[SPARK-12004] Preserve the RDD partitioner through RDD checkpointing

The solution is the save the RDD partitioner in a separate file in the RDD checkpoint directory. That is, `<checkpoint dir>/_partitioner`.  In most cases, whether the RDD partitioner was recovered or not, does not affect the correctness, only reduces performance. So this solution makes a best-effort attempt to save and recover the partitioner. If either fails, the checkpointing is not affected. This makes this patch safe and backward compatible.

Author: Tathagata Das <[email protected]>

Closes #9983 from tdas/SPARK-12004.

Author: tdas

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

@@ -20,21 +20,22 @@ package org.apache.spark.rdd
 import java.io.IOException
 
 import scala.reflect.ClassTag
+import scala.util.control.NonFatal
 
 import org.apache.hadoop.fs.Path
 
 import org.apache.spark._
 import org.apache.spark.broadcast.Broadcast
-import org.apache.spark.deploy.SparkHadoopUtil
 import org.apache.spark.util.{SerializableConfiguration, Utils}
 
 /**
  * An RDD that reads from checkpoint files previously written to reliable storage.
  */
 private[spark] class ReliableCheckpointRDD[T: ClassTag](
     sc: SparkContext,
-    val checkpointPath: String)
-  extends CheckpointRDD[T](sc) {
+    val checkpointPath: String,
+    _partitioner: Option[Partitioner] = None
+  ) extends CheckpointRDD[T](sc) {
 
   @transient private val hadoopConf = sc.hadoopConfiguration
   @transient private val cpath = new Path(checkpointPath)
@@ -47,7 +48,13 @@ private[spark] class ReliableCheckpointRDD[T: ClassTag](
   /**
    * Return the path of the checkpoint directory this RDD reads data from.
    */
-  override def getCheckpointFile: Option[String] = Some(checkpointPath)
+  override val getCheckpointFile: Option[String] = Some(checkpointPath)
+
+  override val partitioner: Option[Partitioner] = {
+    _partitioner.orElse {
+      ReliableCheckpointRDD.readCheckpointedPartitionerFile(context, checkpointPath)
+    }
+  }
 
   /**
    * Return partitions described by the files in the checkpoint directory.
@@ -100,10 +107,52 @@ private[spark] object ReliableCheckpointRDD extends Logging {
     "part-%05d".format(partitionIndex)
   }
 
+  private def checkpointPartitionerFileName(): String = {
+    "_partitioner"
+  }
+
+  /**
+   * Write RDD to checkpoint files and return a ReliableCheckpointRDD representing the RDD.
+   */
+  def writeRDDToCheckpointDirectory[T: ClassTag](
+      originalRDD: RDD[T],
+      checkpointDir: String,
+      blockSize: Int = -1): ReliableCheckpointRDD[T] = {
+
+    val sc = originalRDD.sparkContext
+
+    // Create the output path for the checkpoint
+    val checkpointDirPath = new Path(checkpointDir)
+    val fs = checkpointDirPath.getFileSystem(sc.hadoopConfiguration)
+    if (!fs.mkdirs(checkpointDirPath)) {
+      throw new SparkException(s"Failed to create checkpoint path $checkpointDirPath")
+    }
+
+    // Save to file, and reload it as an RDD
+    val broadcastedConf = sc.broadcast(
+      new SerializableConfiguration(sc.hadoopConfiguration))
+    // TODO: This is expensive because it computes the RDD again unnecessarily (SPARK-8582)
+    sc.runJob(originalRDD,
+      writePartitionToCheckpointFile[T](checkpointDirPath.toString, broadcastedConf) _)
+
+    if (originalRDD.partitioner.nonEmpty) {
+      writePartitionerToCheckpointDir(sc, originalRDD.partitioner.get, checkpointDirPath)
+    }
+
+    val newRDD = new ReliableCheckpointRDD[T](
+      sc, checkpointDirPath.toString, originalRDD.partitioner)
+    if (newRDD.partitions.length != originalRDD.partitions.length) {
+      throw new SparkException(
+        s"Checkpoint RDD $newRDD(${newRDD.partitions.length}) has different " +
+          s"number of partitions from original RDD $originalRDD(${originalRDD.partitions.length})")
+    }
+    newRDD
+  }
+
   /**
-   * Write this partition's values to a checkpoint file.
+   * Write a RDD partition's data to a checkpoint file.
    */
-  def writeCheckpointFile[T: ClassTag](
+  def writePartitionToCheckpointFile[T: ClassTag](
       path: String,
       broadcastedConf: Broadcast[SerializableConfiguration],
       blockSize: Int = -1)(ctx: TaskContext, iterator: Iterator[T]) {
@@ -151,6 +200,67 @@ private[spark] object ReliableCheckpointRDD extends Logging {
     }
   }
 
+  /**
+   * Write a partitioner to the given RDD checkpoint directory. This is done on a best-effort
+   * basis; any exception while writing the partitioner is caught, logged and ignored.
+   */
+  private def writePartitionerToCheckpointDir(
+    sc: SparkContext, partitioner: Partitioner, checkpointDirPath: Path): Unit = {
+    try {
+      val partitionerFilePath = new Path(checkpointDirPath, checkpointPartitionerFileName)
+      val bufferSize = sc.conf.getInt("spark.buffer.size", 65536)
+      val fs = partitionerFilePath.getFileSystem(sc.hadoopConfiguration)
+      val fileOutputStream = fs.create(partitionerFilePath, false, bufferSize)
+      val serializer = SparkEnv.get.serializer.newInstance()
+      val serializeStream = serializer.serializeStream(fileOutputStream)
+      Utils.tryWithSafeFinally {
+        serializeStream.writeObject(partitioner)
+      } {
+        serializeStream.close()
+      }
+      logDebug(s"Written partitioner to $partitionerFilePath")
+    } catch {
+      case NonFatal(e) =>
+        logWarning(s"Error writing partitioner $partitioner to $checkpointDirPath")
+    }
+  }
+
+
+  /**
+   * Read a partitioner from the given RDD checkpoint directory, if it exists.
+   * This is done on a best-effort basis; any exception while reading the partitioner is
+   * caught, logged and ignored.
+   */
+  private def readCheckpointedPartitionerFile(
+      sc: SparkContext,
+      checkpointDirPath: String): Option[Partitioner] = {
+    try {
+      val bufferSize = sc.conf.getInt("spark.buffer.size", 65536)
+      val partitionerFilePath = new Path(checkpointDirPath, checkpointPartitionerFileName)
+      val fs = partitionerFilePath.getFileSystem(sc.hadoopConfiguration)
+      if (fs.exists(partitionerFilePath)) {
+        val fileInputStream = fs.open(partitionerFilePath, bufferSize)
+        val serializer = SparkEnv.get.serializer.newInstance()
+        val deserializeStream = serializer.deserializeStream(fileInputStream)
+        val partitioner = Utils.tryWithSafeFinally[Partitioner] {
+          deserializeStream.readObject[Partitioner]
+        } {
+          deserializeStream.close()
+        }
+        logDebug(s"Read partitioner from $partitionerFilePath")
+        Some(partitioner)
+      } else {
+        logDebug("No partitioner file")
+        None
+      }
+    } catch {
+      case NonFatal(e) =>
+        logWarning(s"Error reading partitioner from $checkpointDirPath, " +
+            s"partitioner will not be recovered which may lead to performance loss", e)
+        None
+    }
+  }
+
   /**
    * Read the content of the specified checkpoint file.
    */

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

@@ -55,25 +55,7 @@ private[spark] class ReliableRDDCheckpointData[T: ClassTag](@transient private v
    * This is called immediately after the first action invoked on this RDD has completed.
    */
   protected override def doCheckpoint(): CheckpointRDD[T] = {
-
-    // Create the output path for the checkpoint
-    val path = new Path(cpDir)
-    val fs = path.getFileSystem(rdd.context.hadoopConfiguration)
-    if (!fs.mkdirs(path)) {
-      throw new SparkException(s"Failed to create checkpoint path $cpDir")
-    }
-
-    // Save to file, and reload it as an RDD
-    val broadcastedConf = rdd.context.broadcast(
-      new SerializableConfiguration(rdd.context.hadoopConfiguration))
-    // TODO: This is expensive because it computes the RDD again unnecessarily (SPARK-8582)
-    rdd.context.runJob(rdd, ReliableCheckpointRDD.writeCheckpointFile[T](cpDir, broadcastedConf) _)
-    val newRDD = new ReliableCheckpointRDD[T](rdd.context, cpDir)
-    if (newRDD.partitions.length != rdd.partitions.length) {
-      throw new SparkException(
-        s"Checkpoint RDD $newRDD(${newRDD.partitions.length}) has different " +
-          s"number of partitions from original RDD $rdd(${rdd.partitions.length})")
-    }
+    val newRDD = ReliableCheckpointRDD.writeRDDToCheckpointDirectory(rdd, cpDir)
 
     // Optionally clean our checkpoint files if the reference is out of scope
     if (rdd.conf.getBoolean("spark.cleaner.referenceTracking.cleanCheckpoints", false)) {
@@ -83,7 +65,6 @@ private[spark] class ReliableRDDCheckpointData[T: ClassTag](@transient private v
     }
 
     logInfo(s"Done checkpointing RDD ${rdd.id} to $cpDir, new parent is RDD ${newRDD.id}")
-
     newRDD
   }
 

core/src/test/scala/org/apache/spark/CheckpointSuite.scala

@@ -21,7 +21,8 @@ import java.io.File
 
 import scala.reflect.ClassTag
 
-import org.apache.spark.CheckpointSuite._
+import org.apache.hadoop.fs.Path
+
 import org.apache.spark.rdd._
 import org.apache.spark.storage.{BlockId, StorageLevel, TestBlockId}
 import org.apache.spark.util.Utils
@@ -74,8 +75,10 @@ trait RDDCheckpointTester { self: SparkFunSuite =>
 
     // Test whether the checkpoint file has been created
     if (reliableCheckpoint) {
-      assert(
-        collectFunc(sparkContext.checkpointFile[U](operatedRDD.getCheckpointFile.get)) === result)
+      assert(operatedRDD.getCheckpointFile.nonEmpty)
+      val recoveredRDD = sparkContext.checkpointFile[U](operatedRDD.getCheckpointFile.get)
+      assert(collectFunc(recoveredRDD) === result)
+      assert(recoveredRDD.partitioner === operatedRDD.partitioner)
     }
 
     // Test whether dependencies have been changed from its earlier parent RDD
@@ -211,9 +214,14 @@ trait RDDCheckpointTester { self: SparkFunSuite =>
   }
 
   /** Run a test twice, once for local checkpointing and once for reliable checkpointing. */
-  protected def runTest(name: String)(body: Boolean => Unit): Unit = {
+  protected def runTest(
+      name: String,
+      skipLocalCheckpoint: Boolean = false
+    )(body: Boolean => Unit): Unit = {
     test(name + " [reliable checkpoint]")(body(true))
-    test(name + " [local checkpoint]")(body(false))
+    if (!skipLocalCheckpoint) {
+      test(name + " [local checkpoint]")(body(false))
+    }
   }
 
   /**
@@ -264,6 +272,49 @@ class CheckpointSuite extends SparkFunSuite with RDDCheckpointTester with LocalS
     assert(flatMappedRDD.collect() === result)
   }
 
+  runTest("checkpointing partitioners", skipLocalCheckpoint = true) { _: Boolean =>
+
+    def testPartitionerCheckpointing(
+        partitioner: Partitioner,
+        corruptPartitionerFile: Boolean = false
+      ): Unit = {
+      val rddWithPartitioner = sc.makeRDD(1 to 4).map { _ -> 1 }.partitionBy(partitioner)
+      rddWithPartitioner.checkpoint()
+      rddWithPartitioner.count()
+      assert(rddWithPartitioner.getCheckpointFile.get.nonEmpty,
+        "checkpointing was not successful")
+
+      if (corruptPartitionerFile) {
+        // Overwrite the partitioner file with garbage data
+        val checkpointDir = new Path(rddWithPartitioner.getCheckpointFile.get)
+        val fs = checkpointDir.getFileSystem(sc.hadoopConfiguration)
+        val partitionerFile = fs.listStatus(checkpointDir)
+            .find(_.getPath.getName.contains("partitioner"))
+            .map(_.getPath)
+        require(partitionerFile.nonEmpty, "could not find the partitioner file for testing")
+        val output = fs.create(partitionerFile.get, true)
+        output.write(100)
+        output.close()
+      }
+
+      val newRDD = sc.checkpointFile[(Int, Int)](rddWithPartitioner.getCheckpointFile.get)
+      assert(newRDD.collect().toSet === rddWithPartitioner.collect().toSet, "RDD not recovered")
+
+      if (!corruptPartitionerFile) {
+        assert(newRDD.partitioner != None, "partitioner not recovered")
+        assert(newRDD.partitioner === rddWithPartitioner.partitioner,
+          "recovered partitioner does not match")
+      } else {
+        assert(newRDD.partitioner == None, "partitioner unexpectedly recovered")
+      }
+    }
+
+    testPartitionerCheckpointing(partitioner)
+
+    // Test that corrupted partitioner file does not prevent recovery of RDD
+    testPartitionerCheckpointing(partitioner, corruptPartitionerFile = true)
+  }
+
   runTest("RDDs with one-to-one dependencies") { reliableCheckpoint: Boolean =>
     testRDD(_.map(x => x.toString), reliableCheckpoint)
     testRDD(_.flatMap(x => 1 to x), reliableCheckpoint)