[SPARK-2568] RangePartitioner should run only one job if data is balanced

As of Spark 1.0, RangePartitioner goes through data twice: once to compute the count and once to do sampling. As a result, to do sortByKey, Spark goes through data 3 times (once to count, once to sample, and once to sort).

`RangePartitioner` should go through data only once, collecting samples from input partitions as well as counting. If the data is balanced, this should give us a good sketch. If we see big partitions, we re-sample from them in order to collect enough items.

The downside is that we need to collect more from each partition in the first pass. An alternative solution is caching the intermediate result and decide whether to fetch the data after.

Author: Xiangrui Meng <[email protected]>
Author: Reynold Xin <[email protected]>

Closes apache#1562 from mengxr/range-partitioner and squashes the following commits:

6cc2551 [Xiangrui Meng] change foreach to for
eb39b08 [Xiangrui Meng] Merge branch 'master' into range-partitioner
eb95dd8 [Xiangrui Meng] separate sketching and determining bounds impl
c436d30 [Xiangrui Meng] fix binary metrics unit tests
db58a55 [Xiangrui Meng] add unit tests
a6e35d6 [Xiangrui Meng] minor update
60be09e [Xiangrui Meng] remove importance sampler
9ee9992 [Xiangrui Meng] update range partitioner to run only one job on roughly balanced data
cc12f47 [Xiangrui Meng] Merge remote-tracking branch 'apache/master' into range-part
06ac2ec [Xiangrui Meng] Merge remote-tracking branch 'apache/master' into range-part
17bcbf3 [Reynold Xin] Added seed.
badf20d [Reynold Xin] Renamed the method.
6940010 [Reynold Xin] Reservoir sampling implementation.

Author: mengxr

core/src/main/scala/org/apache/spark/Partitioner.scala

@@ -19,11 +19,15 @@ package org.apache.spark
 
 import java.io.{IOException, ObjectInputStream, ObjectOutputStream}
 
-import scala.reflect.ClassTag
+import scala.collection.mutable
+import scala.collection.mutable.ArrayBuffer
+import scala.reflect.{ClassTag, classTag}
+import scala.util.hashing.byteswap32
 
-import org.apache.spark.rdd.RDD
+import org.apache.spark.rdd.{PartitionPruningRDD, RDD}
 import org.apache.spark.serializer.JavaSerializer
 import org.apache.spark.util.{CollectionsUtils, Utils}
+import org.apache.spark.util.random.{XORShiftRandom, SamplingUtils}
 
 /**
  * An object that defines how the elements in a key-value pair RDD are partitioned by key.
@@ -103,26 +107,49 @@ class RangePartitioner[K : Ordering : ClassTag, V](
     private var ascending: Boolean = true)
   extends Partitioner {
 
+  // We allow partitions = 0, which happens when sorting an empty RDD under the default settings.
+  require(partitions >= 0, s"Number of partitions cannot be negative but found $partitions.")
+
   private var ordering = implicitly[Ordering[K]]
 
   // An array of upper bounds for the first (partitions - 1) partitions
   private var rangeBounds: Array[K] = {
-    if (partitions == 1) {
-      Array()
+    if (partitions <= 1) {
+      Array.empty
     } else {
-      val rddSize = rdd.count()
-      val maxSampleSize = partitions * 20.0
-      val frac = math.min(maxSampleSize / math.max(rddSize, 1), 1.0)
-      val rddSample = rdd.sample(false, frac, 1).map(_._1).collect().sorted
-      if (rddSample.length == 0) {
-        Array()
+      // This is the sample size we need to have roughly balanced output partitions, capped at 1M.
+      val sampleSize = math.min(20.0 * partitions, 1e6)
+      // Assume the input partitions are roughly balanced and over-sample a little bit.
+      val sampleSizePerPartition = math.ceil(3.0 * sampleSize / rdd.partitions.size).toInt
+      val (numItems, sketched) = RangePartitioner.sketch(rdd.map(_._1), sampleSizePerPartition)
+      if (numItems == 0L) {
+        Array.empty
       } else {
-        val bounds = new Array[K](partitions - 1)
-        for (i <- 0 until partitions - 1) {
-          val index = (rddSample.length - 1) * (i + 1) / partitions
-          bounds(i) = rddSample(index)
+        // If a partition contains much more than the average number of items, we re-sample from it
+        // to ensure that enough items are collected from that partition.
+        val fraction = math.min(sampleSize / math.max(numItems, 1L), 1.0)
+        val candidates = ArrayBuffer.empty[(K, Float)]
+        val imbalancedPartitions = mutable.Set.empty[Int]
+        sketched.foreach { case (idx, n, sample) =>
+          if (fraction * n > sampleSizePerPartition) {
+            imbalancedPartitions += idx
+          } else {
+            // The weight is 1 over the sampling probability.
+            val weight = (n.toDouble / sample.size).toFloat
+            for (key <- sample) {
+              candidates += ((key, weight))
+            }
+          }
+        }
+        if (imbalancedPartitions.nonEmpty) {
+          // Re-sample imbalanced partitions with the desired sampling probability.
+          val imbalanced = new PartitionPruningRDD(rdd.map(_._1), imbalancedPartitions.contains)
+          val seed = byteswap32(-rdd.id - 1)
+          val reSampled = imbalanced.sample(withReplacement = false, fraction, seed).collect()
+          val weight = (1.0 / fraction).toFloat
+          candidates ++= reSampled.map(x => (x, weight))
         }
-        bounds
+        RangePartitioner.determineBounds(candidates, partitions)
       }
     }
   }
@@ -212,3 +239,67 @@ class RangePartitioner[K : Ordering : ClassTag, V](
     }
   }
 }
+
+private[spark] object RangePartitioner {
+
+  /**
+   * Sketches the input RDD via reservoir sampling on each partition.
+   *
+   * @param rdd the input RDD to sketch
+   * @param sampleSizePerPartition max sample size per partition
+   * @return (total number of items, an array of (partitionId, number of items, sample))
+   */
+  def sketch[K:ClassTag](
+      rdd: RDD[K],
+      sampleSizePerPartition: Int): (Long, Array[(Int, Int, Array[K])]) = {
+    val shift = rdd.id
+    // val classTagK = classTag[K] // to avoid serializing the entire partitioner object
+    val sketched = rdd.mapPartitionsWithIndex { (idx, iter) =>
+      val seed = byteswap32(idx ^ (shift << 16))
+      val (sample, n) = SamplingUtils.reservoirSampleAndCount(
+        iter, sampleSizePerPartition, seed)
+      Iterator((idx, n, sample))
+    }.collect()
+    val numItems = sketched.map(_._2.toLong).sum
+    (numItems, sketched)
+  }
+
+  /**
+   * Determines the bounds for range partitioning from candidates with weights indicating how many
+   * items each represents. Usually this is 1 over the probability used to sample this candidate.
+   *
+   * @param candidates unordered candidates with weights
+   * @param partitions number of partitions
+   * @return selected bounds
+   */
+  def determineBounds[K:Ordering:ClassTag](
+      candidates: ArrayBuffer[(K, Float)],
+      partitions: Int): Array[K] = {
+    val ordering = implicitly[Ordering[K]]
+    val ordered = candidates.sortBy(_._1)
+    val numCandidates = ordered.size
+    val sumWeights = ordered.map(_._2.toDouble).sum
+    val step = sumWeights / partitions
+    var cumWeight = 0.0
+    var target = step
+    val bounds = ArrayBuffer.empty[K]
+    var i = 0
+    var j = 0
+    var previousBound = Option.empty[K]
+    while ((i < numCandidates) && (j < partitions - 1)) {
+      val (key, weight) = ordered(i)
+      cumWeight += weight
+      if (cumWeight > target) {
+        // Skip duplicate values.
+        if (previousBound.isEmpty || ordering.gt(key, previousBound.get)) {
+          bounds += key
+          target += step
+          j += 1
+          previousBound = Some(key)
+        }
+      }
+      i += 1
+    }
+    bounds.toArray
+  }
+}

core/src/test/scala/org/apache/spark/PartitioningSuite.scala

@@ -17,6 +17,7 @@
 
 package org.apache.spark
 
+import scala.collection.mutable.ArrayBuffer
 import scala.math.abs
 
 import org.scalatest.{FunSuite, PrivateMethodTester}
@@ -52,14 +53,12 @@ class PartitioningSuite extends FunSuite with SharedSparkContext with PrivateMet
 
     assert(p2 === p2)
     assert(p4 === p4)
-    assert(p2 != p4)
-    assert(p4 != p2)
+    assert(p2 === p4)
     assert(p4 === anotherP4)
     assert(anotherP4 === p4)
     assert(descendingP2 === descendingP2)
     assert(descendingP4 === descendingP4)
-    assert(descendingP2 != descendingP4)
-    assert(descendingP4 != descendingP2)
+    assert(descendingP2 === descendingP4)
     assert(p2 != descendingP2)
     assert(p4 != descendingP4)
     assert(descendingP2 != p2)
@@ -102,6 +101,63 @@ class PartitioningSuite extends FunSuite with SharedSparkContext with PrivateMet
     partitioner.getPartition(Row(100))
   }
 
+  test("RangPartitioner.sketch") {
+    val rdd = sc.makeRDD(0 until 20, 20).flatMap { i =>
+      val random = new java.util.Random(i)
+      Iterator.fill(i)(random.nextDouble())
+    }.cache()
+    val sampleSizePerPartition = 10
+    val (count, sketched) = RangePartitioner.sketch(rdd, sampleSizePerPartition)
+    assert(count === rdd.count())
+    sketched.foreach { case (idx, n, sample) =>
+      assert(n === idx)
+      assert(sample.size === math.min(n, sampleSizePerPartition))
+    }
+  }
+
+  test("RangePartitioner.determineBounds") {
+    assert(RangePartitioner.determineBounds(ArrayBuffer.empty[(Int, Float)], 10).isEmpty,
+      "Bounds on an empty candidates set should be empty.")
+    val candidates = ArrayBuffer(
+      (0.7, 2.0f), (0.1, 1.0f), (0.4, 1.0f), (0.3, 1.0f), (0.2, 1.0f), (0.5, 1.0f), (1.0, 3.0f))
+    assert(RangePartitioner.determineBounds(candidates, 3) === Array(0.4, 0.7))
+  }
+
+  test("RangePartitioner should run only one job if data is roughly balanced") {
+    val rdd = sc.makeRDD(0 until 20, 20).flatMap { i =>
+      val random = new java.util.Random(i)
+      Iterator.fill(5000 * i)((random.nextDouble() + i, i))
+    }.cache()
+    for (numPartitions <- Seq(10, 20, 40)) {
+      val partitioner = new RangePartitioner(numPartitions, rdd)
+      assert(partitioner.numPartitions === numPartitions)
+      val counts = rdd.keys.map(key => partitioner.getPartition(key)).countByValue().values
+      assert(counts.max < 3.0 * counts.min)
+    }
+  }
+
+  test("RangePartitioner should work well on unbalanced data") {
+    val rdd = sc.makeRDD(0 until 20, 20).flatMap { i =>
+      val random = new java.util.Random(i)
+      Iterator.fill(20 * i * i * i)((random.nextDouble() + i, i))
+    }.cache()
+    for (numPartitions <- Seq(2, 4, 8)) {
+      val partitioner = new RangePartitioner(numPartitions, rdd)
+      assert(partitioner.numPartitions === numPartitions)
+      val counts = rdd.keys.map(key => partitioner.getPartition(key)).countByValue().values
+      assert(counts.max < 3.0 * counts.min)
+    }
+  }
+
+  test("RangePartitioner should return a single partition for empty RDDs") {
+    val empty1 = sc.emptyRDD[(Int, Double)]
+    val partitioner1 = new RangePartitioner(0, empty1)
+    assert(partitioner1.numPartitions === 1)
+    val empty2 = sc.makeRDD(0 until 2, 2).flatMap(i => Seq.empty[(Int, Double)])
+    val partitioner2 = new RangePartitioner(2, empty2)
+    assert(partitioner2.numPartitions === 1)
+  }
+
   test("HashPartitioner not equal to RangePartitioner") {
     val rdd = sc.parallelize(1 to 10).map(x => (x, x))
     val rangeP2 = new RangePartitioner(2, rdd)

core/src/test/scala/org/apache/spark/rdd/RDDSuite.scala

@@ -613,6 +613,11 @@ class RDDSuite extends FunSuite with SharedSparkContext {
     }
   }
 
+  test("sort an empty RDD") {
+    val data = sc.emptyRDD[Int]
+    assert(data.sortBy(x => x).collect() === Array.empty)
+  }
+
   test("sortByKey") {
     val data = sc.parallelize(Seq("5|50|A","4|60|C", "6|40|B"))