SPARK-2043: ExternalAppendOnlyMap doesn't always find matching keys

mateiz · mateiz · commit 51f677eb9a4f · 2014-06-05T23:02:11.000-07:00
The current implementation reads one key with the next hash code as it finishes reading the keys with the current hash code, which may cause it to miss some matches of the next key. This can cause operations like join to give the wrong result when reduce tasks spill to disk and there are hash collisions, as values won't be matched together. This PR fixes it by not reading in that next key, using a peeking iterator instead. Author: Matei Zaharia <matei@databricks.com> Closes apache#986 from mateiz/spark-2043 and squashes the following commits: 0959514 [Matei Zaharia] Added unit test for having many hash collisions 892debb [Matei Zaharia] SPARK-2043: don't read a key with the next hash code in ExternalAppendOnlyMap, instead use a buffered iterator to only read values with the current hash code. (cherry picked from commit b45c13e) Signed-off-by: Matei Zaharia <matei@databricks.com>
diff --git a/core/src/main/scala/org/apache/spark/util/collection/ExternalAppendOnlyMap.scala b/core/src/main/scala/org/apache/spark/util/collection/ExternalAppendOnlyMap.scala
@@ -20,6 +20,7 @@ package org.apache.spark.util.collection
 import java.io._
 import java.util.Comparator
 
+import scala.collection.BufferedIterator
 import scala.collection.mutable
 import scala.collection.mutable.ArrayBuffer
 
@@ -230,7 +231,7 @@ private[spark] class ExternalAppendOnlyMap[K, V, C](
     // Input streams are derived both from the in-memory map and spilled maps on disk
     // The in-memory map is sorted in place, while the spilled maps are already in sorted order
     private val sortedMap = currentMap.destructiveSortedIterator(comparator)
-    private val inputStreams = Seq(sortedMap) ++ spilledMaps
+    private val inputStreams = (Seq(sortedMap) ++ spilledMaps).map(it => it.buffered)
 
     inputStreams.foreach { it =>
       val kcPairs = getMorePairs(it)
@@ -245,13 +246,13 @@ private[spark] class ExternalAppendOnlyMap[K, V, C](
      * In the event of key hash collisions, this ensures no pairs are hidden from being merged.
      * Assume the given iterator is in sorted order.
      */
-    private def getMorePairs(it: Iterator[(K, C)]): ArrayBuffer[(K, C)] = {
+    private def getMorePairs(it: BufferedIterator[(K, C)]): ArrayBuffer[(K, C)] = {
       val kcPairs = new ArrayBuffer[(K, C)]
       if (it.hasNext) {
         var kc = it.next()
         kcPairs += kc
         val minHash = kc._1.hashCode()
-        while (it.hasNext && kc._1.hashCode() == minHash) {
+        while (it.hasNext && it.head._1.hashCode() == minHash) {
           kc = it.next()
           kcPairs += kc
         }
@@ -324,7 +325,8 @@ private[spark] class ExternalAppendOnlyMap[K, V, C](
      *
      * StreamBuffers are ordered by the minimum key hash found across all of their own pairs.
      */
-    private case class StreamBuffer(iterator: Iterator[(K, C)], pairs: ArrayBuffer[(K, C)])
+    private class StreamBuffer(
+        val iterator: BufferedIterator[(K, C)], val pairs: ArrayBuffer[(K, C)])
       extends Comparable[StreamBuffer] {
 
       def isEmpty = pairs.length == 0
diff --git a/core/src/test/scala/org/apache/spark/util/collection/ExternalAppendOnlyMapSuite.scala b/core/src/test/scala/org/apache/spark/util/collection/ExternalAppendOnlyMapSuite.scala
@@ -277,6 +277,11 @@ class ExternalAppendOnlyMapSuite extends FunSuite with LocalSparkContext {
       ("pomatoes", "eructation")      // 568647356
     )
 
+    collisionPairs.foreach { case (w1, w2) =>
+      // String.hashCode is documented to use a specific algorithm, but check just in case
+      assert(w1.hashCode === w2.hashCode)
+    }
+
     (1 to 100000).map(_.toString).foreach { i => map.insert(i, i) }
     collisionPairs.foreach { case (w1, w2) =>
       map.insert(w1, w2)
@@ -296,7 +301,32 @@ class ExternalAppendOnlyMapSuite extends FunSuite with LocalSparkContext {
       assert(kv._2.equals(expectedValue))
       count += 1
     }
-    assert(count == 100000 + collisionPairs.size * 2)
+    assert(count === 100000 + collisionPairs.size * 2)
+  }
+
+  test("spilling with many hash collisions") {
+    val conf = new SparkConf(true)
+    conf.set("spark.shuffle.memoryFraction", "0.0001")
+    sc = new SparkContext("local-cluster[1,1,512]", "test", conf)
+
+    val map = new ExternalAppendOnlyMap[FixedHashObject, Int, Int](_ => 1, _ + _, _ + _)
+
+    // Insert 10 copies each of lots of objects whose hash codes are either 0 or 1. This causes
+    // problems if the map fails to group together the objects with the same code (SPARK-2043).
+    for (i <- 1 to 10) {
+      for (j <- 1 to 10000) {
+        map.insert(FixedHashObject(j, j % 2), 1)
+      }
+    }
+
+    val it = map.iterator
+    var count = 0
+    while (it.hasNext) {
+      val kv = it.next()
+      assert(kv._2 === 10)
+      count += 1
+    }
+    assert(count === 10000)
   }
 
   test("spilling with hash collisions using the Int.MaxValue key") {
@@ -317,3 +347,10 @@ class ExternalAppendOnlyMapSuite extends FunSuite with LocalSparkContext {
     }
   }
 }
+
+/**
+ * A dummy class that always returns the same hash code, to easily test hash collisions
+ */
+case class FixedHashObject(val v: Int, val h: Int) extends Serializable {
+  override def hashCode(): Int = h
+}
