[SPARK-18916] Possible bug in Pregel / mergeMsg with hashmaps

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Not A Problem
• Affects Version/s: 2.0.2
• Component/s: GraphX
• Labels: error graphx pregel

Environment

OSX / IntelliJ IDEA 2016.3 CE EAP, Scala 2.11.8, Spark 2.0.2

• Assignee: Unassigned
• Reporter: Seth Bromberger
• Created: 18/Dec/16 03:56
• Updated: 23/Dec/16 13:29
• Resolved: 23/Dec/16 13:29

Description

Consider the following (rough) code that attempts to calculate all-pairs shortest paths via pregel:

    def allPairsShortestPaths: RDD[(VertexId, HashMap[VertexId, ParentDist])] = {
      val initialMsg = HashMap(-1L -> ParentDist(-1L, -1L))
      val pregelg = g.mapVertices((vid, vd) => (vd, HashMap[VertexId, ParentDist](vid -> ParentDist(vid, 0L)))).reverse

      def vprog(v: VertexId, value: (VD, HashMap[VertexId, ParentDist]), message: HashMap[VertexId, ParentDist]): (VD, HashMap[VertexId, ParentDist]) = {
        val updatedValues = mm2(value._2, message).filter(v => v._2.dist >= 0)
        (value._1, updatedValues)
      }

      def sendMsg(triplet: EdgeTriplet[(VD, HashMap[VertexId, ParentDist]), ED]): Iterator[(VertexId, HashMap[VertexId, ParentDist])] = {
        val dstVertexId = triplet.dstId
        val srcMap = triplet.srcAttr._2
        val dstMap = triplet.dstAttr._2  // guaranteed to have dstVertexId as a key

        val updatesToSend : HashMap[VertexId, ParentDist] = srcMap.filter {
          case (vid, srcPD) => dstMap.get(vid) match {
            case Some(dstPD) => dstPD.dist > srcPD.dist + 1   // if it exists, is it cheaper?
            case _ => true // not found - new update
          }
        }.map(u => u._1 -> ParentDist(triplet.srcId, u._2.dist +1))


        if (updatesToSend.nonEmpty)
          Iterator[(VertexId, HashMap[VertexId, ParentDist])]((dstVertexId, updatesToSend))
        else
          Iterator.empty
      }

      def mergeMsg(m1: HashMap[VertexId, ParentDist], m2: HashMap[VertexId, ParentDist]): HashMap[VertexId, ParentDist] = {

        // when the following two lines are commented out, the program fails with
        // 16/12/17 19:53:50 INFO DAGScheduler: Job 24 failed: reduce at VertexRDDImpl.scala:88, took 0.244042 s
        // Exception in thread "main" org.apache.spark.SparkException: Job aborted due to stage failure: Task 0 in stage 1099.0 failed 1 times, most recent failure: Lost task 0.0 in stage 1099.0 (TID 129, localhost): scala.MatchError: (null,null) (of class scala.Tuple2)

        m1.foreach(_ => ())
        m2.foreach(_ => ())

        m1.merged(m2) {
          case ((k1, v1), (_, v2)) => (k1, v1.min(v2))
        }
      }

      // mm2 is here just to provide a separate function for vprog. Ideally we'd just re-use mergeMsg.
      def mm2(m1: HashMap[VertexId, ParentDist], m2: HashMap[VertexId, ParentDist]): HashMap[VertexId, ParentDist] = {
        m1.merged(m2) {
          case ((k1, v1), (_, v2)) => (k1, v1.min(v2))
          case n => throw new Exception("we've got a problem: " + n)
        }
      }

      val pregelRun = pregelg.pregel(initialMsg)(vprog, sendMsg, mergeMsg)
      val sps = pregelRun.vertices.map(v => v._1 -> v._2._2)
      sps
    }
  }

Note the comment in the mergeMsg function: when the messages are explicitly accessed prior to the .merged statement, the code works. If these side-effect statements are removed / commented out, the error message in the comments is generated.

This fails consistently on a 50-node undirected cyclegraph.