[SPARK-4092] [CORE] Fix InputMetrics for coalesce'd Rdds

core/src/main/scala/org/apache/spark/CacheManager.scala

@@ -44,7 +44,11 @@ private[spark] class CacheManager(blockManager: BlockManager) extends Logging {
     blockManager.get(key) match {
       case Some(blockResult) =>
         // Partition is already materialized, so just return its values
-        context.taskMetrics.inputMetrics = Some(blockResult.inputMetrics)
+        val inputMetrics = blockResult.inputMetrics
+        val existingMetrics = context.taskMetrics
+          .getInputMetricsForReadMethod(inputMetrics.readMethod)
+        existingMetrics.addBytesRead(inputMetrics.bytesRead)
+
         new InterruptibleIterator(context, blockResult.data.asInstanceOf[Iterator[T]])
 
       case None =>

core/src/main/scala/org/apache/spark/executor/Executor.scala

@@ -370,6 +370,7 @@ private[spark] class Executor(
             if (!taskRunner.attemptedTask.isEmpty) {
               Option(taskRunner.task).flatMap(_.metrics).foreach { metrics =>
                 metrics.updateShuffleReadMetrics
+                metrics.updateInputMetrics()
                 metrics.jvmGCTime = curGCTime - taskRunner.startGCTime
                 if (isLocal) {
                   // JobProgressListener will hold an reference of it during

core/src/main/scala/org/apache/spark/executor/TaskMetrics.scala

@@ -17,6 +17,11 @@
 
 package org.apache.spark.executor
 
+import java.util.concurrent.atomic.AtomicLong
+
+import org.apache.spark.executor.DataReadMethod
+import org.apache.spark.executor.DataReadMethod.DataReadMethod
+
 import scala.collection.mutable.ArrayBuffer
 
 import org.apache.spark.annotation.DeveloperApi
@@ -80,7 +85,17 @@ class TaskMetrics extends Serializable {
    * If this task reads from a HadoopRDD or from persisted data, metrics on how much data was read
    * are stored here.
    */
-  var inputMetrics: Option[InputMetrics] = None
+  private var _inputMetrics: Option[InputMetrics] = None
+
+  def inputMetrics = _inputMetrics
+
+  /**
+   * This should only be used when recreating TaskMetrics, not when updating input metrics in
+   * executors
+   */
+  private[spark] def setInputMetrics(inputMetrics: Option[InputMetrics]) {
+    _inputMetrics = inputMetrics
+  }
 
   /**
    * If this task writes data externally (e.g. to a distributed filesystem), metrics on how much
@@ -133,6 +148,30 @@ class TaskMetrics extends Serializable {
     readMetrics
   }
 
+  /**
+   * Returns the input metrics object that the task should use. Currently, if
+   * there exists an input metric with the same readMethod, we return that one
+   * so the caller can accumulate bytes read. If the readMethod is different
+   * than previously seen by this task, we return a new InputMetric but don't
+   * record it.
+   *
+   * Once https://issues.apache.org/jira/browse/SPARK-5225 is addressed,
+   * we can store all the different inputMetrics (one per readMethod).
+   */
+  private[spark] def getInputMetricsForReadMethod(readMethod: DataReadMethod):
+    InputMetrics =synchronized {
+    _inputMetrics match {
+      case None =>
+        val metrics = new InputMetrics(readMethod)
+        _inputMetrics = Some(metrics)
+        metrics
+      case Some(metrics @ InputMetrics(method)) if method == readMethod =>
+        metrics
+      case Some(InputMetrics(method)) =>
+       new InputMetrics(readMethod)
+    }
+  }
+
   /**
    * Aggregates shuffle read metrics for all registered dependencies into shuffleReadMetrics.
    */
@@ -146,6 +185,10 @@ class TaskMetrics extends Serializable {
     }
     _shuffleReadMetrics = Some(merged)
   }
+
+  private[spark] def updateInputMetrics() = synchronized {
+    inputMetrics.foreach(_.updateBytesRead())
+  }
 }
 
 private[spark] object TaskMetrics {
@@ -179,10 +222,38 @@ object DataWriteMethod extends Enumeration with Serializable {
  */
 @DeveloperApi
 case class InputMetrics(readMethod: DataReadMethod.Value) {
+
+  private val _bytesRead: AtomicLong = new AtomicLong()
+
   /**
    * Total bytes read.
    */
-  var bytesRead: Long = 0L
+  def bytesRead: Long = _bytesRead.get()
+  @volatile @transient var bytesReadCallback: Option[() => Long] = None
+
+  /**
+   * Adds additional bytes read for this read method.
+   */
+  def addBytesRead(bytes: Long) = {
+    _bytesRead.addAndGet(bytes)
+  }
+
+  /**
+   * Invoke the bytesReadCallback and mutate bytesRead.
+   */
+  def updateBytesRead() {
+    bytesReadCallback.foreach { c =>
+      _bytesRead.set(c())
+    }
+  }
+
+ /**
+  * Register a function that can be called to get up-to-date information on how many bytes the task
+  * has read from an input source.
+  */
+  def setBytesReadCallback(f: Option[() => Long]) {
+    bytesReadCallback = f
+  }
 }
 
 /**

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

@@ -213,18 +213,19 @@ class HadoopRDD[K, V](
       logInfo("Input split: " + split.inputSplit)
       val jobConf = getJobConf()
 
-      val inputMetrics = new InputMetrics(DataReadMethod.Hadoop)
+      val inputMetrics = context.taskMetrics
+        .getInputMetricsForReadMethod(DataReadMethod.Hadoop)
+
       // Find a function that will return the FileSystem bytes read by this thread. Do this before
       // creating RecordReader, because RecordReader's constructor might read some bytes
-      val bytesReadCallback = if (split.inputSplit.value.isInstanceOf[FileSplit]) {
-        SparkHadoopUtil.get.getFSBytesReadOnThreadCallback(
-          split.inputSplit.value.asInstanceOf[FileSplit].getPath, jobConf)
-      } else {
-        None
-      }
-      if (bytesReadCallback.isDefined) {
-        context.taskMetrics.inputMetrics = Some(inputMetrics)
-      }
+      val bytesReadCallback = inputMetrics.bytesReadCallback.orElse(
+        split.inputSplit.value match {
+          case split: FileSplit =>
+            SparkHadoopUtil.get.getFSBytesReadOnThreadCallback(split.getPath, jobConf)
+          case _ => None
+        }
+      )
+      inputMetrics.setBytesReadCallback(bytesReadCallback)
 
       var reader: RecordReader[K, V] = null
       val inputFormat = getInputFormat(jobConf)
@@ -237,40 +238,26 @@ class HadoopRDD[K, V](
       val key: K = reader.createKey()
       val value: V = reader.createValue()
 
-      var recordsSinceMetricsUpdate = 0
-
       override def getNext() = {
         try {
           finished = !reader.next(key, value)
         } catch {
           case eof: EOFException =>
             finished = true
         }
-
-        // Update bytes read metric every few records
-        if (recordsSinceMetricsUpdate == HadoopRDD.RECORDS_BETWEEN_BYTES_READ_METRIC_UPDATES
-            && bytesReadCallback.isDefined) {
-          recordsSinceMetricsUpdate = 0
-          val bytesReadFn = bytesReadCallback.get
-          inputMetrics.bytesRead = bytesReadFn()
-        } else {
-          recordsSinceMetricsUpdate += 1
-        }
         (key, value)
       }
 
       override def close() {
         try {
           reader.close()
           if (bytesReadCallback.isDefined) {
-            val bytesReadFn = bytesReadCallback.get
-            inputMetrics.bytesRead = bytesReadFn()
+            inputMetrics.updateBytesRead()
           } else if (split.inputSplit.value.isInstanceOf[FileSplit]) {
             // If we can't get the bytes read from the FS stats, fall back to the split size,
             // which may be inaccurate.
             try {
-              inputMetrics.bytesRead = split.inputSplit.value.getLength
-              context.taskMetrics.inputMetrics = Some(inputMetrics)
+              inputMetrics.addBytesRead(split.inputSplit.value.getLength)
             } catch {
               case e: java.io.IOException =>
                 logWarning("Unable to get input size to set InputMetrics for task", e)

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

@@ -109,18 +109,19 @@ class NewHadoopRDD[K, V](
       logInfo("Input split: " + split.serializableHadoopSplit)
       val conf = confBroadcast.value.value
 
-      val inputMetrics = new InputMetrics(DataReadMethod.Hadoop)
+      val inputMetrics = context.taskMetrics
+        .getInputMetricsForReadMethod(DataReadMethod.Hadoop)
+
       // Find a function that will return the FileSystem bytes read by this thread. Do this before
       // creating RecordReader, because RecordReader's constructor might read some bytes
-      val bytesReadCallback = if (split.serializableHadoopSplit.value.isInstanceOf[FileSplit]) {
-        SparkHadoopUtil.get.getFSBytesReadOnThreadCallback(
-          split.serializableHadoopSplit.value.asInstanceOf[FileSplit].getPath, conf)
-      } else {
-        None
-      }
-      if (bytesReadCallback.isDefined) {
-        context.taskMetrics.inputMetrics = Some(inputMetrics)
-      }
+      val bytesReadCallback = inputMetrics.bytesReadCallback.orElse(
+        split.serializableHadoopSplit.value match {
+          case split: FileSplit =>
+            SparkHadoopUtil.get.getFSBytesReadOnThreadCallback(split.getPath, conf)
+          case _ => None
+        }
+      )
+      inputMetrics.setBytesReadCallback(bytesReadCallback)
 
       val attemptId = newTaskAttemptID(jobTrackerId, id, isMap = true, split.index, 0)
       val hadoopAttemptContext = newTaskAttemptContext(conf, attemptId)
@@ -153,34 +154,19 @@ class NewHadoopRDD[K, V](
           throw new java.util.NoSuchElementException("End of stream")
         }
         havePair = false
-
-        // Update bytes read metric every few records
-        if (recordsSinceMetricsUpdate == HadoopRDD.RECORDS_BETWEEN_BYTES_READ_METRIC_UPDATES
-            && bytesReadCallback.isDefined) {
-          recordsSinceMetricsUpdate = 0
-          val bytesReadFn = bytesReadCallback.get
-          inputMetrics.bytesRead = bytesReadFn()
-        } else {
-          recordsSinceMetricsUpdate += 1
-        }
-
         (reader.getCurrentKey, reader.getCurrentValue)
       }
 
       private def close() {
         try {
           reader.close()
-
-          // Update metrics with final amount
           if (bytesReadCallback.isDefined) {
-            val bytesReadFn = bytesReadCallback.get
-            inputMetrics.bytesRead = bytesReadFn()
+            inputMetrics.updateBytesRead()
           } else if (split.serializableHadoopSplit.value.isInstanceOf[FileSplit]) {
             // If we can't get the bytes read from the FS stats, fall back to the split size,
             // which may be inaccurate.
             try {
-              inputMetrics.bytesRead = split.serializableHadoopSplit.value.getLength
-              context.taskMetrics.inputMetrics = Some(inputMetrics)
+              inputMetrics.addBytesRead(split.serializableHadoopSplit.value.getLength)
             } catch {
               case e: java.io.IOException =>
                 logWarning("Unable to get input size to set InputMetrics for task", e)

core/src/main/scala/org/apache/spark/storage/BlockManager.scala

@@ -54,7 +54,7 @@ private[spark] class BlockResult(
     readMethod: DataReadMethod.Value,
     bytes: Long) {
   val inputMetrics = new InputMetrics(readMethod)
-  inputMetrics.bytesRead = bytes
+  inputMetrics.addBytesRead(bytes)
 }
 
 /**

core/src/main/scala/org/apache/spark/util/JsonProtocol.scala

@@ -604,8 +604,8 @@ private[spark] object JsonProtocol {
       Utils.jsonOption(json \ "Shuffle Read Metrics").map(shuffleReadMetricsFromJson))
     metrics.shuffleWriteMetrics =
       Utils.jsonOption(json \ "Shuffle Write Metrics").map(shuffleWriteMetricsFromJson)
-    metrics.inputMetrics =
-      Utils.jsonOption(json \ "Input Metrics").map(inputMetricsFromJson)
+    metrics.setInputMetrics(
+      Utils.jsonOption(json \ "Input Metrics").map(inputMetricsFromJson))
     metrics.outputMetrics =
       Utils.jsonOption(json \ "Output Metrics").map(outputMetricsFromJson)
     metrics.updatedBlocks =
@@ -638,7 +638,7 @@ private[spark] object JsonProtocol {
   def inputMetricsFromJson(json: JValue): InputMetrics = {
     val metrics = new InputMetrics(
       DataReadMethod.withName((json \ "Data Read Method").extract[String]))
-    metrics.bytesRead = (json \ "Bytes Read").extract[Long]
+    metrics.addBytesRead((json \ "Bytes Read").extract[Long])
     metrics
   }