refactor, combine TransformedRDD, fix reuse PythonRDD, fix union

Author: davies

python/pyspark/streaming/context.py

@@ -15,12 +15,7 @@
 # limitations under the License.
 #
 
-import sys
-from signal import signal, SIGTERM, SIGINT
-import atexit
-import time
-
-from pyspark.serializers import PickleSerializer, BatchedSerializer, UTF8Deserializer
+from pyspark.serializers import UTF8Deserializer
 from pyspark.context import SparkContext
 from pyspark.streaming.dstream import DStream
 from pyspark.streaming.duration import Duration, Seconds

python/pyspark/streaming/dstream.py

@@ -15,21 +15,15 @@
 # limitations under the License.
 #
 
-from collections import defaultdict
 from itertools import chain, ifilter, imap
 import operator
 
 from pyspark import RDD
-from pyspark.serializers import NoOpSerializer,\
-    BatchedSerializer, CloudPickleSerializer, pack_long,\
-    CompressedSerializer
 from pyspark.storagelevel import StorageLevel
-from pyspark.resultiterable import ResultIterable
-from pyspark.streaming.util import rddToFileName, RDDFunction
-from pyspark.rdd import portable_hash, _parse_memory
-from pyspark.traceback_utils import SCCallSiteSync
+from pyspark.streaming.util import rddToFileName, RDDFunction, RDDFunction2
+from pyspark.rdd import portable_hash
+from pyspark.streaming.duration import Seconds
 
-from py4j.java_collections import ListConverter, MapConverter
 
 __all__ = ["DStream"]
 
@@ -42,7 +36,6 @@ def __init__(self, jdstream, ssc, jrdd_deserializer):
         self._jrdd_deserializer = jrdd_deserializer
         self.is_cached = False
         self.is_checkpointed = False
-        self._partitionFunc = None
 
     def context(self):
         """
@@ -159,7 +152,7 @@ def foreachRDD(self, func):
         This is an output operator, so this DStream will be registered as an output
         stream and there materialized.
         """
-        jfunc = RDDFunction(self.ctx, lambda a, b, t: func(a, t), self._jrdd_deserializer)
+        jfunc = RDDFunction(self.ctx, func, self._jrdd_deserializer)
         self.ctx._jvm.PythonForeachDStream(self._jdstream.dstream(), jfunc)
 
     def pyprint(self):
@@ -306,19 +299,19 @@ def get_output(rdd, time):
         return result
 
     def transform(self, func):
-        return TransformedRDD(self, lambda a, b, t: func(a), cache=True)
-
-    def transformWith(self, func, other):
-        return TransformedRDD(self, lambda a, b, t: func(a, b), other)
+        return TransformedRDD(self, lambda a, t: func(a), True)
 
     def transformWithTime(self, func):
-        return TransformedRDD(self, lambda a, b, t: func(a, t))
+        return TransformedRDD(self, func, False)
+
+    def transformWith(self, func, other, keepSerializer=False):
+        return Transformed2RDD(self, lambda a, b, t: func(a, b), other, keepSerializer)
 
     def repartitions(self, numPartitions):
         return self.transform(lambda rdd: rdd.repartition(numPartitions))
 
     def union(self, other):
-        return self.transformWith(lambda a, b: a.union(b), other)
+        return self.transformWith(lambda a, b: a.union(b), other, True)
 
     def cogroup(self, other):
         return self.transformWith(lambda a, b: a.cogroup(b), other)
@@ -329,32 +322,79 @@ def leftOuterJoin(self, other):
     def rightOuterJoin(self, other):
         return self.transformWith(lambda a, b: a.rightOuterJoin(b), other)
 
-    def slice(self, fromTime, toTime):
-        jrdds = self._jdstream.slice(fromTime._jtime, toTime._jtime)
-        # FIXME: serializer
-        return [RDD(jrdd, self.ctx, self.ctx.serializer) for jrdd in jrdds]
+    def _jtime(self, milliseconds):
+        return self.ctx._jvm.Time(milliseconds)
+
+    def slice(self, begin, end):
+        jrdds = self._jdstream.slice(self._jtime(begin), self._jtime(end))
+        return [RDD(jrdd, self.ctx, self._jrdd_deserializer) for jrdd in jrdds]
+
+    def window(self, windowDuration, slideDuration=None):
+        d = Seconds(windowDuration)
+        if slideDuration is None:
+            return DStream(self._jdstream.window(d), self._ssc, self._jrdd_deserializer)
+        s = Seconds(slideDuration)
+        return DStream(self._jdstream.window(d, s), self._ssc, self._jrdd_deserializer)
+
+    def reduceByWindow(self, reduceFunc, inReduceFunc, windowDuration, slideDuration):
+        pass
+
+    def countByWindow(self, window, slide):
+        pass
+
+    def countByValueAndWindow(self, window, slide, numPartitions=None):
+        pass
+
+    def groupByKeyAndWindow(self, window, slide, numPartitions=None):
+        pass
+
+    def reduceByKeyAndWindow(self, reduceFunc, inReduceFunc, window, slide, numPartitions=None):
+        pass
 
     def updateStateByKey(self, updateFunc):
         # FIXME: convert updateFunc to java JFunction2
         jFunc = updateFunc
         return self._jdstream.updateStateByKey(jFunc)
 
 
-# Window Operations
-# TODO: implement window
-# TODO: implement groupByKeyAndWindow
-# TODO: implement reduceByKeyAndWindow
-# TODO: implement countByValueAndWindow
-# TODO: implement countByWindow
-# TODO: implement reduceByWindow
+class TransformedRDD(DStream):
+    def __init__(self, prev, func, reuse=False):
+        ssc = prev._ssc
+        self._ssc = ssc
+        self.ctx = ssc._sc
+        self._jrdd_deserializer = self.ctx.serializer
+        self.is_cached = False
+        self.is_checkpointed = False
+
+        if isinstance(prev, TransformedRDD) and not prev.is_cached and not prev.is_checkpointed:
+            prev_func = prev.func
+            old_func = func
+            func = lambda rdd, t: old_func(prev_func(rdd, t), t)
+            reuse = reuse and prev.reuse
+            prev = prev.prev
 
+        self.prev = prev
+        self.func = func
+        self.reuse = reuse
+        self._jdstream_val = None
 
-class TransformedRDD(DStream):
-    # TODO: better name for cache
-    def __init__(self, prev, func, other=None, cache=False):
-        # TODO: combine transformed RDD
+    @property
+    def _jdstream(self):
+        if self._jdstream_val is not None:
+            return self._jdstream_val
+
+        jfunc = RDDFunction(self.ctx, self.func, self.prev._jrdd_deserializer)
+        jdstream = self.ctx._jvm.PythonTransformedDStream(self.prev._jdstream.dstream(),
+                                                          jfunc, self.reuse).asJavaDStream()
+        self._jdstream_val = jdstream
+        return jdstream
+
+
+class Transformed2RDD(DStream):
+    def __init__(self, prev, func, other, keepSerializer=False):
         ssc = prev._ssc
-        t = RDDFunction(ssc._sc, func, prev._jrdd_deserializer)
-        jdstream = ssc._jvm.PythonTransformedDStream(prev._jdstream.dstream(),
-                                                     other and other._jdstream, t, cache)
-        DStream.__init__(self, jdstream.asJavaDStream(), ssc, ssc._sc.serializer)
+        jfunc = RDDFunction2(ssc._sc, func, prev._jrdd_deserializer)
+        jdstream = ssc._jvm.PythonTransformed2DStream(prev._jdstream.dstream(),
+                                                      other._jdstream.dstream(), jfunc)
+        jrdd_serializer = prev._jrdd_deserializer if keepSerializer else ssc._sc.serializer
+        DStream.__init__(self, jdstream.asJavaDStream(), ssc, jrdd_serializer)

python/pyspark/streaming/tests.py

@@ -213,6 +213,32 @@ def add(a, b):
                     [("a", "11"), ("b", "1"), ("", "111")]]
         self._test_func(input, func, expected, sort=True)
 
+    def test_union(self):
+        input1 = [range(3), range(5), range(1)]
+        input2 = [range(3, 6), range(5, 6), range(1, 6)]
+
+        d1 = self.ssc._makeStream(input1)
+        d2 = self.ssc._makeStream(input2)
+        d = d1.union(d2)
+        result = d.collect()
+        expected = [range(6), range(6), range(6)]
+
+        self.ssc.start()
+        start_time = time.time()
+        # Loop until get the expected the number of the result from the stream.
+        while True:
+            current_time = time.time()
+            # Check time out.
+            if (current_time - start_time) > self.timeout * 2:
+                break
+            # StreamingContext.awaitTermination is not used to wait because
+            # if py4j server is called every 50 milliseconds, it gets an error.
+            time.sleep(0.05)
+            # Check if the output is the same length of expected output.
+            if len(expected) == len(result):
+                break
+        self.assertEqual(expected, result)
+
     def _sort_result_based_on_key(self, outputs):
         """Sort the list base onf first value."""
         for output in outputs:

python/pyspark/streaming/util.py

@@ -28,9 +28,36 @@ def __init__(self, ctx, func, jrdd_deserializer):
         self.func = func
         self.deserializer = jrdd_deserializer
 
-    def call(self, jrdd, jrdd2, milliseconds):
+    def call(self, jrdd, milliseconds):
         try:
             rdd = RDD(jrdd, self.ctx, self.deserializer)
+            r = self.func(rdd, milliseconds)
+            if r:
+                return r._jrdd
+        except:
+            import traceback
+            traceback.print_exc()
+
+    def __repr__(self):
+        return "RDDFunction(%s, %s)" % (str(self.deserializer), str(self.func))
+
+    class Java:
+        implements = ['org.apache.spark.streaming.api.python.PythonRDDFunction']
+
+
+class RDDFunction2(object):
+    """
+    This class is for py4j callback. This class is related with
+    org.apache.spark.streaming.api.python.PythonRDDFunction2.
+    """
+    def __init__(self, ctx, func, jrdd_deserializer):
+        self.ctx = ctx
+        self.func = func
+        self.deserializer = jrdd_deserializer
+
+    def call(self, jrdd, jrdd2, milliseconds):
+        try:
+            rdd = RDD(jrdd, self.ctx, self.deserializer) if jrdd else None
             other = RDD(jrdd2, self.ctx, self.deserializer) if jrdd2 else None
             r = self.func(rdd, other, milliseconds)
             if r:
@@ -43,7 +70,7 @@ def __repr__(self):
         return "RDDFunction(%s, %s)" % (str(self.deserializer), str(self.func))
 
     class Java:
-        implements = ['org.apache.spark.streaming.api.python.PythonRDDFunction']
+        implements = ['org.apache.spark.streaming.api.python.PythonRDDFunction2']
 
 
 def rddToFileName(prefix, suffix, time):

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

@@ -28,77 +28,91 @@ import org.apache.spark.streaming.api.java._
 
 
 /**
- * Interface for Python callback function
+ * Interface for Python callback function with two arguments
  */
 trait PythonRDDFunction {
-  def call(rdd: JavaRDD[_], rdd2: JavaRDD[_], time: Long): JavaRDD[Array[Byte]]
+  def call(rdd: JavaRDD[_], time: Long): JavaRDD[Array[Byte]]
 }
 
+/**
+ * Interface for Python callback function with three arguments
+ */
+trait PythonRDDFunction2 {
+  def call(rdd: JavaRDD[_], rdd2: JavaRDD[_], time: Long): JavaRDD[Array[Byte]]
+}
 
 /**
  * Transformed DStream in Python.
  *
  * If the result RDD is PythonRDD, then it will cache it as an template for future use,
  * this can reduce the Python callbacks.
- *
- * @param parent
- * @param parent2
- * @param func
- * @param cache
  */
-class PythonTransformedDStream (parent: DStream[_], parent2: DStream[_], func: PythonRDDFunction,
-                                cache: Boolean = false)
+class PythonTransformedDStream (parent: DStream[_], func: PythonRDDFunction,
+                                var reuse: Boolean = false)
   extends DStream[Array[Byte]] (parent.ssc) {
 
   var lastResult: PythonRDD = _
 
-  override def dependencies = {
-    if (parent2 == null) {
-      List(parent)
-    } else {
-      List(parent, parent2)
-    }
-  }
+  override def dependencies = List(parent)
 
   override def slideDuration: Duration = parent.slideDuration
 
   override def compute(validTime: Time): Option[RDD[Array[Byte]]] = {
     val rdd1 = parent.getOrCompute(validTime).getOrElse(null)
-    val rdd2 = if (parent2 != null) parent2.getOrCompute(validTime).getOrElse(null) else null
-
-    val r = if (rdd2 != null) {
-      func.call(JavaRDD.fromRDD(rdd1), JavaRDD.fromRDD(rdd2), validTime.milliseconds)
-    } else if (cache && lastResult != null) {
-      lastResult.copyTo(rdd1).asJavaRDD
+    if (reuse && lastResult != null) {
+      Some(lastResult.copyTo(rdd1))
     } else {
-      func.call(JavaRDD.fromRDD(rdd1), null, validTime.milliseconds)
-    }
-    if (r != null) {
-      if (lastResult == null && r.isInstanceOf[PythonRDD]) {
-        lastResult = r.asInstanceOf[PythonRDD]
+      val r = func.call(JavaRDD.fromRDD(rdd1), validTime.milliseconds).rdd
+      if (reuse && lastResult == null) {
+        r match {
+          case rdd: PythonRDD =>
+            if (rdd.parent(0) == rdd1) {
+              // only one PythonRDD
+              lastResult = rdd
+            } else {
+              // may have multiple stages
+              reuse = false
+            }
+        }
       }
       Some(r)
-    } else {
-      None
     }
   }
 
   val asJavaDStream  = JavaDStream.fromDStream(this)
 }
 
+/**
+ * Transformed from two DStreams in Python.
+ */
+class PythonTransformed2DStream (parent: DStream[_], parent2: DStream[_], func: PythonRDDFunction2)
+  extends DStream[Array[Byte]] (parent.ssc) {
+
+  override def dependencies = List(parent, parent2)
+
+  override def slideDuration: Duration = parent.slideDuration
+
+  override def compute(validTime: Time): Option[RDD[Array[Byte]]] = {
+    def resultRdd(stream: DStream[_]): JavaRDD[_] = stream.getOrCompute(validTime) match {
+      case Some(rdd) => JavaRDD.fromRDD(rdd)
+      case None => null
+    }
+    Some(func.call(resultRdd(parent), resultRdd(parent2), validTime.milliseconds))
+  }
+
+  val asJavaDStream  = JavaDStream.fromDStream(this)
+}
 
 /**
  * This is used for foreachRDD() in Python
- * @param prev
- * @param foreachFunction
  */
 class PythonForeachDStream(
     prev: DStream[Array[Byte]],
     foreachFunction: PythonRDDFunction
   ) extends ForEachDStream[Array[Byte]](
     prev,
     (rdd: RDD[Array[Byte]], time: Time) => {
-      foreachFunction.call(rdd.toJavaRDD(), null, time.milliseconds)
+      foreachFunction.call(rdd.toJavaRDD(), time.milliseconds)
     }
   ) {