Spark RDD checkpoint on S3 exits with exception intermittently

  /**
   * 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 checkpointStartTimeNs = System.nanoTime()

    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) _)  //<<<----- Writing checkpoint file

    if (originalRDD.partitioner.nonEmpty) {
      writePartitionerToCheckpointDir(sc, originalRDD.partitioner.get, checkpointDirPath)
    }

    val checkpointDurationMs =
      TimeUnit.NANOSECONDS.toMillis(System.nanoTime() - checkpointStartTimeNs)
    logInfo(s"Checkpointing took $checkpointDurationMs ms.")

    val newRDD = new ReliableCheckpointRDD[T](   //<---- Here it will reload the checkpoint file
      sc, checkpointDirPath.toString, originalRDD.partitioner)  
    if (newRDD.partitions.length != originalRDD.partitions.length) {
      throw new SparkException(
        "Checkpoint RDD has a different number of partitions from original RDD. Original " +
          s"RDD [ID: ${originalRDD.id}, num of partitions: ${originalRDD.partitions.length}]; " +
          s"Checkpoint RDD [ID: ${newRDD.id}, num of partitions: " +
          s"${newRDD.partitions.length}].")
    }
    newRDD
}