[SPARK-3803] ArrayIndexOutOfBoundsException found in executing computePrincipalComponents

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.1.0
• Fix Version/s: 1.2.0
• Component/s: MLlib
• Assignee: Sean R. Owen
• Reporter: Masaru Dobashi
• Created: 05/Oct/14 11:27
• Updated: 15/Jan/15 09:08
• Resolved: 14/Oct/14 21:42

Description

When I executed computePrincipalComponents method of RowMatrix, I got java.lang.ArrayIndexOutOfBoundsException.

14/10/05 20:16:31 INFO DAGScheduler: Failed to run reduce at RDDFunctions.scala:111
org.apache.spark.SparkException: Job aborted due to stage failure: Task 0 in stage 31.0 failed 1 times, most recent failure: Lost task 0.0 in stage 31.0 (TID 611, localhost): java.lang.ArrayIndexOutOfBoundsException: 4878161
        org.apache.spark.mllib.linalg.distributed.RowMatrix$.org$apache$spark$mllib$linalg$distributed$RowMatrix$$dspr(RowMatrix.scala:460)
        org.apache.spark.mllib.linalg.distributed.RowMatrix$$anonfun$3.apply(RowMatrix.scala:114)
        org.apache.spark.mllib.linalg.distributed.RowMatrix$$anonfun$3.apply(RowMatrix.scala:113)
        scala.collection.TraversableOnce$$anonfun$foldLeft$1.apply(TraversableOnce.scala:144)
        scala.collection.TraversableOnce$$anonfun$foldLeft$1.apply(TraversableOnce.scala:144)
        scala.collection.Iterator$class.foreach(Iterator.scala:727)
        scala.collection.AbstractIterator.foreach(Iterator.scala:1157)
        scala.collection.TraversableOnce$class.foldLeft(TraversableOnce.scala:144)
        scala.collection.AbstractIterator.foldLeft(Iterator.scala:1157)
        scala.collection.TraversableOnce$class.aggregate(TraversableOnce.scala:201)
        scala.collection.AbstractIterator.aggregate(Iterator.scala:1157)
        org.apache.spark.mllib.rdd.RDDFunctions$$anonfun$4.apply(RDDFunctions.scala:99)
        org.apache.spark.mllib.rdd.RDDFunctions$$anonfun$4.apply(RDDFunctions.scala:99)
        org.apache.spark.mllib.rdd.RDDFunctions$$anonfun$5.apply(RDDFunctions.scala:100)
        org.apache.spark.mllib.rdd.RDDFunctions$$anonfun$5.apply(RDDFunctions.scala:100)
        org.apache.spark.rdd.RDD$$anonfun$13.apply(RDD.scala:596)
        org.apache.spark.rdd.RDD$$anonfun$13.apply(RDD.scala:596)
        org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:35)
        org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:262)
        org.apache.spark.rdd.RDD.iterator(RDD.scala:229)
        org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:35)
        org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:262)
        org.apache.spark.rdd.RDD.iterator(RDD.scala:229)
        org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:68)
        org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:41)
        org.apache.spark.scheduler.Task.run(Task.scala:54)
        org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:177)
        java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
        java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
        java.lang.Thread.run(Thread.java:745)

The RowMatrix instance was generated from the result of TF-IDF like the following.

scala> val hashingTF = new HashingTF()
scala> val tf = hashingTF.transform(texts)
scala> import org.apache.spark.mllib.feature.IDF
scala> tf.cache()
scala> val idf = new IDF().fit(tf)
scala> val tfidf: RDD[Vector] = idf.transform(tf)

scala> import org.apache.spark.mllib.linalg.distributed.RowMatrix
scala> val mat = new RowMatrix(tfidf)
scala> val pc = mat.computePrincipalComponents(2)

I think this was because I created HashingTF instance with default numFeatures and Array is used in RowMatrix#computeGramianMatrix method
like the following.

  /**
   * Computes the Gramian matrix `A^T A`.
   */
  def computeGramianMatrix(): Matrix = {
    val n = numCols().toInt
    val nt: Int = n * (n + 1) / 2

    // Compute the upper triangular part of the gram matrix.
    val GU = rows.treeAggregate(new BDV[Double](new Array[Double](nt)))(
      seqOp = (U, v) => {
        RowMatrix.dspr(1.0, v, U.data)
        U
      }, combOp = (U1, U2) => U1 += U2)

    RowMatrix.triuToFull(n, GU.data)
  }

When the size of Vectors generated by TF-IDF is too large, it makes "nt" to have undesirable value (and undesirable size of Array used in treeAggregate),
since n * (n + 1) / 2 exceeded Int.MaxValue.

Is this surmise correct?

And, of course, I could avoid this situation by creating instance of HashingTF with smaller numFeatures.
But this may not be fundamental solution.

Issue Links

• links to: [Github] Pull Request #2801 (srowen)