Enabled ReceivedBlockTracker by default, stored block metadata and optimized block fetching in WALBackedBlockRDD

Author: tdas

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

@@ -31,8 +31,9 @@ private[spark]
 class BlockRDD[T: ClassTag](@transient sc: SparkContext, @transient val blockIds: Array[BlockId])
   extends RDD[T](sc, Nil) {
 
-  @transient lazy val locations_ = BlockManager.blockIdsToHosts(blockIds, SparkEnv.get)
+  @transient lazy val _locations = BlockManager.blockIdsToHosts(blockIds, SparkEnv.get)
   @volatile private var _isValid = true
+  @volatile private var _setInvalid = true
 
   override def getPartitions: Array[Partition] = {
     assertValid()
@@ -54,7 +55,7 @@ class BlockRDD[T: ClassTag](@transient sc: SparkContext, @transient val blockIds
 
   override def getPreferredLocations(split: Partition): Seq[String] = {
     assertValid()
-    locations_(split.asInstanceOf[BlockRDDPartition].blockId)
+    _locations(split.asInstanceOf[BlockRDDPartition].blockId)
   }
 
   /**
@@ -66,7 +67,9 @@ class BlockRDD[T: ClassTag](@transient sc: SparkContext, @transient val blockIds
     blockIds.foreach { blockId =>
       sc.env.blockManager.master.removeBlock(blockId)
     }
-    _isValid = false
+    if (_setInvalid) {
+      _isValid = false
+    }
   }
 
   /**
@@ -85,8 +88,12 @@ class BlockRDD[T: ClassTag](@transient sc: SparkContext, @transient val blockIds
     }
   }
 
+  protected def setInvalidIfBlocksRemoved(setInvalid: Boolean): Unit = {
+    _setInvalid = setInvalid
+  }
+
   protected def getBlockIdLocations(): Map[BlockId, Seq[String]] = {
-    locations_
+    _locations
   }
 }
 

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

@@ -67,27 +67,26 @@ abstract class ReceiverInputDStream[T: ClassTag](@transient ssc_ : StreamingCont
       } else {
         // Otherwise, ask the tracker for all the blocks that have been allocated to this stream
         // for this batch
-        val blockInfos =
-          ssc.scheduler.receiverTracker.getBlocksOfBatch(validTime).get(id).getOrElse(Seq.empty)
-        val blockStoreResults = blockInfos.map { _.blockStoreResult }
-        val blockIds = blockStoreResults.map { _.blockId.asInstanceOf[BlockId] }.toArray
+        val receiverTracker = ssc.scheduler.receiverTracker
+        val blockInfos = receiverTracker.getBlocksOfBatch(validTime).getOrElse(id, Seq.empty)
+        val blockIds = blockInfos.map { _.blockId.asInstanceOf[BlockId] }.toArray
 
-        // Check whether all the results are of the same type
-        val resultTypes = blockStoreResults.map { _.getClass }.distinct
-        if (resultTypes.size > 1) {
-          logWarning("Multiple result types in block information, WAL information will be ignored.")
-        }
+        // Is WAL segment info present with all the blocks
+        val isWALSegmentInfoPresent = blockInfos.forall { _.writeAheadLogSegmentOption.nonEmpty }
 
-        // If all the results are of type WriteAheadLogBasedStoreResult, then create
-        // WriteAheadLogBackedBlockRDD else create simple BlockRDD.
-        if (resultTypes.size == 1 && resultTypes.head == classOf[WriteAheadLogBasedStoreResult]) {
-          val logSegments = blockStoreResults.map {
-            _.asInstanceOf[WriteAheadLogBasedStoreResult].segment
-          }.toArray
-          // Since storeInBlockManager = false, the storage level does not matter.
-          new WriteAheadLogBackedBlockRDD[T](ssc.sparkContext,
-            blockIds, logSegments, storeInBlockManager = false, StorageLevel.MEMORY_ONLY_SER)
+        if (isWALSegmentInfoPresent) {
+          // If all the blocks have WAL segment info, then create a WALBackedBlockRDD
+          val isBlockIdValid = blockInfos.map { _.isBlockIdValid() }.toArray
+          val blockWALSegments = blockInfos.map { _.writeAheadLogSegmentOption.get }.toArray
+          new WriteAheadLogBackedBlockRDD[T](
+            ssc.sparkContext, blockIds, blockWALSegments, isBlockIdValid)
         } else {
+          // Else, create a BlockRDD. However, if there are some blocks with WAL info but not others
+          // then that is unexpected and log a warning accordingly.
+          if (blockInfos.find(_.writeAheadLogSegmentOption.nonEmpty).nonEmpty) {
+            logWarning("Could not find Write Ahead Log information on some of the blocks, " +
+              "data may not be recoverable after driver failures")
+          }
           new BlockRDD[T](ssc.sc, blockIds)
         }
       }

streaming/src/main/scala/org/apache/spark/streaming/rdd/WriteAheadLogBackedBlockRDD.scala

@@ -22,7 +22,7 @@ import org.apache.hadoop.conf.Configuration
 
 import org.apache.spark._
 import org.apache.spark.rdd.BlockRDD
-import org.apache.spark.storage.{BlockId, StorageLevel}
+import org.apache.spark.storage.{StreamBlockId, BlockId, StorageLevel}
 import org.apache.spark.streaming.util.{HdfsUtils, WriteAheadLogFileSegment, WriteAheadLogRandomReader}
 
 /**
@@ -37,6 +37,7 @@ private[streaming]
 class WriteAheadLogBackedBlockRDDPartition(
     val index: Int,
     val blockId: BlockId,
+    val isBlockIdValid: Boolean,
     val segment: WriteAheadLogFileSegment)
   extends Partition
 
@@ -45,11 +46,19 @@ class WriteAheadLogBackedBlockRDDPartition(
  * This class represents a special case of the BlockRDD where the data blocks in
  * the block manager are also backed by segments in write ahead logs. For reading
  * the data, this RDD first looks up the blocks by their ids in the block manager.
- * If it does not find them, it looks up the corresponding file segment.
+ * If it does not find them, it looks up the corresponding file segment. The finding
+ * of the blocks by their ids can be skipped by setting the corresponding element in
+ * isBlockIdValid to false. This is a performance optimization which does not affect
+ * correctness, and it can be used in situations where it is known that the block
+ * does not exist in the Spark executors (e.g. after a failed driver is restarted).
+ *
  *
  * @param sc SparkContext
  * @param blockIds Ids of the blocks that contains this RDD's data
  * @param segments Segments in write ahead logs that contain this RDD's data
+ * @param isBlockIdValid Whether the block Ids are valid (i.e., the blocks are present in the Spark
+ *                         executors). If not, then block lookups by the block ids will be skipped.
+ *                         By default, this is an empty array signifying true for all the blocks.
  * @param storeInBlockManager Whether to store in the block manager after reading from the segment
  * @param storageLevel storage level to store when storing in block manager
  *                     (applicable when storeInBlockManager = true)
@@ -59,23 +68,32 @@ class WriteAheadLogBackedBlockRDD[T: ClassTag](
     @transient sc: SparkContext,
     @transient blockIds: Array[BlockId],
     @transient segments: Array[WriteAheadLogFileSegment],
-    storeInBlockManager: Boolean,
-    storageLevel: StorageLevel)
+    @transient isBlockIdValid: Array[Boolean] = Array.empty,
+    storeInBlockManager: Boolean = false,
+    storageLevel: StorageLevel = StorageLevel.MEMORY_ONLY_SER)
   extends BlockRDD[T](sc, blockIds) {
 
   require(
     blockIds.length == segments.length,
-    s"Number of block ids (${blockIds.length}) must be " +
-      s"the same as number of segments (${segments.length}})!")
+    s"Number of block Ids (${blockIds.length}) must be " +
+      s" same as number of segments (${segments.length}})")
+
+  require(
+    isBlockIdValid.isEmpty || isBlockIdValid.length == blockIds.length,
+    s"Number of elements in isBlockIdValid (${isBlockIdValid.length}) must be " +
+      s" same as number of block Ids (${blockIds.length})")
 
   // Hadoop configuration is not serializable, so broadcast it as a serializable.
   @transient private val hadoopConfig = sc.hadoopConfiguration
   private val broadcastedHadoopConf = new SerializableWritable(hadoopConfig)
 
+  setInvalidIfBlocksRemoved(false)
+
   override def getPartitions: Array[Partition] = {
     assertValid()
-    Array.tabulate(blockIds.size) { i =>
-      new WriteAheadLogBackedBlockRDDPartition(i, blockIds(i), segments(i))
+    Array.tabulate(blockIds.length) { i =>
+      val isValid = if (isBlockIdValid.length == 0) true else isBlockIdValid(i)
+      new WriteAheadLogBackedBlockRDDPartition(i, blockIds(i), isValid, segments(i))
     }
   }
 
@@ -90,22 +108,29 @@ class WriteAheadLogBackedBlockRDD[T: ClassTag](
     val blockManager = SparkEnv.get.blockManager
     val partition = split.asInstanceOf[WriteAheadLogBackedBlockRDDPartition]
     val blockId = partition.blockId
-    blockManager.get(blockId) match {
-      case Some(block) => // Data is in Block Manager
-        val iterator = block.data.asInstanceOf[Iterator[T]]
-        logDebug(s"Read partition data of $this from block manager, block $blockId")
-        iterator
-      case None => // Data not found in Block Manager, grab it from write ahead log file
-        val reader = new WriteAheadLogRandomReader(partition.segment.path, hadoopConf)
-        val dataRead = reader.read(partition.segment)
-        reader.close()
-        logInfo(s"Read partition data of $this from write ahead log, segment ${partition.segment}")
-        if (storeInBlockManager) {
-          blockManager.putBytes(blockId, dataRead, storageLevel)
-          logDebug(s"Stored partition data of $this into block manager with level $storageLevel")
-          dataRead.rewind()
-        }
-        blockManager.dataDeserialize(blockId, dataRead).asInstanceOf[Iterator[T]]
+    val segment = partition.segment
+
+    def getBlockFromBlockManager(): Option[Iterator[T]] = {
+      blockManager.get(blockId).map(_.data.asInstanceOf[Iterator[T]])
+    }
+
+    def getBlockFromWriteAheadLog(): Iterator[T] = {
+      val reader = new WriteAheadLogRandomReader(segment.path, hadoopConf)
+      val dataRead = reader.read(segment)
+      reader.close()
+      logDebug(s"Read partition data of $this from write ahead log, segment ${partition.segment}")
+      if (storeInBlockManager) {
+        blockManager.putBytes(blockId, dataRead, storageLevel)
+        logDebug(s"Stored partition data of $this into block manager with level $storageLevel")
+        dataRead.rewind()
+      }
+      blockManager.dataDeserialize(blockId, dataRead).asInstanceOf[Iterator[T]]
+    }
+
+    if (partition.isBlockIdValid) {
+      getBlockFromBlockManager().getOrElse { getBlockFromWriteAheadLog() }
+    } else {
+      getBlockFromWriteAheadLog()
     }
   }
 
@@ -116,7 +141,12 @@ class WriteAheadLogBackedBlockRDD[T: ClassTag](
    */
   override def getPreferredLocations(split: Partition): Seq[String] = {
     val partition = split.asInstanceOf[WriteAheadLogBackedBlockRDDPartition]
-    val blockLocations = getBlockIdLocations().get(partition.blockId)
+    val blockLocations = if (partition.isBlockIdValid) {
+      getBlockIdLocations().get(partition.blockId)
+    } else {
+      None
+    }
+
     blockLocations.getOrElse(
       HdfsUtils.getFileSegmentLocations(
         partition.segment.path, partition.segment.offset, partition.segment.length, hadoopConfig))

streaming/src/main/scala/org/apache/spark/streaming/receiver/ReceiverSupervisorImpl.scala

@@ -145,7 +145,7 @@ private[streaming] class ReceiverSupervisorImpl(
     val blockStoreResult = receivedBlockHandler.storeBlock(blockId, receivedBlock)
     logDebug(s"Pushed block $blockId in ${(System.currentTimeMillis - time)} ms")
 
-    val blockInfo = ReceivedBlockInfo(streamId, numRecords, blockStoreResult)
+    val blockInfo = ReceivedBlockInfo(streamId, numRecords, metadataOption, blockStoreResult)
     trackerEndpoint.askWithReply[Boolean](AddBlock(blockInfo))
     logDebug(s"Reported block $blockId")
   }

streaming/src/main/scala/org/apache/spark/streaming/scheduler/ReceivedBlockInfo.scala

@@ -17,12 +17,38 @@
 
 package org.apache.spark.streaming.scheduler
 
-import org.apache.spark.streaming.receiver.ReceivedBlockStoreResult
+import org.apache.spark.storage.StreamBlockId
+import org.apache.spark.streaming.receiver.{WriteAheadLogBasedStoreResult, ReceivedBlockStoreResult}
+import org.apache.spark.streaming.util.WriteAheadLogFileSegment
 
 /** Information about blocks received by the receiver */
 private[streaming] case class ReceivedBlockInfo(
     streamId: Int,
     numRecords: Long,
+    metadataOption: Option[Any],
     blockStoreResult: ReceivedBlockStoreResult
-  )
+  ) {
+
+  @volatile private var _isBlockIdValid = true
+
+  def blockId: StreamBlockId = blockStoreResult.blockId
+
+  def writeAheadLogSegmentOption: Option[WriteAheadLogFileSegment] = {
+    blockStoreResult match {
+      case walStoreResult: WriteAheadLogBasedStoreResult => Some(walStoreResult.segment)
+      case _ => None
+    }
+  }
+
+  /** Is the block ID valid, that is, is the block present in the Spark executors. */
+  def isBlockIdValid(): Boolean = _isBlockIdValid
+
+  /**
+   * Set the block ID as invalid. This is useful when it is known that the block is not present
+   * in the Spark executors.
+   */
+  def setBlockIdInvalid(): Unit = {
+    _isBlockIdValid = false
+  }
+}
 

streaming/src/main/scala/org/apache/spark/streaming/scheduler/ReceivedBlockTracker.scala

@@ -25,10 +25,10 @@ import scala.language.implicitConversions
 import org.apache.hadoop.conf.Configuration
 import org.apache.hadoop.fs.Path
 
-import org.apache.spark.{SparkException, Logging, SparkConf}
 import org.apache.spark.streaming.Time
 import org.apache.spark.streaming.util.WriteAheadLogManager
 import org.apache.spark.util.{Clock, Utils}
+import org.apache.spark.{Logging, SparkConf, SparkException}
 
 /** Trait representing any event in the ReceivedBlockTracker that updates its state. */
 private[streaming] sealed trait ReceivedBlockTrackerLogEvent
@@ -45,7 +45,7 @@ private[streaming] case class BatchCleanupEvent(times: Seq[Time])
 private[streaming]
 case class AllocatedBlocks(streamIdToAllocatedBlocks: Map[Int, Seq[ReceivedBlockInfo]]) {
   def getBlocksOfStream(streamId: Int): Seq[ReceivedBlockInfo] = {
-    streamIdToAllocatedBlocks.get(streamId).getOrElse(Seq.empty)
+    streamIdToAllocatedBlocks.getOrElse(streamId, Seq.empty)
   }
 }
 
@@ -171,6 +171,7 @@ private[streaming] class ReceivedBlockTracker(
     // Insert the recovered block information
     def insertAddedBlock(receivedBlockInfo: ReceivedBlockInfo) {
       logTrace(s"Recovery: Inserting added block $receivedBlockInfo")
+      receivedBlockInfo.setBlockIdInvalid()
       getReceivedBlockQueue(receivedBlockInfo.streamId) += receivedBlockInfo
     }
 
@@ -223,22 +224,13 @@ private[streaming] class ReceivedBlockTracker(
 
   /** Optionally create the write ahead log manager only if the feature is enabled */
   private def createLogManager(): Option[WriteAheadLogManager] = {
-    if (conf.getBoolean("spark.streaming.receiver.writeAheadLog.enable", false)) {
-      if (checkpointDirOption.isEmpty) {
-        throw new SparkException(
-          "Cannot enable receiver write-ahead log without checkpoint directory set. " +
-            "Please use streamingContext.checkpoint() to set the checkpoint directory. " +
-            "See documentation for more details.")
-      }
-      val logDir = ReceivedBlockTracker.checkpointDirToLogDir(checkpointDirOption.get)
+    checkpointDirOption.map { checkpointDir =>
+      val logDir = ReceivedBlockTracker.checkpointDirToLogDir(checkpointDir)
       val rollingIntervalSecs = conf.getInt(
         "spark.streaming.receivedBlockTracker.writeAheadLog.rotationIntervalSecs", 60)
-      val logManager = new WriteAheadLogManager(logDir, hadoopConf,
+      new WriteAheadLogManager(logDir, hadoopConf,
         rollingIntervalSecs = rollingIntervalSecs, clock = clock,
         callerName = "ReceivedBlockHandlerMaster")
-      Some(logManager)
-    } else {
-      None
     }
   }
 

streaming/src/test/scala/org/apache/spark/streaming/ReceivedBlockTrackerSuite.scala

@@ -88,7 +88,7 @@ class ReceivedBlockTrackerSuite
     receivedBlockTracker.getUnallocatedBlocks(streamId) shouldEqual blockInfos
   }
 
-  test("block addition, block to batch allocation and cleanup with write ahead log") {
+  test("recovery and cleanup with write ahead logs") {
     val manualClock = new ManualClock
     // Set the time increment level to twice the rotation interval so that every increment creates
     // a new log file
@@ -114,7 +114,6 @@ class ReceivedBlockTrackerSuite
     }
 
     // Start tracker and add blocks
-    conf.set("spark.streaming.receiver.writeAheadLog.enable", "true")
     conf.set("spark.streaming.receivedBlockTracker.writeAheadLog.rotationIntervalSecs", "1")
     val tracker1 = createTracker(clock = manualClock)
     tracker1.isLogManagerEnabled should be (true)
@@ -130,7 +129,11 @@ class ReceivedBlockTrackerSuite
     // Restart tracker and verify recovered list of unallocated blocks
     incrementTime()
     val tracker2 = createTracker(clock = manualClock)
-    tracker2.getUnallocatedBlocks(streamId).toList shouldEqual blockInfos1
+    val unallocatedBlocks = tracker2.getUnallocatedBlocks(streamId).toList
+    unallocatedBlocks shouldEqual blockInfos1
+    unallocatedBlocks.foreach { block =>
+      block.isBlockIdValid() should be (false)
+    }
 
     // Allocate blocks to batch and verify whether the unallocated blocks got allocated
     val batchTime1 = manualClock.getTimeMillis()
@@ -182,22 +185,10 @@ class ReceivedBlockTrackerSuite
     tracker4.getBlocksOfBatchAndStream(batchTime2, streamId) shouldEqual blockInfos2
   }
 
-  test("enabling write ahead log but not setting checkpoint dir") {
-    conf.set("spark.streaming.receiver.writeAheadLog.enable", "true")
-    intercept[SparkException] {
-      createTracker(setCheckpointDir = false)
-    }
-  }
-
-  test("setting checkpoint dir but not enabling write ahead log") {
-    // When WAL config is not set, log manager should not be enabled
-    val tracker1 = createTracker(setCheckpointDir = true)
+  test("write ahead log disabled when not checkpoint directory is set") {
+    // When checkpoint is not enabled, then the write ahead log is also disabled
+    val tracker1 = createTracker(setCheckpointDir = false)
     tracker1.isLogManagerEnabled should be (false)
-
-    // When WAL is explicitly disabled, log manager should not be enabled
-    conf.set("spark.streaming.receiver.writeAheadLog.enable", "false")
-    val tracker2 = createTracker(setCheckpointDir = true)
-    tracker2.isLogManagerEnabled should be(false)
   }
 
   /**
@@ -215,7 +206,7 @@ class ReceivedBlockTrackerSuite
 
   /** Generate blocks infos using random ids */
   def generateBlockInfos(): Seq[ReceivedBlockInfo] = {
-    List.fill(5)(ReceivedBlockInfo(streamId, 0,
+    List.fill(5)(ReceivedBlockInfo(streamId, 0, None,
       BlockManagerBasedStoreResult(StreamBlockId(streamId, math.abs(Random.nextInt)))))
   }