使用Clojure / Flambo在Spark中进行Secondar排序

Question

I have a scala program in which I have implemented a secondary sort which works perfectly. The way I have written that program is:

object rfmc {
  // Custom Key and partitioner

  case class RFMCKey(cId: String, R: Double, F: Double, M: Double, C: Double)
  class RFMCPartitioner(partitions: Int) extends Partitioner {
    require(partitions >= 0, "Number of partitions ($partitions) cannot be negative.")
    override def numPartitions: Int = partitions
    override def getPartition(key: Any): Int = {
      val k = key.asInstanceOf[RFMCKey]
      k.cId.hashCode() % numPartitions
    }
  }
  object RFMCKey {
    implicit def orderingBycId[A <: RFMCKey] : Ordering[A] = {
      Ordering.by(k => (k.R, k.F * -1, k.M * -1, k.C * -1))
    }
  }
  // The body of the code
  //
  //
  val x = rdd.map(RFMCKey(cust,r,f,m,c), r+","+f+","+m+","+c)
  val y = x.repartitionAndSortWithinPartitions(new RFMCPartitioner(1))
}

I wanted to implement the same thing using clojure's DSL for spark called flambo. Since I can't write partitioner using clojure, I re-used the code defind above, compiled it and used it as a dependency in my Clojure code.

Now I am importing the partitioner and the key in my clojure code the following way:

(ns xyz
  (:import
    [package RFMCPartitioner]
    [package RFMCKey]
    )
  )

But when I try to create RFMCKey by doing (RFMCKey. cust_id rfmc) , it throws the following error:

java.lang.ClassCastException: org.formcept.wisdom.RFMCKey cannot be cast to java.lang.Comparable
    at org.spark-project.guava.collect.NaturalOrdering.compare(NaturalOrdering.java:28)
    at scala.math.LowPriorityOrderingImplicits$$anon$7.compare(Ordering.scala:153)
    at org.apache.spark.util.collection.ExternalSorter$$anon$8.compare(ExternalSorter.scala:170)
    at org.apache.spark.util.collection.ExternalSorter$$anon$8.compare(ExternalSorter.scala:164)
    at org.apache.spark.util.collection.TimSort.countRunAndMakeAscending(TimSort.java:252)
    at org.apache.spark.util.collection.TimSort.sort(TimSort.java:110)
    at org.apache.spark.util.collection.Sorter.sort(Sorter.scala:37)
    at org.apache.spark.util.collection.SizeTrackingPairBuffer.destructiveSortedIterator(SizeTrackingPairBuffer.scala:83)
    at org.apache.spark.util.collection.ExternalSorter.partitionedIterator(ExternalSorter.scala:687)
    at org.apache.spark.util.collection.ExternalSorter.iterator(ExternalSorter.scala:705)
    at org.apache.spark.shuffle.hash.HashShuffleReader.read(HashShuffleReader.scala:64)
    at org.apache.spark.rdd.ShuffledRDD.compute(ShuffledRDD.scala:92)
    at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:277)
    at org.apache.spark.CacheManager.getOrCompute(CacheManager.scala:70)
    at org.apache.spark.rdd.RDD.iterator(RDD.scala:242)
    at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:35)
    at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:277)
    at org.apache.spark.rdd.RDD.iterator(RDD.scala:244)
    at org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:61)
    at org.apache.spark.scheduler.Task.run(Task.scala:64)
    at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:203)
    at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
    at java.lang.Thread.run(Thread.java:745)

My guess is that its not able to find the ordering that I have defined after the partitioner. But if it works in Scala, why doesn't it work in Clojure?

Answer 1

So I finally figured it out on my own. I had to basically write my custom ordering function as a separate scala project and then call that in clojure.

I had my scala file written in this manner:

import org.apache.spark.Partitioner
import org.apache.spark.rdd.RDD

case class RFMCKey(cId: String, R: Double, F: Long, M: Double, C: Double)
class RFMCPartitioner(partitions: Int) extends Partitioner {
  require(partitions >= 0, "Number of partitions ($partitions) cannot be negative.")
  override def numPartitions: Int = partitions
  override def getPartition(key: Any): Int = {
    val k = key.asInstanceOf[RFMCKey]
    k.cId.hashCode() % numPartitions
  }
}
object RFMCKey {
  implicit def orderingBycId[A <: RFMCKey] : Ordering[A] = {
    Ordering.by(k => (k.R, k.F * -1, k.M * -1, k.C * -1))
  }
}

class rfmcSort {
  def sortWithRFMC(a: RDD[(String, (((Double, Long), Double), Double))], parts: Int): RDD[(RFMCKey, String)] = {
    val x = a.map(v => v match {
                case (custId, (((rVal, fVal), mVal),cVal)) => (RFMCKey(custId, rVal, fVal, mVal, cVal), rVal+","+fVal+","+mVal+","+cVal)
            }).repartitionAndSortWithinPartitions(new RFMCPartitioner(parts))
    x
  }
}

I compiled it as ascala project and used it in my clojure code this way:

(:import [org.formcept.wisdom rfmcSort]
         [org.apache.spark.rdd.RDD])

sorted-rfmc-records (.toJavaRDD (.sortWithRFMC (rfmcSort.) (.rdd rfmc-records) num_partitions))

Please notice the way I am calling the sortWithRFMC function from the rfmcSort object that I created. Also one very important thing to note here is when you pass your JavaPairRDD to your scala function, you have to convert it into a normal spark RDD first by calling the .rdd method on it. And then you have to convert the spark RDD back to JavaPairRDD to work with it in clojure.