Spark: Get top N by key

Question

Say I have a PairRDD as such (Obviously much more data in real life, assume millions of records):

val scores = sc.parallelize(Array(
      ("a", 1),  
      ("a", 2), 
      ("a", 3), 
      ("b", 3), 
      ("b", 1), 
      ("a", 4),  
      ("b", 4), 
      ("b", 2)
))

What is the most efficient way to generate a RDD with the top 2 scores per key?

val top2ByKey = ...
res3: Array[(String, Int)] = Array((a,4), (a,3), (b,4), (b,3))

Answer 1

I think this should be quite efficient:

Edited according to OP comments:

scores.mapValues(p => (p, p)).reduceByKey((u, v) => {
  val values = List(u._1, u._2, v._1, v._2).sorted(Ordering[Int].reverse).distinct
  if (values.size > 1) (values(0), values(1))
  else (values(0), values(0))
}).collect().foreach(println)

Answer 2

Since version 1.4, there is a built-in way to do this using MLLib: https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/mllib/rdd/MLPairRDDFunctions.scala

import org.apache.spark.mllib.rdd.MLPairRDDFunctions.fromPairRDD
scores.topByKey(2)