Apache Spark RDD filter into two RDDs

I need to split an RDD into 2 parts:

1 part which satisfies a condition; another part which does not. I can do filter twice on the original RDD but it seems inefficient. Is there a way that can do what I'm after? I can't find anything in the API nor in the literature.

Marius Soutier

Spark doesn't support this by default. Filtering on the same data twice isn't that bad if you cache it beforehand, and the filtering itself is quick.

If it's really just two different types, you can use a helper method:

implicit class RDDOps[T](rdd: RDD[T]) {
  def partitionBy(f: T => Boolean): (RDD[T], RDD[T]) = {
    val passes = rdd.filter(f)
    val fails = rdd.filter(e => !f(e)) // Spark doesn't have filterNot
    (passes, fails)
  }
}

val (matches, matchesNot) = sc.parallelize(1 to 100).cache().partitionBy(_ % 2 == 0)

But as soon as you have multiple types of data, just assign the filtered to a new val.

Spark RDD does not have such api.

Here is a version based on a pull request for rdd.span that should work:

import scala.reflect.ClassTag
import org.apache.spark.rdd._

def split[T:ClassTag](rdd: RDD[T], p: T => Boolean): (RDD[T], RDD[T]) = {

    val splits = rdd.mapPartitions { iter =>
        val (left, right) = iter.partition(p)
        val iterSeq = Seq(left, right)
        iterSeq.iterator
    }

    val left = splits.mapPartitions { iter => iter.next().toIterator}

    val right = splits.mapPartitions { iter => 
        iter.next()
        iter.next().toIterator
    }
    (left, right)
}

val rdd = sc.parallelize(0 to 10, 2)

val (first, second) = split[Int](rdd, _ % 2 == 0 )

first.collect
// Array[Int] = Array(0, 2, 4, 6, 8, 10)

The point is, you do not want to do a filter, but a map.

(T) -> (Boolean, T)

Sorry, I am inefficient in Scala Syntax. But the idea is that you split your answer set by mapping it to Key/Value pairs. The Key can be a boolean indicating wether or not it was passing the 'Filter' predicate.

You can control output to different targets by doing partition wise processing. Just make sure that you don’t restrict parallel processing to just two partitions downstream.

Justin Pihony

If you are ok with a T instead of an RDD[T], then you can do this. Otherwise, you could maybe do something like this:

val data = sc.parallelize(1 to 100)
val splitData = data.mapPartitions{iter => {
    val splitList = (iter.toList).partition(_%2 == 0)
    Tuple1(splitList).productIterator
  }
}.map(_.asInstanceOf[Tuple2[List[Int],List[Int]]])

And, then you will probably need to reduce this down to merge the lists when you go to perform an action

Priyanshu Ranjan

You can use subtract function (If filter operation is too expensive).

PySpark code:

rdd1 = data.filter(filterFunction)

rdd2 = data.subtract(rdd1)

来源：https://stackoverflow.com/questions/29547185/apache-spark-rdd-filter-into-two-rdds

标签

apache-spark

rdd