Spark accumulableCollection does not work with mut

Question

I am using Spark to do employee record accumulation and for that I use Spark's accumulator. I am using Map[empId, emp] as accumulableCollection so that I can search employee by their ids. I have tried everything but it does not work. Can someone point if there is any logical issues with the way I am using accumulableCollection or Map is not supported. Following is my code

package demo

import org.apache.spark.{SparkContext, SparkConf, Logging}

import org.apache.spark.SparkContext._
import scala.collection.mutable


object MapAccuApp extends App with Logging {
  case class Employee(id:String, name:String, dept:String)

  val conf = new SparkConf().setAppName("Employees") setMaster ("local[4]")
  val sc = new SparkContext(conf)

  implicit def empMapToSet(empIdToEmp: mutable.Map[String, Employee]): mutable.MutableList[Employee] = {
    empIdToEmp.foldLeft(mutable.MutableList[Employee]()) { (l, e) => l += e._2}
  }

  val empAccu = sc.accumulableCollection[mutable.Map[String, Employee], Employee](mutable.Map[String,Employee]())

  val employees = List(
    Employee("10001", "Tom", "Eng"),
    Employee("10002", "Roger", "Sales"),
    Employee("10003", "Rafael", "Sales"),
    Employee("10004", "David", "Sales"),
    Employee("10005", "Moore", "Sales"),
    Employee("10006", "Dawn", "Sales"),
    Employee("10007", "Stud", "Marketing"),
    Employee("10008", "Brown", "QA")
  )

  System.out.println("employee count " + employees.size)


  sc.parallelize(employees).foreach(e => {
    empAccu += e
  })

  System.out.println("empAccumulator size " + empAccu.value.size)
}

Answer 1

Using accumulableCollection seems like overkill for your problem, as the following demonstrates:

import org.apache.spark.{AccumulableParam, Accumulable, SparkContext, SparkConf}

import scala.collection.mutable

case class Employee(id:String, name:String, dept:String)

val conf = new SparkConf().setAppName("Employees") setMaster ("local[4]")
val sc = new SparkContext(conf)

implicit def mapAccum =
    new AccumulableParam[mutable.Map[String,Employee], Employee]
{
  def addInPlace(t1: mutable.Map[String,Employee],
                 t2: mutable.Map[String,Employee])
      : mutable.Map[String,Employee] = {
    t1 ++= t2
    t1
  }
  def addAccumulator(t1: mutable.Map[String,Employee], e: Employee)
      : mutable.Map[String,Employee] = {
    t1 += (e.id -> e)
    t1
  }
  def zero(t: mutable.Map[String,Employee])
      : mutable.Map[String,Employee] = {
    mutable.Map[String,Employee]()
  }
}

val empAccu = sc.accumulable(mutable.Map[String,Employee]())

val employees = List(
  Employee("10001", "Tom", "Eng"),
  Employee("10002", "Roger", "Sales"),
  Employee("10003", "Rafael", "Sales"),
  Employee("10004", "David", "Sales"),
  Employee("10005", "Moore", "Sales"),
  Employee("10006", "Dawn", "Sales"),
  Employee("10007", "Stud", "Marketing"),
  Employee("10008", "Brown", "QA")
)

System.out.println("employee count " + employees.size)

sc.parallelize(employees).foreach(e => {
  empAccu += e
})

println("empAccumulator size " + empAccu.value.size)
empAccu.value.foreach(entry =>
  println("emp id = " + entry._1 + " name = " + entry._2.name))

While this is poorly documented right now, the relevant test in the Spark codebase is quite illuminating.

Edit: It turns out that using accumulableCollection does have value: you don't need to define an AccumulableParam and the following works. I'm leaving both solutions in case they're useful to people.

case class Employee(id:String, name:String, dept:String)

val conf = new SparkConf().setAppName("Employees") setMaster ("local[4]")
val sc = new SparkContext(conf)

val empAccu = sc.accumulableCollection(mutable.HashMap[String,Employee]())

val employees = List(
  Employee("10001", "Tom", "Eng"),
  Employee("10002", "Roger", "Sales"),
  Employee("10003", "Rafael", "Sales"),
  Employee("10004", "David", "Sales"),
  Employee("10005", "Moore", "Sales"),
  Employee("10006", "Dawn", "Sales"),
  Employee("10007", "Stud", "Marketing"),
  Employee("10008", "Brown", "QA")
)

System.out.println("employee count " + employees.size)

sc.parallelize(employees).foreach(e => {
  // notice this is different from the previous solution
  empAccu += e.id -> e
})

println("empAccumulator size " + empAccu.value.size)
empAccu.value.foreach(entry =>
  println("emp id = " + entry._1 + " name = " + entry._2.name))

Both solutions tested using Spark 1.0.2.