[SPARK-3721] [PySpark] broadcast objects larger than 2G

This patch will bring support for broadcasting objects larger than 2G.

pickle, zlib, FrameSerializer and Array[Byte] all can not support objects larger than 2G, so this patch introduce LargeObjectSerializer to serialize broadcast objects, the object will be serialized and compressed into small chunks, it also change the type of Broadcast[Array[Byte]]] into Broadcast[Array[Array[Byte]]]].

Testing for support broadcast objects larger than 2G is slow and memory hungry, so this is tested manually, could be added into SparkPerf.

Author: Davies Liu <davies@databricks.com>
Author: Davies Liu <davies.liu@gmail.com>

Closes apache#2659 from davies/huge and squashes the following commits:

7b57a14 [Davies Liu] add more tests for broadcast
28acff9 [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
a2f6a02 [Davies Liu] bug fix
4820613 [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
5875c73 [Davies Liu] address comments
10a349b [Davies Liu] address comments
0c33016 [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
6182c8f [Davies Liu] Merge branch 'master' into huge
d94b68f [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
2514848 [Davies Liu] address comments
fda395b [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
1c2d928 [Davies Liu] fix scala style
091b107 [Davies Liu] broadcast objects larger than 2G

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

@@ -47,7 +47,7 @@ private[spark] class PythonRDD(
     pythonIncludes: JList[String],
     preservePartitoning: Boolean,
     pythonExec: String,
-    broadcastVars: JList[Broadcast[Array[Byte]]],
+    broadcastVars: JList[Broadcast[Array[Array[Byte]]]],
     accumulator: Accumulator[JList[Array[Byte]]])
   extends RDD[Array[Byte]](parent) {
 
@@ -230,8 +230,8 @@ private[spark] class PythonRDD(
           if (!oldBids.contains(broadcast.id)) {
             // send new broadcast
             dataOut.writeLong(broadcast.id)
-            dataOut.writeInt(broadcast.value.length)
-            dataOut.write(broadcast.value)
+            dataOut.writeLong(broadcast.value.map(_.length.toLong).sum)
+            broadcast.value.foreach(dataOut.write)
             oldBids.add(broadcast.id)
           }
         }
@@ -368,16 +368,24 @@ private[spark] object PythonRDD extends Logging {
     }
   }
 
-  def readBroadcastFromFile(sc: JavaSparkContext, filename: String): Broadcast[Array[Byte]] = {
+  def readBroadcastFromFile(
+      sc: JavaSparkContext,
+      filename: String): Broadcast[Array[Array[Byte]]] = {
+    val size = new File(filename).length()
     val file = new DataInputStream(new FileInputStream(filename))
+    val blockSize = 1 << 20
+    val n = ((size + blockSize - 1) / blockSize).toInt
+    val obj = new Array[Array[Byte]](n)
     try {
-      val length = file.readInt()
-      val obj = new Array[Byte](length)
-      file.readFully(obj)
-      sc.broadcast(obj)
+      for (i <- 0 until n) {
+        val length = if (i < (n - 1)) blockSize else (size % blockSize).toInt
+        obj(i) = new Array[Byte](length)
+        file.readFully(obj(i))
+      }
     } finally {
       file.close()
     }
+    sc.broadcast(obj)
   }
 
   def writeIteratorToStream[T](iter: Iterator[T], dataOut: DataOutputStream) {

python/pyspark/broadcast.py

@@ -29,7 +29,7 @@
 """
 import os
 
-from pyspark.serializers import CompressedSerializer, PickleSerializer
+from pyspark.serializers import LargeObjectSerializer
 
 
 __all__ = ['Broadcast']
@@ -73,7 +73,7 @@ def value(self):
         """ Return the broadcasted value
         """
         if not hasattr(self, "_value") and self.path is not None:
-            ser = CompressedSerializer(PickleSerializer())
+            ser = LargeObjectSerializer()
             self._value = ser.load_stream(open(self.path)).next()
         return self._value
 

python/pyspark/context.py

@@ -29,7 +29,7 @@
 from pyspark.files import SparkFiles
 from pyspark.java_gateway import launch_gateway
 from pyspark.serializers import PickleSerializer, BatchedSerializer, UTF8Deserializer, \
-    PairDeserializer, CompressedSerializer, AutoBatchedSerializer, NoOpSerializer
+    PairDeserializer, AutoBatchedSerializer, NoOpSerializer, LargeObjectSerializer
 from pyspark.storagelevel import StorageLevel
 from pyspark.rdd import RDD
 from pyspark.traceback_utils import CallSite, first_spark_call
@@ -624,7 +624,8 @@ def broadcast(self, value):
         object for reading it in distributed functions. The variable will
         be sent to each cluster only once.
         """
-        ser = CompressedSerializer(PickleSerializer())
+        ser = LargeObjectSerializer()
+
         # pass large object by py4j is very slow and need much memory
         tempFile = NamedTemporaryFile(delete=False, dir=self._temp_dir)
         ser.dump_stream([value], tempFile)

python/pyspark/serializers.py

@@ -133,6 +133,8 @@ def load_stream(self, stream):
 
     def _write_with_length(self, obj, stream):
         serialized = self.dumps(obj)
+        if len(serialized) > (1 << 31):
+            raise ValueError("can not serialize object larger than 2G")
         write_int(len(serialized), stream)
         if self._only_write_strings:
             stream.write(str(serialized))
@@ -446,20 +448,184 @@ def loads(self, obj):
             raise ValueError("invalid sevialization type: %s" % _type)
 
 
-class CompressedSerializer(FramedSerializer):
+class SizeLimitedStream(object):
     """
-    Compress the serialized data
+    Read at most `limit` bytes from underlying stream
+
+    >>> from StringIO import StringIO
+    >>> io = StringIO()
+    >>> io.write("Hello world")
+    >>> io.seek(0)
+    >>> lio = SizeLimitedStream(io, 5)
+    >>> lio.read()
+    'Hello'
+    """
+    def __init__(self, stream, limit):
+        self.stream = stream
+        self.limit = limit
+
+    def read(self, n=0):
+        if n > self.limit or n == 0:
+            n = self.limit
+        buf = self.stream.read(n)
+        self.limit -= len(buf)
+        return buf
+
+
+class CompressedStream(object):
+    """
+    Compress the data using zlib
+
+    >>> from StringIO import StringIO
+    >>> io = StringIO()
+    >>> wio = CompressedStream(io, 'w')
+    >>> wio.write("Hello world")
+    >>> wio.flush()
+    >>> io.seek(0)
+    >>> rio = CompressedStream(io, 'r')
+    >>> rio.read()
+    'Hello world'
+    >>> rio.read()
+    ''
+    """
+    MAX_BATCH = 1 << 20  # 1MB
+
+    def __init__(self, stream, mode='w', level=1):
+        self.stream = stream
+        self.mode = mode
+        if mode == 'w':
+            self.compresser = zlib.compressobj(level)
+        elif mode == 'r':
+            self.decompresser = zlib.decompressobj()
+            self.buf = ''
+        else:
+            raise ValueError("can only support mode 'w' or 'r' ")
+
+    def write(self, buf):
+        assert self.mode == 'w', "It's not opened for write"
+        if len(buf) > self.MAX_BATCH:
+            # zlib can not compress string larger than 2G
+            batches = len(buf) / self.MAX_BATCH + 1  # last one may be empty
+            for i in xrange(batches):
+                self.write(buf[i * self.MAX_BATCH:(i + 1) * self.MAX_BATCH])
+        else:
+            compressed = self.compresser.compress(buf)
+            self.stream.write(compressed)
+
+    def flush(self, mode=zlib.Z_FULL_FLUSH):
+        if self.mode == 'w':
+            d = self.compresser.flush(mode)
+            self.stream.write(d)
+            self.stream.flush()
+
+    def close(self):
+        if self.mode == 'w':
+            self.flush(zlib.Z_FINISH)
+            self.stream.close()
+
+    def read(self, size=0):
+        assert self.mode == 'r', "It's not opened for read"
+        if not size:
+            data = self.stream.read()
+            result = self.decompresser.decompress(data)
+            last = self.decompresser.flush()
+            return self.buf + result + last
+
+        # fast path for small read()
+        if size <= len(self.buf):
+            result = self.buf[:size]
+            self.buf = self.buf[size:]
+            return result
+
+        result = [self.buf]
+        size -= len(self.buf)
+        self.buf = ''
+        while size:
+            need = min(size, self.MAX_BATCH)
+            input = self.stream.read(need)
+            if input:
+                buf = self.decompresser.decompress(input)
+            else:
+                buf = self.decompresser.flush()
+
+            if len(buf) >= size:
+                self.buf = buf[size:]
+                result.append(buf[:size])
+                return ''.join(result)
+
+            size -= len(buf)
+            result.append(buf)
+            if not input:
+                return ''.join(result)
+
+    def readline(self):
+        """
+        This is needed for pickle, but not used in protocol 2
+        """
+        line = []
+        b = self.read(1)
+        while b and b != '\n':
+            line.append(b)
+            b = self.read(1)
+        line.append(b)
+        return ''.join(line)
+
+
+class LargeObjectSerializer(Serializer):
+    """
+    Serialize large object which could be larger than 2G
+
+    It uses cPickle to serialize the objects
     """
+    def dump_stream(self, iterator, stream):
+        stream = CompressedStream(stream, 'w')
+        for value in iterator:
+            if isinstance(value, basestring):
+                if isinstance(value, unicode):
+                    stream.write('U')
+                    value = value.encode("utf-8")
+                else:
+                    stream.write('S')
+                write_long(len(value), stream)
+                stream.write(value)
+            else:
+                stream.write('P')
+                cPickle.dump(value, stream, 2)
+        stream.flush()
 
+    def load_stream(self, stream):
+        stream = CompressedStream(stream, 'r')
+        while True:
+            type = stream.read(1)
+            if not type:
+                return
+            if type in ('S', 'U'):
+                length = read_long(stream)
+                value = stream.read(length)
+                if type == 'U':
+                    value = value.decode('utf-8')
+                yield value
+            elif type == 'P':
+                yield cPickle.load(stream)
+            else:
+                raise ValueError("unknown type: %s" % type)
+
+
+class CompressedSerializer(Serializer):
+    """
+    Compress the serialized data
+    """
     def __init__(self, serializer):
-        FramedSerializer.__init__(self)
         self.serializer = serializer
 
-    def dumps(self, obj):
-        return zlib.compress(self.serializer.dumps(obj), 1)
+    def load_stream(self, stream):
+        stream = CompressedStream(stream, "r")
+        return self.serializer.load_stream(stream)
 
-    def loads(self, obj):
-        return self.serializer.loads(zlib.decompress(obj))
+    def dump_stream(self, iterator, stream):
+        stream = CompressedStream(stream, "w")
+        self.serializer.dump_stream(iterator, stream)
+        stream.flush()
 
 
 class UTF8Deserializer(Serializer):
@@ -517,3 +683,8 @@ def write_int(value, stream):
 def write_with_length(obj, stream):
     write_int(len(obj), stream)
     stream.write(obj)
+
+
+if __name__ == '__main__':
+    import doctest
+    doctest.testmod()

python/pyspark/tests.py

@@ -32,6 +32,7 @@
 import zipfile
 import random
 import threading
+import hashlib
 
 if sys.version_info[:2] <= (2, 6):
     try:
@@ -47,7 +48,7 @@
 from pyspark.context import SparkContext
 from pyspark.files import SparkFiles
 from pyspark.serializers import read_int, BatchedSerializer, MarshalSerializer, PickleSerializer, \
-    CloudPickleSerializer
+    CloudPickleSerializer, SizeLimitedStream, CompressedSerializer, LargeObjectSerializer
 from pyspark.shuffle import Aggregator, InMemoryMerger, ExternalMerger, ExternalSorter
 from pyspark.sql import SQLContext, IntegerType, Row, ArrayType, StructType, StructField, \
     UserDefinedType, DoubleType
@@ -236,6 +237,27 @@ def foo():
         self.assertTrue("exit" in foo.func_code.co_names)
         ser.dumps(foo)
 
+    def _test_serializer(self, ser):
+        from StringIO import StringIO
+        io = StringIO()
+        ser.dump_stream(["abc", u"123", range(5)], io)
+        io.seek(0)
+        self.assertEqual(["abc", u"123", range(5)], list(ser.load_stream(io)))
+        size = io.tell()
+        ser.dump_stream(range(1000), io)
+        io.seek(0)
+        first = SizeLimitedStream(io, size)
+        self.assertEqual(["abc", u"123", range(5)], list(ser.load_stream(first)))
+        self.assertEqual(range(1000), list(ser.load_stream(io)))
+
+    def test_compressed_serializer(self):
+        ser = CompressedSerializer(PickleSerializer())
+        self._test_serializer(ser)
+
+    def test_large_object_serializer(self):
+        ser = LargeObjectSerializer()
+        self._test_serializer(ser)
+
 
 class PySparkTestCase(unittest.TestCase):
 
@@ -440,7 +462,7 @@ def test_sampling_default_seed(self):
         subset = data.takeSample(False, 10)
         self.assertEqual(len(subset), 10)
 
-    def testAggregateByKey(self):
+    def test_aggregate_by_key(self):
         data = self.sc.parallelize([(1, 1), (1, 1), (3, 2), (5, 1), (5, 3)], 2)
 
         def seqOp(x, y):
@@ -478,6 +500,32 @@ def test_large_broadcast(self):
         m = self.sc.parallelize(range(1), 1).map(lambda x: len(bdata.value)).sum()
         self.assertEquals(N, m)
 
+    def test_multiple_broadcasts(self):
+        N = 1 << 21
+        b1 = self.sc.broadcast(set(range(N)))  # multiple blocks in JVM
+        r = range(1 << 15)
+        random.shuffle(r)
+        s = str(r)
+        checksum = hashlib.md5(s).hexdigest()
+        b2 = self.sc.broadcast(s)
+        r = list(set(self.sc.parallelize(range(10), 10).map(
+            lambda x: (len(b1.value), hashlib.md5(b2.value).hexdigest())).collect()))
+        self.assertEqual(1, len(r))
+        size, csum = r[0]
+        self.assertEqual(N, size)
+        self.assertEqual(checksum, csum)
+
+        random.shuffle(r)
+        s = str(r)
+        checksum = hashlib.md5(s).hexdigest()
+        b2 = self.sc.broadcast(s)
+        r = list(set(self.sc.parallelize(range(10), 10).map(
+            lambda x: (len(b1.value), hashlib.md5(b2.value).hexdigest())).collect()))
+        self.assertEqual(1, len(r))
+        size, csum = r[0]
+        self.assertEqual(N, size)
+        self.assertEqual(checksum, csum)
+
     def test_large_closure(self):
         N = 1000000
         data = [float(i) for i in xrange(N)]