recover from checkpoint

Author: davies

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

@@ -42,7 +42,7 @@ import org.apache.spark.rdd.RDD
 import org.apache.spark.util.Utils
 
 private[spark] class PythonRDD(
-    parent: RDD[_],
+    @transient parent: RDD[_],
     command: Array[Byte],
     envVars: JMap[String, String],
     pythonIncludes: JList[String],
@@ -61,9 +61,9 @@ private[spark] class PythonRDD(
   val bufferSize = conf.getInt("spark.buffer.size", 65536)
   val reuse_worker = conf.getBoolean("spark.python.worker.reuse", true)
 
-  override def getPartitions = parent.partitions
+  override def getPartitions = firstParent.partitions
 
-  override val partitioner = if (preservePartitoning) parent.partitioner else None
+  override val partitioner = if (preservePartitoning) firstParent.partitioner else None
 
   override def compute(split: Partition, context: TaskContext): Iterator[Array[Byte]] = {
     val startTime = System.currentTimeMillis
@@ -241,7 +241,7 @@ private[spark] class PythonRDD(
         dataOut.writeInt(command.length)
         dataOut.write(command)
         // Data values
-        PythonRDD.writeIteratorToStream(parent.iterator(split, context), dataOut)
+        PythonRDD.writeIteratorToStream(firstParent.iterator(split, context), dataOut)
         dataOut.writeInt(SpecialLengths.END_OF_DATA_SECTION)
         dataOut.flush()
       } catch {

core/src/main/scala/org/apache/spark/rdd/ParallelCollectionRDD.scala

@@ -84,7 +84,7 @@ private[spark] class ParallelCollectionPartition[T: ClassTag](
 
 private[spark] class ParallelCollectionRDD[T: ClassTag](
     @transient sc: SparkContext,
-    @transient data: Seq[T],
+    data: Seq[T],
     numSlices: Int,
     locationPrefs: Map[Int, Seq[String]])
     extends RDD[T](sc, Nil) {

core/src/main/scala/org/apache/spark/rdd/RDD.scala

@@ -82,6 +82,14 @@ abstract class RDD[T: ClassTag](
   def this(@transient oneParent: RDD[_]) =
     this(oneParent.context , List(new OneToOneDependency(oneParent)))
 
+  // setContext after loading from checkpointing
+  private[spark] def setContext(s: SparkContext) = {
+    if (sc != null && sc != s) {
+      throw new SparkException("Context is already set in " + this + ", cannot set it again")
+    }
+    sc = s
+  }
+
   private[spark] def conf = sc.conf
   // =======================================================================
   // Methods that should be implemented by subclasses of RDD

python/pyspark/context.py

@@ -68,7 +68,7 @@ class SparkContext(object):
 
     def __init__(self, master=None, appName=None, sparkHome=None, pyFiles=None,
                  environment=None, batchSize=1024, serializer=PickleSerializer(), conf=None,
-                 gateway=None):
+                 gateway=None, jsc=None):
         """
         Create a new SparkContext. At least the master and app name should be set,
         either through the named parameters here or through C{conf}.
@@ -103,14 +103,14 @@ def __init__(self, master=None, appName=None, sparkHome=None, pyFiles=None,
         SparkContext._ensure_initialized(self, gateway=gateway)
         try:
             self._do_init(master, appName, sparkHome, pyFiles, environment, batchSize, serializer,
-                          conf)
+                          conf, jsc)
         except:
             # If an error occurs, clean up in order to allow future SparkContext creation:
             self.stop()
             raise
 
     def _do_init(self, master, appName, sparkHome, pyFiles, environment, batchSize, serializer,
-                 conf):
+                 conf, jsc):
         self.environment = environment or {}
         self._conf = conf or SparkConf(_jvm=self._jvm)
         self._batchSize = batchSize  # -1 represents an unlimited batch size
@@ -151,7 +151,7 @@ def _do_init(self, master, appName, sparkHome, pyFiles, environment, batchSize,
                 self.environment[varName] = v
 
         # Create the Java SparkContext through Py4J
-        self._jsc = self._initialize_context(self._conf._jconf)
+        self._jsc = jsc or self._initialize_context(self._conf._jconf)
 
         # Create a single Accumulator in Java that we'll send all our updates through;
         # they will be passed back to us through a TCP server

python/pyspark/streaming/context.py

@@ -14,11 +14,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
+import os
+import sys
 
 from py4j.java_collections import ListConverter
 from py4j.java_gateway import java_import
 
-from pyspark import RDD
+from pyspark import RDD, SparkConf
 from pyspark.serializers import UTF8Deserializer, CloudPickleSerializer
 from pyspark.context import SparkContext
 from pyspark.storagelevel import StorageLevel
@@ -75,41 +77,81 @@ class StreamingContext(object):
     respectively. `context.awaitTransformation()` allows the current thread
     to wait for the termination of the context by `stop()` or by an exception.
     """
+    _transformerSerializer = None
 
-    def __init__(self, sparkContext, duration):
+    def __init__(self, sparkContext, duration=None, jssc=None):
         """
         Create a new StreamingContext.
 
         @param sparkContext: L{SparkContext} object.
         @param duration: number of seconds.
         """
+
         self._sc = sparkContext
         self._jvm = self._sc._jvm
-        self._start_callback_server()
-        self._jssc = self._initialize_context(self._sc, duration)
+        self._jssc = jssc or self._initialize_context(self._sc, duration)
+
+    def _initialize_context(self, sc, duration):
+        self._ensure_initialized()
+        return self._jvm.JavaStreamingContext(sc._jsc, self._jduration(duration))
+
+    def _jduration(self, seconds):
+        """
+        Create Duration object given number of seconds
+        """
+        return self._jvm.Duration(int(seconds * 1000))
 
-    def _start_callback_server(self):
-        gw = self._sc._gateway
+    @classmethod
+    def _ensure_initialized(cls):
+        SparkContext._ensure_initialized()
+        gw = SparkContext._gateway
+        # start callback server
         # getattr will fallback to JVM
         if "_callback_server" not in gw.__dict__:
             _daemonize_callback_server()
             gw._start_callback_server(gw._python_proxy_port)
-            gw._python_proxy_port = gw._callback_server.port  # update port with real port
 
-    def _initialize_context(self, sc, duration):
-        java_import(self._jvm, "org.apache.spark.streaming.*")
-        java_import(self._jvm, "org.apache.spark.streaming.api.java.*")
-        java_import(self._jvm, "org.apache.spark.streaming.api.python.*")
+        java_import(gw.jvm, "org.apache.spark.streaming.*")
+        java_import(gw.jvm, "org.apache.spark.streaming.api.java.*")
+        java_import(gw.jvm, "org.apache.spark.streaming.api.python.*")
         # register serializer for RDDFunction
-        ser = RDDFunctionSerializer(self._sc, CloudPickleSerializer())
-        self._jvm.PythonDStream.registerSerializer(ser)
-        return self._jvm.JavaStreamingContext(sc._jsc, self._jduration(duration))
+        # it happens before creating SparkContext when loading from checkpointing
+        cls._transformerSerializer = RDDFunctionSerializer(SparkContext._active_spark_context,
+                                                           CloudPickleSerializer(), gw)
+        gw.jvm.PythonDStream.registerSerializer(cls._transformerSerializer)
 
-    def _jduration(self, seconds):
+    @classmethod
+    def getOrCreate(cls, path, setupFunc):
         """
-        Create Duration object given number of seconds
+        Get the StreamingContext from checkpoint file at `path`, or setup
+        it by `setupFunc`.
+
+        :param path: directory of checkpoint
+        :param setupFunc: a function used to create StreamingContext and
+                          setup DStreams.
+        :return: a StreamingContext
         """
-        return self._jvm.Duration(int(seconds * 1000))
+        if not os.path.exists(path) or not os.path.isdir(path) or not os.listdir(path):
+            ssc = setupFunc()
+            ssc.checkpoint(path)
+            return ssc
+
+        cls._ensure_initialized()
+        gw = SparkContext._gateway
+
+        try:
+            jssc = gw.jvm.JavaStreamingContext(path)
+        except Exception:
+            print >>sys.stderr, "failed to load StreamingContext from checkpoint"
+            raise
+
+        jsc = jssc.sparkContext()
+        conf = SparkConf(_jconf=jsc.getConf())
+        sc = SparkContext(conf=conf, gateway=gw, jsc=jsc)
+        # update ctx in serializer
+        SparkContext._active_spark_context = sc
+        cls._transformerSerializer.ctx = sc
+        return StreamingContext(sc, None, jssc)
 
     @property
     def sparkContext(self):

python/pyspark/streaming/tests.py

@@ -493,5 +493,38 @@ def func(rdds):
         self.assertEqual([2, 3, 1], self._take(dstream, 3))
 
 
+class TestCheckpoint(PySparkStreamingTestCase):
+
+    def setUp(self):
+        pass
+
+    def tearDown(self):
+        pass
+
+    def test_get_or_create(self):
+        result = [0]
+
+        def setup():
+            conf = SparkConf().set("spark.default.parallelism", 1)
+            sc = SparkContext(conf=conf)
+            ssc = StreamingContext(sc, .2)
+            rdd = sc.parallelize(range(10), 1)
+            dstream = ssc.queueStream([rdd], default=rdd)
+            result[0] = self._collect(dstream.countByWindow(1, .2))
+            return ssc
+        tmpd = tempfile.mkdtemp("test_streaming_cps")
+        ssc = StreamingContext.getOrCreate(tmpd, setup)
+        ssc.start()
+        ssc.awaitTermination(4)
+        ssc.stop()
+        expected = [[i * 10 + 10] for i in range(5)] + [[50]] * 5
+        self.assertEqual(expected, result[0][:10])
+
+        ssc = StreamingContext.getOrCreate(tmpd, setup)
+        ssc.start()
+        ssc.awaitTermination(2)
+        ssc.stop()
+
+
 if __name__ == "__main__":
     unittest.main()

python/pyspark/streaming/util.py

@@ -18,24 +18,31 @@
 from datetime import datetime
 import traceback
 
-from pyspark.rdd import RDD
+from pyspark import SparkContext, RDD
 
 
 class RDDFunction(object):
     """
     This class is for py4j callback.
     """
+    _emptyRDD = None
+
     def __init__(self, ctx, func, *deserializers):
         self.ctx = ctx
         self.func = func
         self.deserializers = deserializers
-        emptyRDD = getattr(self.ctx, "_emptyRDD", None)
-        if emptyRDD is None:
-            self.ctx._emptyRDD = emptyRDD = self.ctx.parallelize([]).cache()
-        self.emptyRDD = emptyRDD
+
+    @property
+    def emptyRDD(self):
+        if self._emptyRDD is None and self.ctx:
+            self._emptyRDD = self.ctx.parallelize([]).cache()
+        return self._emptyRDD
 
     def call(self, milliseconds, jrdds):
         try:
+            if self.ctx is None:
+                self.ctx = SparkContext._active_spark_context
+
             # extend deserializers with the first one
             sers = self.deserializers
             if len(sers) < len(jrdds):
@@ -51,20 +58,21 @@ def call(self, milliseconds, jrdds):
             traceback.print_exc()
 
     def __repr__(self):
-        return "RDDFunction(%s)" % (str(self.func))
+        return "RDDFunction(%s)" % self.func
 
     class Java:
         implements = ['org.apache.spark.streaming.api.python.PythonRDDFunction']
 
 
 class RDDFunctionSerializer(object):
-    def __init__(self, ctx, serializer):
+    def __init__(self, ctx, serializer, gateway=None):
         self.ctx = ctx
         self.serializer = serializer
+        self.gateway = gateway or self.ctx._gateway
 
     def dumps(self, id):
         try:
-            func = self.ctx._gateway.gateway_property.pool[id]
+            func = self.gateway.gateway_property.pool[id]
             return bytearray(self.serializer.dumps((func.func, func.deserializers)))
         except Exception:
             traceback.print_exc()

streaming/src/main/scala/org/apache/spark/streaming/api/python/PythonDStream.scala

@@ -77,7 +77,7 @@ private[python] class RDDFunction(@transient var pfunc: PythonRDDFunction)
 }
 
 /**
- * Inferface for Python Serializer to serialize PythonRDDFunction
+ * Interface for Python Serializer to serialize PythonRDDFunction
  */
 private[python] trait PythonRDDFunctionSerializer {
   def dumps(id: String): Array[Byte]  //
@@ -91,9 +91,9 @@ private[python] class RDDFunctionSerializer(pser: PythonRDDFunctionSerializer) {
   def serialize(func: PythonRDDFunction): Array[Byte] = {
     // get the id of PythonRDDFunction in py4j
     val h = Proxy.getInvocationHandler(func.asInstanceOf[Proxy])
-    val f = h.getClass().getDeclaredField("id");
-    f.setAccessible(true);
-    val id = f.get(h).asInstanceOf[String];
+    val f = h.getClass().getDeclaredField("id")
+    f.setAccessible(true)
+    val id = f.get(h).asInstanceOf[String]
     pser.dumps(id)
   }
 

streaming/src/main/scala/org/apache/spark/streaming/dstream/QueueInputDStream.scala

@@ -17,6 +17,7 @@
 
 package org.apache.spark.streaming.dstream
 
+import org.apache.spark.SparkException
 import org.apache.spark.rdd.RDD
 import org.apache.spark.rdd.UnionRDD
 import scala.collection.mutable.Queue
@@ -32,6 +33,12 @@ class QueueInputDStream[T: ClassTag](
     defaultRDD: RDD[T]
   ) extends InputDStream[T](ssc) {
 
+  private[streaming] override def setContext(s: StreamingContext) {
+    super.setContext(s)
+    queue.map(_.setContext(s.sparkContext))
+    defaultRDD.setContext(s.sparkContext)
+  }
+
   override def start() { }
 
   override def stop() { }