SPARK-1939: Refactor takeSample method in RDD

Reviewer comments addressed: - commons-math3 is now a test-only dependency. bumped up to v3.3 - comments added to explain what computeFraction is doing - fixed the unit for computeFraction to use BinomialDitro for without replacement sampling - stylistic fixes
markhamstra · May 30, 2014 · ffea61a · ffea61a
1 parent 1441977
commit ffea61a
Show file tree

Hide file tree

Showing 6 changed files with 44 additions and 32 deletions.
diff --git a/core/pom.xml b/core/pom.xml
@@ -70,6 +70,7 @@
     <dependency>
       <groupId>org.apache.commons</groupId>
       <artifactId>commons-math3</artifactId>
+      <scope>test</scope>
     </dependency>
     <dependency>
       <groupId>com.google.code.findbugs</groupId>

diff --git a/core/src/main/scala/org/apache/spark/rdd/RDD.scala b/core/src/main/scala/org/apache/spark/rdd/RDD.scala
@@ -388,8 +388,8 @@ abstract class RDD[T: ClassTag](
    * @return sample of specified size in an array
    */
   def takeSample(withReplacement: Boolean,
-                 num: Int,
-                 seed: Long = Utils.random.nextLong): Array[T] = {
+      num: Int,
+      seed: Long = Utils.random.nextLong): Array[T] = {
     var fraction = 0.0
     var total = 0
     val multiplier = 3.0
@@ -431,18 +431,31 @@ abstract class RDD[T: ClassTag](
     Utils.randomizeInPlace(samples, rand).take(total)
   }
 
-  private[spark] def computeFraction(num: Int, total: Long, withReplacement: Boolean) : Double = {
+  /**
+   * Let p = num / total, where num is the sample size and total is the total number of
+   * datapoints in the RDD. We're trying to compute q > p such that
+   *   - when sampling with replacement, we're drawing each datapoint with prob_i ~ Pois(q),
+   *     where we want to guarantee Pr[s < num] < 0.0001 for s = sum(prob_i for i from 0 to total),
+   *     i.e. the failure rate of not having a sufficiently large sample < 0.0001.
+   *     Setting q = p + 5 * sqrt(p/total) is sufficient to guarantee 0.9999 success rate for
+   *     num > 12, but we need a slightly larger q (9 empirically determined).
+   *   - when sampling without replacement, we're drawing each datapoint with prob_i
+   *     ~ Binomial(total, fraction) and our choice of q guarantees 1-delta, or 0.9999 success
+   *     rate, where success rate is defined the same as in sampling with replacement.
+   *
+   * @param num sample size
+   * @param total size of RDD
+   * @param withReplacement whether sampling with replacement
+   * @return a sampling rate that guarantees sufficient sample size with 99.99% success rate
+   */
+  private[rdd] def computeFraction(num: Int, total: Long, withReplacement: Boolean): Double = {
     val fraction = num.toDouble / total
     if (withReplacement) {
-      var numStDev = 5
-      if (num < 12) {
-        // special case to guarantee sample size for small s
-        numStDev = 9
-      }
+      val numStDev = if (num < 12) 9 else 5
       fraction + numStDev * math.sqrt(fraction / total)
     } else {
-      val delta = 0.00005
-      val gamma = - math.log(delta)/total
+      val delta = 1e-4
+      val gamma = - math.log(delta) / total
       math.min(1, fraction + gamma + math.sqrt(gamma * gamma + 2 * gamma * fraction))
     }
   }

diff --git a/core/src/main/scala/org/apache/spark/util/random/RandomSampler.scala b/core/src/main/scala/org/apache/spark/util/random/RandomSampler.scala
@@ -70,7 +70,7 @@ class BernoulliSampler[T](lb: Double, ub: Double, complement: Boolean = false)
   }
 
   /**
-   *  Return a sampler which is the complement of the range specified of the current sampler.
+   *  Return a sampler that is the complement of the range specified of the current sampler.
    */
   def cloneComplement():  BernoulliSampler[T] = new BernoulliSampler[T](lb, ub, !complement)
 

diff --git a/core/src/test/scala/org/apache/spark/rdd/RDDSuite.scala b/core/src/test/scala/org/apache/spark/rdd/RDDSuite.scala
@@ -22,7 +22,9 @@ import scala.reflect.ClassTag
 
 import org.scalatest.FunSuite
 
+import org.apache.commons.math3.distribution.BinomialDistribution
 import org.apache.commons.math3.distribution.PoissonDistribution
+
 import org.apache.spark._
 import org.apache.spark.SparkContext._
 import org.apache.spark.rdd._
@@ -496,29 +498,25 @@ class RDDSuite extends FunSuite with SharedSparkContext {
   }
 
   test("computeFraction") {
-    // test that the computed fraction guarantees enough datapoints in the sample with a failure rate <= 0.0001
+    // test that the computed fraction guarantees enough datapoints
+    // in the sample with a failure rate <= 0.0001
     val data = new EmptyRDD[Int](sc)
     val n = 100000
 
     for (s <- 1 to 15) {
       val frac = data.computeFraction(s, n, true)
-      val qpois = new PoissonDistribution(frac * n)
-      assert(qpois.inverseCumulativeProbability(0.0001) >= s, "Computed fraction is too low")
+      val poisson = new PoissonDistribution(frac * n)
+      assert(poisson.inverseCumulativeProbability(0.0001) >= s, "Computed fraction is too low")
     }
-    for (s <- 1 to 15) {
-      val frac = data.computeFraction(s, n, false)
-      val qpois = new PoissonDistribution(frac * n)
-      assert(qpois.inverseCumulativeProbability(0.0001) >= s, "Computed fraction is too low")
-    }
-    for (s <- List(1, 10, 100, 1000)) {
+    for (s <- List(20, 100, 1000)) {
       val frac = data.computeFraction(s, n, true)
-      val qpois = new PoissonDistribution(frac * n)
-      assert(qpois.inverseCumulativeProbability(0.0001) >= s, "Computed fraction is too low")
+      val poisson = new PoissonDistribution(frac * n)
+      assert(poisson.inverseCumulativeProbability(0.0001) >= s, "Computed fraction is too low")
     }
     for (s <- List(1, 10, 100, 1000)) {
       val frac = data.computeFraction(s, n, false)
-      val qpois = new PoissonDistribution(frac * n)
-      assert(qpois.inverseCumulativeProbability(0.0001) >= s, "Computed fraction is too low")
+      val binomial = new BinomialDistribution(n, frac)
+      assert(binomial.inverseCumulativeProbability(0.0001)*n >= s, "Computed fraction is too low")
     }
   }
 
@@ -530,37 +528,37 @@ class RDDSuite extends FunSuite with SharedSparkContext {
       val sample = data.takeSample(withReplacement=false, num=num)
       assert(sample.size === num)        // Got exactly num elements
       assert(sample.toSet.size === num)  // Elements are distinct
-      assert(sample.forall(x => 1 <= x && x <= n), "elements not in [1, 100]")
+      assert(sample.forall(x => 1 <= x && x <= n), s"elements not in [1, $n]")
     }
     for (seed <- 1 to 5) {
       val sample = data.takeSample(withReplacement=false, 20, seed)
       assert(sample.size === 20)        // Got exactly 20 elements
       assert(sample.toSet.size === 20)  // Elements are distinct
-      assert(sample.forall(x => 1 <= x && x <= n), "elements not in [1, 100]")
+      assert(sample.forall(x => 1 <= x && x <= n), s"elements not in [1, $n]")
     }
     for (seed <- 1 to 5) {
       val sample = data.takeSample(withReplacement=false, 100, seed)
       assert(sample.size === 100)        // Got only 100 elements
       assert(sample.toSet.size === 100)  // Elements are distinct
-      assert(sample.forall(x => 1 <= x && x <= n), "elements not in [1, 100]")
+      assert(sample.forall(x => 1 <= x && x <= n), s"elements not in [1, $n]")
     }
     for (seed <- 1 to 5) {
       val sample = data.takeSample(withReplacement=true, 20, seed)
       assert(sample.size === 20)        // Got exactly 20 elements
-      assert(sample.forall(x => 1 <= x && x <= n), "elements not in [1, 100]")
+      assert(sample.forall(x => 1 <= x && x <= n), s"elements not in [1, $n]")
     }
     {
       val sample = data.takeSample(withReplacement=true, num=20)
       assert(sample.size === 20)        // Got exactly 100 elements
       assert(sample.toSet.size <= 20, "sampling with replacement returned all distinct elements")
-      assert(sample.forall(x => 1 <= x && x <= n), "elements not in [1, 100]")
+      assert(sample.forall(x => 1 <= x && x <= n), s"elements not in [1, $n]")
     }
     {
       val sample = data.takeSample(withReplacement=true, num=n)
       assert(sample.size === n)        // Got exactly 100 elements
       // Chance of getting all distinct elements is astronomically low, so test we got < 100
       assert(sample.toSet.size < n, "sampling with replacement returned all distinct elements")
-      assert(sample.forall(x => 1 <= x && x <= n), "elements not in [1, 100]")
+      assert(sample.forall(x => 1 <= x && x <= n), s"elements not in [1, $n]")
     }
     for (seed <- 1 to 5) {
       val sample = data.takeSample(withReplacement=true, n, seed)

diff --git a/pom.xml b/pom.xml
@@ -248,7 +248,7 @@
       <dependency>
         <groupId>org.apache.commons</groupId>
         <artifactId>commons-math3</artifactId>
-        <version>3.2</version>
+        <version>3.3</version>
       </dependency>
       <dependency>
         <groupId>com.google.code.findbugs</groupId>

diff --git a/project/SparkBuild.scala b/project/SparkBuild.scala
@@ -331,7 +331,7 @@ object SparkBuild extends Build {
     libraryDependencies ++= Seq(
         "com.google.guava"           % "guava"            % "14.0.1",
         "org.apache.commons"         % "commons-lang3"    % "3.3.2",
-        "org.apache.commons"         % "commons-math3"    % "3.2",
+        "org.apache.commons"         % "commons-math3"    % "3.3" % "test",
         "com.google.code.findbugs"   % "jsr305"           % "1.3.9",
         "log4j"                      % "log4j"            % "1.2.17",
         "org.slf4j"                  % "slf4j-api"        % slf4jVersion,
-Original file line number
+Diff line change
@@ Expand Up @@
       }
       /**
-       *  Return a sampler which is the complement of the range specified of the current sampler.
+       *  Return a sampler that is the complement of the range specified of the current sampler.
        */
       def cloneComplement():  BernoulliSampler[T] = new BernoulliSampler[T](lb, ub, !complement)
@@ Expand Down @@