[SPARK-11932][STREAMING] Partition previous TrackStateRDD if partitioner not present

tdas · tdas · commit 5d80d8c6a54b · 2015-12-07T11:03:59.000-08:00
The reason is that TrackStateRDDs generated by trackStateByKey expect the previous batch's TrackStateRDDs to have a partitioner. However, when recovery from DStream checkpoints, the RDDs recovered from RDD checkpoints do not have a partitioner attached to it. This is because RDD checkpoints do not preserve the partitioner (SPARK-12004). While #9983 solves SPARK-12004 by preserving the partitioner through RDD checkpoints, there may be a non-zero chance that the saving and recovery fails. To be resilient, this PR repartitions the previous state RDD if the partitioner is not detected. Author: Tathagata Das <tathagata.das1565@gmail.com> Closes #9988 from tdas/SPARK-11932.
diff --git a/streaming/src/main/scala/org/apache/spark/streaming/Checkpoint.scala b/streaming/src/main/scala/org/apache/spark/streaming/Checkpoint.scala
@@ -277,7 +277,7 @@ class CheckpointWriter(
       val bytes = Checkpoint.serialize(checkpoint, conf)
       executor.execute(new CheckpointWriteHandler(
         checkpoint.checkpointTime, bytes, clearCheckpointDataLater))
-      logDebug("Submitted checkpoint of time " + checkpoint.checkpointTime + " writer queue")
+      logInfo("Submitted checkpoint of time " + checkpoint.checkpointTime + " writer queue")
     } catch {
       case rej: RejectedExecutionException =>
         logError("Could not submit checkpoint task to the thread pool executor", rej)
diff --git a/streaming/src/main/scala/org/apache/spark/streaming/dstream/TrackStateDStream.scala b/streaming/src/main/scala/org/apache/spark/streaming/dstream/TrackStateDStream.scala
@@ -132,22 +132,37 @@ class InternalTrackStateDStream[K: ClassTag, V: ClassTag, S: ClassTag, E: ClassT
   /** Method that generates a RDD for the given time */
   override def compute(validTime: Time): Option[RDD[TrackStateRDDRecord[K, S, E]]] = {
     // Get the previous state or create a new empty state RDD
-    val prevStateRDD = getOrCompute(validTime - slideDuration).getOrElse {
-      TrackStateRDD.createFromPairRDD[K, V, S, E](
-        spec.getInitialStateRDD().getOrElse(new EmptyRDD[(K, S)](ssc.sparkContext)),
-        partitioner, validTime
-      )
+    val prevStateRDD = getOrCompute(validTime - slideDuration) match {
+      case Some(rdd) =>
+        if (rdd.partitioner != Some(partitioner)) {
+          // If the RDD is not partitioned the right way, let us repartition it using the
+          // partition index as the key. This is to ensure that state RDD is always partitioned
+          // before creating another state RDD using it
+          TrackStateRDD.createFromRDD[K, V, S, E](
+            rdd.flatMap { _.stateMap.getAll() }, partitioner, validTime)
+        } else {
+          rdd
+        }
+      case None =>
+        TrackStateRDD.createFromPairRDD[K, V, S, E](
+          spec.getInitialStateRDD().getOrElse(new EmptyRDD[(K, S)](ssc.sparkContext)),
+          partitioner,
+          validTime
+        )
     }
 
+
     // Compute the new state RDD with previous state RDD and partitioned data RDD
-    parent.getOrCompute(validTime).map { dataRDD =>
-      val partitionedDataRDD = dataRDD.partitionBy(partitioner)
-      val timeoutThresholdTime = spec.getTimeoutInterval().map { interval =>
-        (validTime - interval).milliseconds
-      }
-      new TrackStateRDD(
-        prevStateRDD, partitionedDataRDD, trackingFunction, validTime, timeoutThresholdTime)
+    // Even if there is no data RDD, use an empty one to create a new state RDD
+    val dataRDD = parent.getOrCompute(validTime).getOrElse {
+      context.sparkContext.emptyRDD[(K, V)]
+    }
+    val partitionedDataRDD = dataRDD.partitionBy(partitioner)
+    val timeoutThresholdTime = spec.getTimeoutInterval().map { interval =>
+      (validTime - interval).milliseconds
     }
+    Some(new TrackStateRDD(
+      prevStateRDD, partitionedDataRDD, trackingFunction, validTime, timeoutThresholdTime))
   }
 }
 
diff --git a/streaming/src/main/scala/org/apache/spark/streaming/rdd/TrackStateRDD.scala b/streaming/src/main/scala/org/apache/spark/streaming/rdd/TrackStateRDD.scala
@@ -179,22 +179,43 @@ private[streaming] class TrackStateRDD[K: ClassTag, V: ClassTag, S: ClassTag, E:
 
 private[streaming] object TrackStateRDD {
 
-  def createFromPairRDD[K: ClassTag, V: ClassTag, S: ClassTag, T: ClassTag](
+  def createFromPairRDD[K: ClassTag, V: ClassTag, S: ClassTag, E: ClassTag](
       pairRDD: RDD[(K, S)],
       partitioner: Partitioner,
-      updateTime: Time): TrackStateRDD[K, V, S, T] = {
+      updateTime: Time): TrackStateRDD[K, V, S, E] = {
 
     val rddOfTrackStateRecords = pairRDD.partitionBy(partitioner).mapPartitions ({ iterator =>
       val stateMap = StateMap.create[K, S](SparkEnv.get.conf)
       iterator.foreach { case (key, state) => stateMap.put(key, state, updateTime.milliseconds) }
-      Iterator(TrackStateRDDRecord(stateMap, Seq.empty[T]))
+      Iterator(TrackStateRDDRecord(stateMap, Seq.empty[E]))
     }, preservesPartitioning = true)
 
     val emptyDataRDD = pairRDD.sparkContext.emptyRDD[(K, V)].partitionBy(partitioner)
 
     val noOpFunc = (time: Time, key: K, value: Option[V], state: State[S]) => None
 
-    new TrackStateRDD[K, V, S, T](rddOfTrackStateRecords, emptyDataRDD, noOpFunc, updateTime, None)
+    new TrackStateRDD[K, V, S, E](rddOfTrackStateRecords, emptyDataRDD, noOpFunc, updateTime, None)
+  }
+
+  def createFromRDD[K: ClassTag, V: ClassTag, S: ClassTag, E: ClassTag](
+      rdd: RDD[(K, S, Long)],
+      partitioner: Partitioner,
+      updateTime: Time): TrackStateRDD[K, V, S, E] = {
+
+    val pairRDD = rdd.map { x => (x._1, (x._2, x._3)) }
+    val rddOfTrackStateRecords = pairRDD.partitionBy(partitioner).mapPartitions({ iterator =>
+      val stateMap = StateMap.create[K, S](SparkEnv.get.conf)
+      iterator.foreach { case (key, (state, updateTime)) =>
+        stateMap.put(key, state, updateTime)
+      }
+      Iterator(TrackStateRDDRecord(stateMap, Seq.empty[E]))
+    }, preservesPartitioning = true)
+
+    val emptyDataRDD = pairRDD.sparkContext.emptyRDD[(K, V)].partitionBy(partitioner)
+
+    val noOpFunc = (time: Time, key: K, value: Option[V], state: State[S]) => None
+
+    new TrackStateRDD[K, V, S, E](rddOfTrackStateRecords, emptyDataRDD, noOpFunc, updateTime, None)
   }
 }
 
diff --git a/streaming/src/test/scala/org/apache/spark/streaming/CheckpointSuite.scala b/streaming/src/test/scala/org/apache/spark/streaming/CheckpointSuite.scala
@@ -33,17 +33,149 @@ import org.mockito.Mockito.mock
 import org.scalatest.concurrent.Eventually._
 import org.scalatest.time.SpanSugar._
 
-import org.apache.spark.TestUtils
+import org.apache.spark.{SparkConf, SparkContext, SparkFunSuite, TestUtils}
 import org.apache.spark.streaming.dstream.{DStream, FileInputDStream}
 import org.apache.spark.streaming.scheduler._
 import org.apache.spark.util.{MutableURLClassLoader, Clock, ManualClock, Utils}
 
+/**
+ * A trait of that can be mixed in to get methods for testing DStream operations under
+ * DStream checkpointing. Note that the implementations of this trait has to implement
+ * the `setupCheckpointOperation`
+ */
+trait DStreamCheckpointTester { self: SparkFunSuite =>
+
+  /**
+   * Tests a streaming operation under checkpointing, by restarting the operation
+   * from checkpoint file and verifying whether the final output is correct.
+   * The output is assumed to have come from a reliable queue which an replay
+   * data as required.
+   *
+   * NOTE: This takes into consideration that the last batch processed before
+   * master failure will be re-processed after restart/recovery.
+   */
+  protected def testCheckpointedOperation[U: ClassTag, V: ClassTag](
+      input: Seq[Seq[U]],
+      operation: DStream[U] => DStream[V],
+      expectedOutput: Seq[Seq[V]],
+      numBatchesBeforeRestart: Int,
+      batchDuration: Duration = Milliseconds(500),
+      stopSparkContextAfterTest: Boolean = true
+    ) {
+    require(numBatchesBeforeRestart < expectedOutput.size,
+      "Number of batches before context restart less than number of expected output " +
+        "(i.e. number of total batches to run)")
+    require(StreamingContext.getActive().isEmpty,
+      "Cannot run test with already active streaming context")
+
+    // Current code assumes that number of batches to be run = number of inputs
+    val totalNumBatches = input.size
+    val batchDurationMillis = batchDuration.milliseconds
+
+    // Setup the stream computation
+    val checkpointDir = Utils.createTempDir(this.getClass.getSimpleName()).toString
+    logDebug(s"Using checkpoint directory $checkpointDir")
+    val ssc = createContextForCheckpointOperation(batchDuration)
+    require(ssc.conf.get("spark.streaming.clock") === classOf[ManualClock].getName,
+      "Cannot run test without manual clock in the conf")
+
+    val inputStream = new TestInputStream(ssc, input, numPartitions = 2)
+    val operatedStream = operation(inputStream)
+    operatedStream.print()
+    val outputStream = new TestOutputStreamWithPartitions(operatedStream,
+      new ArrayBuffer[Seq[Seq[V]]] with SynchronizedBuffer[Seq[Seq[V]]])
+    outputStream.register()
+    ssc.checkpoint(checkpointDir)
+
+    // Do the computation for initial number of batches, create checkpoint file and quit
+    val beforeRestartOutput = generateOutput[V](ssc,
+      Time(batchDurationMillis * numBatchesBeforeRestart), checkpointDir, stopSparkContextAfterTest)
+    assertOutput(beforeRestartOutput, expectedOutput, beforeRestart = true)
+    // Restart and complete the computation from checkpoint file
+    logInfo(
+      "\n-------------------------------------------\n" +
+        "        Restarting stream computation          " +
+        "\n-------------------------------------------\n"
+    )
+
+    val restartedSsc = new StreamingContext(checkpointDir)
+    val afterRestartOutput = generateOutput[V](restartedSsc,
+      Time(batchDurationMillis * totalNumBatches), checkpointDir, stopSparkContextAfterTest)
+    assertOutput(afterRestartOutput, expectedOutput, beforeRestart = false)
+  }
+
+  protected def createContextForCheckpointOperation(batchDuration: Duration): StreamingContext = {
+    val conf = new SparkConf().setMaster("local").setAppName(this.getClass.getSimpleName)
+    conf.set("spark.streaming.clock", classOf[ManualClock].getName())
+    new StreamingContext(SparkContext.getOrCreate(conf), batchDuration)
+  }
+
+  private def generateOutput[V: ClassTag](
+      ssc: StreamingContext,
+      targetBatchTime: Time,
+      checkpointDir: String,
+      stopSparkContext: Boolean
+    ): Seq[Seq[V]] = {
+    try {
+      val batchDuration = ssc.graph.batchDuration
+      val batchCounter = new BatchCounter(ssc)
+      ssc.start()
+      val clock = ssc.scheduler.clock.asInstanceOf[ManualClock]
+      val currentTime = clock.getTimeMillis()
+
+      logInfo("Manual clock before advancing = " + clock.getTimeMillis())
+      clock.setTime(targetBatchTime.milliseconds)
+      logInfo("Manual clock after advancing = " + clock.getTimeMillis())
+
+      val outputStream = ssc.graph.getOutputStreams().filter { dstream =>
+        dstream.isInstanceOf[TestOutputStreamWithPartitions[V]]
+      }.head.asInstanceOf[TestOutputStreamWithPartitions[V]]
+
+      eventually(timeout(10 seconds)) {
+        ssc.awaitTerminationOrTimeout(10)
+        assert(batchCounter.getLastCompletedBatchTime === targetBatchTime)
+      }
+
+      eventually(timeout(10 seconds)) {
+        val checkpointFilesOfLatestTime = Checkpoint.getCheckpointFiles(checkpointDir).filter {
+          _.toString.contains(clock.getTimeMillis.toString)
+        }
+        // Checkpoint files are written twice for every batch interval. So assert that both
+        // are written to make sure that both of them have been written.
+        assert(checkpointFilesOfLatestTime.size === 2)
+      }
+      outputStream.output.map(_.flatten)
+
+    } finally {
+      ssc.stop(stopSparkContext = stopSparkContext)
+    }
+  }
+
+  private def assertOutput[V: ClassTag](
+      output: Seq[Seq[V]],
+      expectedOutput: Seq[Seq[V]],
+      beforeRestart: Boolean): Unit = {
+    val expectedPartialOutput = if (beforeRestart) {
+      expectedOutput.take(output.size)
+    } else {
+      expectedOutput.takeRight(output.size)
+    }
+    val setComparison = output.zip(expectedPartialOutput).forall {
+      case (o, e) => o.toSet === e.toSet
+    }
+    assert(setComparison, s"set comparison failed\n" +
+      s"Expected output items:\n${expectedPartialOutput.mkString("\n")}\n" +
+      s"Generated output items: ${output.mkString("\n")}"
+    )
+  }
+}
+
 /**
  * This test suites tests the checkpointing functionality of DStreams -
  * the checkpointing of a DStream's RDDs as well as the checkpointing of
  * the whole DStream graph.
  */
-class CheckpointSuite extends TestSuiteBase {
+class CheckpointSuite extends TestSuiteBase with DStreamCheckpointTester {
 
   var ssc: StreamingContext = null
 
@@ -56,7 +188,7 @@ class CheckpointSuite extends TestSuiteBase {
 
   override def afterFunction() {
     super.afterFunction()
-    if (ssc != null) ssc.stop()
+    if (ssc != null) { ssc.stop() }
     Utils.deleteRecursively(new File(checkpointDir))
   }
 
@@ -251,7 +383,9 @@ class CheckpointSuite extends TestSuiteBase {
         Seq(("", 2)),
         Seq(),
         Seq(("a", 2), ("b", 1)),
-        Seq(("", 2)), Seq() ),
+        Seq(("", 2)),
+        Seq()
+      ),
       3
     )
   }
@@ -634,53 +768,6 @@ class CheckpointSuite extends TestSuiteBase {
     checkpointWriter.stop()
   }
 
-  /**
-   * Tests a streaming operation under checkpointing, by restarting the operation
-   * from checkpoint file and verifying whether the final output is correct.
-   * The output is assumed to have come from a reliable queue which an replay
-   * data as required.
-   *
-   * NOTE: This takes into consideration that the last batch processed before
-   * master failure will be re-processed after restart/recovery.
-   */
-  def testCheckpointedOperation[U: ClassTag, V: ClassTag](
-    input: Seq[Seq[U]],
-    operation: DStream[U] => DStream[V],
-    expectedOutput: Seq[Seq[V]],
-    initialNumBatches: Int
-  ) {
-
-    // Current code assumes that:
-    // number of inputs = number of outputs = number of batches to be run
-    val totalNumBatches = input.size
-    val nextNumBatches = totalNumBatches - initialNumBatches
-    val initialNumExpectedOutputs = initialNumBatches
-    val nextNumExpectedOutputs = expectedOutput.size - initialNumExpectedOutputs + 1
-    // because the last batch will be processed again
-
-    // Do the computation for initial number of batches, create checkpoint file and quit
-    ssc = setupStreams[U, V](input, operation)
-    ssc.start()
-    val output = advanceTimeWithRealDelay[V](ssc, initialNumBatches)
-    ssc.stop()
-    verifyOutput[V](output, expectedOutput.take(initialNumBatches), true)
-    Thread.sleep(1000)
-
-    // Restart and complete the computation from checkpoint file
-    logInfo(
-      "\n-------------------------------------------\n" +
-      "        Restarting stream computation          " +
-      "\n-------------------------------------------\n"
-    )
-    ssc = new StreamingContext(checkpointDir)
-    ssc.start()
-    val outputNew = advanceTimeWithRealDelay[V](ssc, nextNumBatches)
-    // the first element will be re-processed data of the last batch before restart
-    verifyOutput[V](outputNew, expectedOutput.takeRight(nextNumExpectedOutputs), true)
-    ssc.stop()
-    ssc = null
-  }
-
   /**
    * Advances the manual clock on the streaming scheduler by given number of batches.
    * It also waits for the expected amount of time for each batch.
diff --git a/streaming/src/test/scala/org/apache/spark/streaming/TestSuiteBase.scala b/streaming/src/test/scala/org/apache/spark/streaming/TestSuiteBase.scala
@@ -142,6 +142,7 @@ class BatchCounter(ssc: StreamingContext) {
   // All access to this state should be guarded by `BatchCounter.this.synchronized`
   private var numCompletedBatches = 0
   private var numStartedBatches = 0
+  private var lastCompletedBatchTime: Time = null
 
   private val listener = new StreamingListener {
     override def onBatchStarted(batchStarted: StreamingListenerBatchStarted): Unit =
@@ -152,6 +153,7 @@ class BatchCounter(ssc: StreamingContext) {
     override def onBatchCompleted(batchCompleted: StreamingListenerBatchCompleted): Unit =
       BatchCounter.this.synchronized {
         numCompletedBatches += 1
+        lastCompletedBatchTime = batchCompleted.batchInfo.batchTime
         BatchCounter.this.notifyAll()
       }
   }
@@ -165,6 +167,10 @@ class BatchCounter(ssc: StreamingContext) {
     numStartedBatches
   }
 
+  def getLastCompletedBatchTime: Time = this.synchronized {
+    lastCompletedBatchTime
+  }
+
   /**
    * Wait until `expectedNumCompletedBatches` batches are completed, or timeout. Return true if
    * `expectedNumCompletedBatches` batches are completed. Otherwise, return false to indicate it's
diff --git a/streaming/src/test/scala/org/apache/spark/streaming/TrackStateByKeySuite.scala b/streaming/src/test/scala/org/apache/spark/streaming/TrackStateByKeySuite.scala
