Merge branch 'master' of github.com:mesos/spark into blockmanager

7 files changed, 218 insertions, 79 deletions

diff --git a/core/src/main/scala/spark/RDD.scala b/core/src/main/scala/spark/RDD.scala
index 09e52ebf3e..fd14ef17f1 100644
--- a/core/src/main/scala/spark/RDD.scala
+++ b/core/src/main/scala/spark/RDD.scala
@@ -107,7 +107,7 @@ abstract class RDD[T: ClassManifest](
   // =======================================================================
 
   /** A unique ID for this RDD (within its SparkContext). */
-  val id = sc.newRddId()
+  val id: Int = sc.newRddId()
 
   /** A friendly name for this RDD */
   var name: String = null
@@ -120,7 +120,8 @@ abstract class RDD[T: ClassManifest](
 
   /**
    * Set this RDD's storage level to persist its values across operations after the first time
-   * it is computed. Can only be called once on each RDD.
+   * it is computed. This can only be used to assign a new storage level if the RDD does not
+   * have a storage level set yet..
    */
   def persist(newLevel: StorageLevel): RDD[T] = {
     // TODO: Handle changes of StorageLevel
@@ -140,6 +141,15 @@ abstract class RDD[T: ClassManifest](
   /** Persist this RDD with the default storage level (`MEMORY_ONLY`). */
   def cache(): RDD[T] = persist()
 
+  /** Mark the RDD as non-persistent, and remove all blocks for it from memory and disk. */
+  def unpersist(): RDD[T] = {
+    logInfo("Removing RDD " + id + " from persistence list")
+    sc.env.blockManager.master.removeRdd(id)
+    sc.persistentRdds.remove(id)
+    storageLevel = StorageLevel.NONE
+    this
+  }
+
   /** Get the RDD's current storage level, or StorageLevel.NONE if none is set. */
   def getStorageLevel = storageLevel
 
diff --git a/core/src/main/scala/spark/SparkContext.scala b/core/src/main/scala/spark/SparkContext.scala
index 5f5ec0b0f4..2ae4ad8659 100644
--- a/core/src/main/scala/spark/SparkContext.scala
+++ b/core/src/main/scala/spark/SparkContext.scala
@@ -1,47 +1,50 @@
 package spark
 
 import java.io._
-import java.util.concurrent.atomic.AtomicInteger
 import java.net.URI
+import java.util.concurrent.ConcurrentHashMap
+import java.util.concurrent.atomic.AtomicInteger
 
+import scala.collection.JavaConversions._
 import scala.collection.Map
 import scala.collection.generic.Growable
-import scala.collection.mutable.HashMap
-import scala.collection.JavaConversions._
+import scala.collection.mutable.{ConcurrentMap, HashMap}
+
+import akka.actor.Actor._
 
-import org.apache.hadoop.fs.Path
 import org.apache.hadoop.conf.Configuration
-import org.apache.hadoop.mapred.InputFormat
-import org.apache.hadoop.mapred.SequenceFileInputFormat
-import org.apache.hadoop.io.Writable
-import org.apache.hadoop.io.IntWritable
-import org.apache.hadoop.io.LongWritable
-import org.apache.hadoop.io.FloatWritable
-import org.apache.hadoop.io.DoubleWritable
+import org.apache.hadoop.fs.Path
+import org.apache.hadoop.io.ArrayWritable
 import org.apache.hadoop.io.BooleanWritable
 import org.apache.hadoop.io.BytesWritable
-import org.apache.hadoop.io.ArrayWritable
+import org.apache.hadoop.io.DoubleWritable
+import org.apache.hadoop.io.FloatWritable
+import org.apache.hadoop.io.IntWritable
+import org.apache.hadoop.io.LongWritable
 import org.apache.hadoop.io.NullWritable
 import org.apache.hadoop.io.Text
+import org.apache.hadoop.io.Writable
 import org.apache.hadoop.mapred.FileInputFormat
+import org.apache.hadoop.mapred.InputFormat
 import org.apache.hadoop.mapred.JobConf
+import org.apache.hadoop.mapred.SequenceFileInputFormat
 import org.apache.hadoop.mapred.TextInputFormat
 import org.apache.hadoop.mapreduce.{InputFormat => NewInputFormat}
-import org.apache.hadoop.mapreduce.lib.input.{FileInputFormat => NewFileInputFormat}
 import org.apache.hadoop.mapreduce.{Job => NewHadoopJob}
+import org.apache.hadoop.mapreduce.lib.input.{FileInputFormat => NewFileInputFormat}
+
 import org.apache.mesos.MesosNativeLibrary
 
-import spark.deploy.{SparkHadoopUtil, LocalSparkCluster}
-import spark.partial.ApproximateEvaluator
-import spark.partial.PartialResult
+import spark.deploy.{LocalSparkCluster, SparkHadoopUtil}
+import spark.partial.{ApproximateEvaluator, PartialResult}
 import spark.rdd.{CheckpointRDD, HadoopRDD, NewHadoopRDD, UnionRDD, ParallelCollectionRDD}
-import spark.scheduler._
+import spark.scheduler.{DAGScheduler, ResultTask, ShuffleMapTask, SparkListener, SplitInfo, Stage, StageInfo, TaskScheduler}
+import spark.scheduler.cluster.{StandaloneSchedulerBackend, SparkDeploySchedulerBackend, ClusterScheduler}
 import spark.scheduler.local.LocalScheduler
-import spark.scheduler.cluster.{StandaloneSchedulerBackend, SparkDeploySchedulerBackend, SchedulerBackend, ClusterScheduler}
 import spark.scheduler.mesos.{CoarseMesosSchedulerBackend, MesosSchedulerBackend}
-import spark.storage.BlockManagerUI
+import spark.storage.{BlockManagerUI, StorageStatus, StorageUtils, RDDInfo}
 import spark.util.{MetadataCleaner, TimeStampedHashMap}
-import spark.storage.{StorageStatus, StorageUtils, RDDInfo}
+
 
 /**
  * Main entry point for Spark functionality. A SparkContext represents the connection to a Spark
@@ -97,7 +100,7 @@ class SparkContext(
   private[spark] val addedJars = HashMap[String, Long]()
 
   // Keeps track of all persisted RDDs
-  private[spark] val persistentRdds = new TimeStampedHashMap[Int, RDD[_]]()
+  private[spark] val persistentRdds = new TimeStampedHashMap[Int, RDD[_]]
   private[spark] val metadataCleaner = new MetadataCleaner("SparkContext", this.cleanup)
 
 
@@ -505,7 +508,7 @@ class SparkContext(
    * Return information about what RDDs are cached, if they are in mem or on disk, how much space
    * they take, etc.
    */
-  def getRDDStorageInfo : Array[RDDInfo] = {
+  def getRDDStorageInfo: Array[RDDInfo] = {
     StorageUtils.rddInfoFromStorageStatus(getExecutorStorageStatus, this)
   }
 
@@ -516,7 +519,7 @@ class SparkContext(
   /**
    * Return information about blocks stored in all of the slaves
    */
-  def getExecutorStorageStatus : Array[StorageStatus] = {
+  def getExecutorStorageStatus: Array[StorageStatus] = {
     env.blockManager.master.getStorageStatus
   }
 
diff --git a/core/src/main/scala/spark/api/java/JavaRDD.scala b/core/src/main/scala/spark/api/java/JavaRDD.scala
index e29f1e5899..eb81ed64cd 100644
--- a/core/src/main/scala/spark/api/java/JavaRDD.scala
+++ b/core/src/main/scala/spark/api/java/JavaRDD.scala
@@ -14,12 +14,18 @@ JavaRDDLike[T, JavaRDD[T]] {
   /** Persist this RDD with the default storage level (`MEMORY_ONLY`). */
   def cache(): JavaRDD[T] = wrapRDD(rdd.cache())
 
-  /** 
+  /**
    * Set this RDD's storage level to persist its values across operations after the first time
-   * it is computed. Can only be called once on each RDD.
+   * it is computed. This can only be used to assign a new storage level if the RDD does not
+   * have a storage level set yet..
    */
   def persist(newLevel: StorageLevel): JavaRDD[T] = wrapRDD(rdd.persist(newLevel))
 
+  /**
+   * Mark the RDD as non-persistent, and remove all blocks for it from memory and disk.
+   */
+  def unpersist(): JavaRDD[T] = wrapRDD(rdd.unpersist())
+
   // Transformations (return a new RDD)
 
   /**
@@ -31,7 +37,7 @@ JavaRDDLike[T, JavaRDD[T]] {
    * Return a new RDD containing the distinct elements in this RDD.
    */
   def distinct(numPartitions: Int): JavaRDD[T] = wrapRDD(rdd.distinct(numPartitions))
-  
+
   /**
    * Return a new RDD containing only the elements that satisfy a predicate.
    */
@@ -54,7 +60,7 @@ JavaRDDLike[T, JavaRDD[T]] {
    */
   def sample(withReplacement: Boolean, fraction: Double, seed: Int): JavaRDD[T] =
     wrapRDD(rdd.sample(withReplacement, fraction, seed))
-    
+
   /**
    * Return the union of this RDD and another one. Any identical elements will appear multiple
    * times (use `.distinct()` to eliminate them).
@@ -63,7 +69,7 @@ JavaRDDLike[T, JavaRDD[T]] {
 
   /**
    * Return an RDD with the elements from `this` that are not in `other`.
-   * 
+   *
    * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting
    * RDD will be <= us.
    */
diff --git a/core/src/main/scala/spark/network/ConnectionManager.scala b/core/src/main/scala/spark/network/ConnectionManager.scala
index 0c6bdb1559..0eb03630d0 100644
--- a/core/src/main/scala/spark/network/ConnectionManager.scala
+++ b/core/src/main/scala/spark/network/ConnectionManager.scala
@@ -188,6 +188,38 @@ private[spark] class ConnectionManager(port: Int) extends Logging {
     } )
   }
 
+  // MUST be called within selector loop - else deadlock.
+  private def triggerForceCloseByException(key: SelectionKey, e: Exception) {
+    try {
+      key.interestOps(0)
+    } catch {
+      // ignore exceptions
+      case e: Exception => logDebug("Ignoring exception", e)
+    }
+
+    val conn = connectionsByKey.getOrElse(key, null)
+    if (conn == null) return
+
+    // Pushing to connect threadpool
+    handleConnectExecutor.execute(new Runnable {
+      override def run() {
+        try {
+          conn.callOnExceptionCallback(e)
+        } catch {
+          // ignore exceptions
+          case e: Exception => logDebug("Ignoring exception", e)
+        }
+        try {
+          conn.close()
+        } catch {
+          // ignore exceptions
+          case e: Exception => logDebug("Ignoring exception", e)
+        }
+      }
+    })
+  }
+
+
   def run() {
     try {
       while(!selectorThread.isInterrupted) {
@@ -200,29 +232,76 @@ private[spark] class ConnectionManager(port: Int) extends Logging {
 
         while(!keyInterestChangeRequests.isEmpty) {
           val (key, ops) = keyInterestChangeRequests.dequeue
-          val connection = connectionsByKey.getOrElse(key, null)
-          if (connection != null) {
-            val lastOps = key.interestOps()
-            key.interestOps(ops)
-
-            // hot loop - prevent materialization of string if trace not enabled.
-            if (isTraceEnabled()) {
-              def intToOpStr(op: Int): String = {
-                val opStrs = ArrayBuffer[String]()
-                if ((op & SelectionKey.OP_READ) != 0) opStrs += "READ"
-                if ((op & SelectionKey.OP_WRITE) != 0) opStrs += "WRITE"
-                if ((op & SelectionKey.OP_CONNECT) != 0) opStrs += "CONNECT"
-                if ((op & SelectionKey.OP_ACCEPT) != 0) opStrs += "ACCEPT"
-                if (opStrs.size > 0) opStrs.reduceLeft(_ + " | " + _) else " "
-              }
 
-              logTrace("Changed key for connection to [" + connection.getRemoteConnectionManagerId()  +
-                "] changed from [" + intToOpStr(lastOps) + "] to [" + intToOpStr(ops) + "]")
+          try {
+            if (key.isValid) {
+              val connection = connectionsByKey.getOrElse(key, null)
+              if (connection != null) {
+                val lastOps = key.interestOps()
+                key.interestOps(ops)
+
+                // hot loop - prevent materialization of string if trace not enabled.
+                if (isTraceEnabled()) {
+                  def intToOpStr(op: Int): String = {
+                    val opStrs = ArrayBuffer[String]()
+                    if ((op & SelectionKey.OP_READ) != 0) opStrs += "READ"
+                    if ((op & SelectionKey.OP_WRITE) != 0) opStrs += "WRITE"
+                    if ((op & SelectionKey.OP_CONNECT) != 0) opStrs += "CONNECT"
+                    if ((op & SelectionKey.OP_ACCEPT) != 0) opStrs += "ACCEPT"
+                    if (opStrs.size > 0) opStrs.reduceLeft(_ + " | " + _) else " "
+                  }
+
+                  logTrace("Changed key for connection to [" + connection.getRemoteConnectionManagerId()  +
+                    "] changed from [" + intToOpStr(lastOps) + "] to [" + intToOpStr(ops) + "]")
+                }
+              }
+            } else {
+              logInfo("Key not valid ? " + key)
+              throw new CancelledKeyException()
+            }
+          } catch {
+            case e: CancelledKeyException => {
+              logInfo("key already cancelled ? " + key, e)
+              triggerForceCloseByException(key, e)
+            }
+            case e: Exception => {
+              logError("Exception processing key " + key, e)
+              triggerForceCloseByException(key, e)
             }
           }
         }
 
-        val selectedKeysCount = selector.select()
+        val selectedKeysCount =
+          try {
+            selector.select()
+          } catch {
+            // Explicitly only dealing with CancelledKeyException here since other exceptions should be dealt with differently.
+            case e: CancelledKeyException => {
+              // Some keys within the selectors list are invalid/closed. clear them.
+              val allKeys = selector.keys().iterator()
+
+              while (allKeys.hasNext()) {
+                val key = allKeys.next()
+                try {
+                  if (! key.isValid) {
+                    logInfo("Key not valid ? " + key)
+                    throw new CancelledKeyException()
+                  }
+                } catch {
+                  case e: CancelledKeyException => {
+                    logInfo("key already cancelled ? " + key, e)
+                    triggerForceCloseByException(key, e)
+                  }
+                  case e: Exception => {
+                    logError("Exception processing key " + key, e)
+                    triggerForceCloseByException(key, e)
+                  }
+                }
+              }
+            }
+            0
+          }
+
         if (selectedKeysCount == 0) {
           logDebug("Selector selected " + selectedKeysCount + " of " + selector.keys.size + " keys")
         }
@@ -230,23 +309,40 @@ private[spark] class ConnectionManager(port: Int) extends Logging {
           logInfo("Selector thread was interrupted!")
           return
         }
-        
-        val selectedKeys = selector.selectedKeys().iterator()
-        while (selectedKeys.hasNext()) {
-          val key = selectedKeys.next
-          selectedKeys.remove()
-          if (key.isValid) {
-            if (key.isAcceptable) {
-              acceptConnection(key)
-            } else 
-            if (key.isConnectable) {
-              triggerConnect(key)
-            } else
-            if (key.isReadable) {
-              triggerRead(key)
-            } else
-            if (key.isWritable) {
-              triggerWrite(key)
+
+        if (0 != selectedKeysCount) {
+          val selectedKeys = selector.selectedKeys().iterator()
+          while (selectedKeys.hasNext()) {
+            val key = selectedKeys.next
+            selectedKeys.remove()
+            try {
+              if (key.isValid) {
+                if (key.isAcceptable) {
+                  acceptConnection(key)
+                } else
+                if (key.isConnectable) {
+                  triggerConnect(key)
+                } else
+                if (key.isReadable) {
+                  triggerRead(key)
+                } else
+                if (key.isWritable) {
+                  triggerWrite(key)
+                }
+              } else {
+                logInfo("Key not valid ? " + key)
+                throw new CancelledKeyException()
+              }
+            } catch {
+              // weird, but we saw this happening - even though key.isValid was true, key.isAcceptable would throw CancelledKeyException.
+              case e: CancelledKeyException => {
+                logInfo("key already cancelled ? " + key, e)
+                triggerForceCloseByException(key, e)
+              }
+              case e: Exception => {
+                logError("Exception processing key " + key, e)
+                triggerForceCloseByException(key, e)
+              }
             }
           }
         }
diff --git a/core/src/main/scala/spark/storage/BlockManagerMaster.scala b/core/src/main/scala/spark/storage/BlockManagerMaster.scala
index 6fae62d373..ac26c16867 100644
--- a/core/src/main/scala/spark/storage/BlockManagerMaster.scala
+++ b/core/src/main/scala/spark/storage/BlockManagerMaster.scala
@@ -15,6 +15,7 @@ import akka.util.duration._
 
 import spark.{Logging, SparkException, Utils}
 
+
 private[spark] class BlockManagerMaster(var driverActor: ActorRef) extends Logging {
 
   val AKKA_RETRY_ATTEMPTS: Int = System.getProperty("spark.akka.num.retries", "3").toInt
@@ -88,6 +89,21 @@ private[spark] class BlockManagerMaster(var driverActor: ActorRef) extends Loggi
   }
 
   /**
+   * Remove all blocks belonging to the given RDD.
+   */
+  def removeRdd(rddId: Int) {
+    val rddBlockPrefix = "rdd_" + rddId + "_"
+    // Get the list of blocks in block manager, and remove ones that are part of this RDD.
+    // The runtime complexity is linear to the number of blocks persisted in the cluster.
+    // It could be expensive if the cluster is large and has a lot of blocks persisted.
+    getStorageStatus.flatMap(_.blocks).foreach { case(blockId, status) =>
+      if (blockId.startsWith(rddBlockPrefix)) {
+        removeBlock(blockId)
+      }
+    }
+  }
+
+  /**
    * Return the memory status for each block manager, in the form of a map from
    * the block manager's id to two long values. The first value is the maximum
    * amount of memory allocated for the block manager, while the second is the
diff --git a/core/src/main/scala/spark/storage/DiskStore.scala b/core/src/main/scala/spark/storage/DiskStore.scala
index 498bc9eeb6..8154b8ca74 100644
--- a/core/src/main/scala/spark/storage/DiskStore.scala
+++ b/core/src/main/scala/spark/storage/DiskStore.scala
@@ -236,8 +236,7 @@ private class DiskStore(blockManager: BlockManager, rootDirs: String)
           localDirId = "%s-%04x".format(dateFormat.format(new Date), rand.nextInt(65536))
           localDir = new File(rootDir, "spark-local-" + localDirId)
           if (!localDir.exists) {
-            localDir.mkdirs()
-            foundLocalDir = true
+            foundLocalDir = localDir.mkdirs()
           }
         } catch {
           case e: Exception =>
diff --git a/core/src/main/scala/spark/storage/StorageUtils.scala b/core/src/main/scala/spark/storage/StorageUtils.scala
index dec47a9d41..8f52168c24 100644
--- a/core/src/main/scala/spark/storage/StorageUtils.scala
+++ b/core/src/main/scala/spark/storage/StorageUtils.scala
@@ -4,9 +4,9 @@ import spark.{Utils, SparkContext}
 import BlockManagerMasterActor.BlockStatus
 
 private[spark]
-case class StorageStatus(blockManagerId: BlockManagerId, maxMem: Long, 
+case class StorageStatus(blockManagerId: BlockManagerId, maxMem: Long,
   blocks: Map[String, BlockStatus]) {
-  
+
   def memUsed(blockPrefix: String = "") = {
     blocks.filterKeys(_.startsWith(blockPrefix)).values.map(_.memSize).
       reduceOption(_+_).getOrElse(0l)
@@ -22,35 +22,40 @@ case class StorageStatus(blockManagerId: BlockManagerId, maxMem: Long,
 }
 
 case class RDDInfo(id: Int, name: String, storageLevel: StorageLevel,
-  numCachedPartitions: Int, numPartitions: Int, memSize: Long, diskSize: Long) {
+  numCachedPartitions: Int, numPartitions: Int, memSize: Long, diskSize: Long)
+  extends Ordered[RDDInfo] {
   override def toString = {
     import Utils.memoryBytesToString
     "RDD \"%s\" (%d) Storage: %s; CachedPartitions: %d; TotalPartitions: %d; MemorySize: %s; DiskSize: %s".format(name, id,
       storageLevel.toString, numCachedPartitions, numPartitions, memoryBytesToString(memSize), memoryBytesToString(diskSize))
   }
+
+  override def compare(that: RDDInfo) = {
+    this.id - that.id
+  }
 }
 
 /* Helper methods for storage-related objects */
 private[spark]
 object StorageUtils {
 
-  /* Given the current storage status of the BlockManager, returns information for each RDD */ 
-  def rddInfoFromStorageStatus(storageStatusList: Array[StorageStatus], 
+  /* Given the current storage status of the BlockManager, returns information for each RDD */
+  def rddInfoFromStorageStatus(storageStatusList: Array[StorageStatus],
     sc: SparkContext) : Array[RDDInfo] = {
-    rddInfoFromBlockStatusList(storageStatusList.flatMap(_.blocks).toMap, sc) 
+    rddInfoFromBlockStatusList(storageStatusList.flatMap(_.blocks).toMap, sc)
   }
 
-  /* Given a list of BlockStatus objets, returns information for each RDD */ 
-  def rddInfoFromBlockStatusList(infos: Map[String, BlockStatus], 
+  /* Given a list of BlockStatus objets, returns information for each RDD */
+  def rddInfoFromBlockStatusList(infos: Map[String, BlockStatus],
     sc: SparkContext) : Array[RDDInfo] = {
 
     // Group by rddId, ignore the partition name
-    val groupedRddBlocks = infos.groupBy { case(k, v) =>
+    val groupedRddBlocks = infos.filterKeys(_.startsWith("rdd_")).groupBy { case(k, v) =>
       k.substring(0,k.lastIndexOf('_'))
     }.mapValues(_.values.toArray)
 
     // For each RDD, generate an RDDInfo object
-    groupedRddBlocks.map { case(rddKey, rddBlocks) =>
+    val rddInfos = groupedRddBlocks.map { case(rddKey, rddBlocks) =>
 
       // Add up memory and disk sizes
       val memSize = rddBlocks.map(_.memSize).reduce(_ + _)
@@ -65,10 +70,14 @@ object StorageUtils {
 
       RDDInfo(rddId, rddName, rddStorageLevel, rddBlocks.length, rdd.partitions.size, memSize, diskSize)
     }.toArray
+
+    scala.util.Sorting.quickSort(rddInfos)
+
+    rddInfos
   }
 
-  /* Removes all BlockStatus object that are not part of a block prefix */ 
-  def filterStorageStatusByPrefix(storageStatusList: Array[StorageStatus], 
+  /* Removes all BlockStatus object that are not part of a block prefix */
+  def filterStorageStatusByPrefix(storageStatusList: Array[StorageStatus],
     prefix: String) : Array[StorageStatus] = {
 
     storageStatusList.map { status =>