How to sort an RDD and limit in Spark?

Question

I have RDD of Foo class : class Foo( name : String, createDate : Date ). I want an other RDD with 10 percent older Foo. My first idea was to sort by createDate and limit by 0.1*count, but there is no limit function.

Have you an idea?

Answer 1

Assuming Foo is a case class like this:

import java.sql.Date
case class Foo(name: String, createDate: java.sql.Date)

1. Using plain RDDs:

   import org.apache.spark.rdd.RDD
   import scala.math.Ordering
   
   val rdd: RDD[Foo] = sc
     .parallelize(Seq(
       ("a", "2015-01-03"), ("b", "2014-11-04"), ("a", "2016-08-10"),
       ("a", "2013-11-11"), ("a", "2015-06-19"), ("a", "2009-11-23")))
     .toDF("name", "createDate")
     .withColumn("createDate", $"createDate".cast("date"))
     .as[Foo].rdd
   
   rdd.cache()
   val  n = scala.math.ceil(0.1 * rdd.count).toInt
   

   • data fits into driver memory:

     • and fraction you want is relatively small

       rdd.takeOrdered(n)(Ordering.by[Foo, Long](_.createDate.getTime))
       // Array[Foo] = Array(Foo(a,2009-11-23))
       

     • fraction you want is relatively large:

       rdd.sortBy(_.createDate.getTime).take(n)
       

   • otherwise

     rdd
       .sortBy(_.createDate.getTime)
       .zipWithIndex
       .filter{case (_, idx) => idx < n}
       .keys
     

2. Using DataFrame (note - this is actually not optimal performance wise due to limit behavior).

   import org.apache.spark.sql.Row
   
   val topN = rdd.toDF.orderBy($"createDate").limit(n)
   topN.show
   
   // +----+----------+
   // |name|createDate|
   // +----+----------+
   // |   a|2009-11-23|
   // +----+----------+
   
   
   // Optionally recreate RDD[Foo]
   topN.map{case Row(name: String, date: Date) => Foo(name, date)}