _gateway = None
     _jvm = None
-    _writeToFile = None
     _next_accum_id = 0
     _active_spark_context = None
     _lock = Lock()
@@ -221,7 +220,6 @@ def _ensure_initialized(cls, instance=None, gateway=None):
             if not SparkContext._gateway:
                 SparkContext._gateway = gateway or launch_gateway()
                 SparkContext._jvm = SparkContext._gateway.jvm
-                SparkContext._writeToFile = SparkContext._jvm.PythonRDD.writeToFile
 
             if instance:
                 if (SparkContext._active_spark_context and
@@ -840,8 +838,9 @@ def runJob(self, rdd, partitionFunc, partitions=None, allowLocal=False):
         # by runJob() in order to avoid having to pass a Python lambda into
         # SparkContext#runJob.
         mappedRDD = rdd.mapPartitions(partitionFunc)
-        it = self._jvm.PythonRDD.runJob(self._jsc.sc(), mappedRDD._jrdd, javaPartitions, allowLocal)
-        return list(mappedRDD._collect_iterator_through_file(it))
+        port = self._jvm.PythonRDD.runJob(self._jsc.sc(), mappedRDD._jrdd, javaPartitions,
+                                          allowLocal)
+        return list(_load_from_socket(port, mappedRDD._jrdd_deserializer))
 
     def show_profiles(self):
         """ Print the profile stats to stdout """

python/pyspark/rdd.py

@@ -19,7 +19,6 @@
 from collections import defaultdict
 from itertools import chain, ifilter, imap
 import operator
-import os
 import sys
 import shlex
 from subprocess import Popen, PIPE
@@ -29,6 +28,7 @@
 import heapq
 import bisect
 import random
+import socket
 from math import sqrt, log, isinf, isnan, pow, ceil
 
 from pyspark.serializers import NoOpSerializer, CartesianDeserializer, \
@@ -111,6 +111,17 @@ def _parse_memory(s):
     return int(float(s[:-1]) * units[s[-1].lower()])
 
 
+def _load_from_socket(port, serializer):
+    sock = socket.socket()
+    try:
+        sock.connect(("localhost", port))
+        rf = sock.makefile("rb", 65536)
+        for item in serializer.load_stream(rf):
+            yield item
+    finally:
+        sock.close()
+
+
 class Partitioner(object):
     def __init__(self, numPartitions, partitionFunc):
         self.numPartitions = numPartitions
@@ -698,21 +709,8 @@ def collect(self):
         Return a list that contains all of the elements in this RDD.
         """
         with SCCallSiteSync(self.context) as css:
-            bytesInJava = self._jrdd.collect().iterator()
-        return list(self._collect_iterator_through_file(bytesInJava))
-
-    def _collect_iterator_through_file(self, iterator):
-        # Transferring lots of data through Py4J can be slow because
-        # socket.readline() is inefficient.  Instead, we'll dump the data to a
-        # file and read it back.
-        tempFile = NamedTemporaryFile(delete=False, dir=self.ctx._temp_dir)
-        tempFile.close()
-        self.ctx._writeToFile(iterator, tempFile.name)
-        # Read the data into Python and deserialize it:
-        with open(tempFile.name, 'rb') as tempFile:
-            for item in self._jrdd_deserializer.load_stream(tempFile):
-                yield item
-        os.unlink(tempFile.name)
+            port = self.ctx._jvm.PythonRDD.collectAndServe(self._jrdd.rdd())
+        return list(_load_from_socket(port, self._jrdd_deserializer))
 
     def reduce(self, f):
         """

python/pyspark/sql/dataframe.py

@@ -19,13 +19,11 @@
 import itertools
 import warnings
 import random
-import os
-from tempfile import NamedTemporaryFile
 
 from py4j.java_collections import ListConverter, MapConverter
 
 from pyspark.context import SparkContext
-from pyspark.rdd import RDD
+from pyspark.rdd import RDD, _load_from_socket
 from pyspark.serializers import BatchedSerializer, PickleSerializer, UTF8Deserializer
 from pyspark.storagelevel import StorageLevel
 from pyspark.traceback_utils import SCCallSiteSync
@@ -310,14 +308,8 @@ def collect(self):
         [Row(age=2, name=u'Alice'), Row(age=5, name=u'Bob')]
         """
         with SCCallSiteSync(self._sc) as css:
-            bytesInJava = self._jdf.javaToPython().collect().iterator()
-        tempFile = NamedTemporaryFile(delete=False, dir=self._sc._temp_dir)
-        tempFile.close()
-        self._sc._writeToFile(bytesInJava, tempFile.name)
-        # Read the data into Python and deserialize it:
-        with open(tempFile.name, 'rb') as tempFile:
-            rs = list(BatchedSerializer(PickleSerializer()).load_stream(tempFile))
-        os.unlink(tempFile.name)
+            port = self._sc._jvm.PythonRDD.collectAndServe(self._jdf.javaToPython().rdd())
+        rs = list(_load_from_socket(port, BatchedSerializer(PickleSerializer())))
         cls = _create_cls(self.schema)
         return [cls(r) for r in rs]