Merge pull request #473 from stephenh/subtract

Add RDD.subtract.

core/src/main/scala/spark/RDD.scala

@@ -30,6 +30,7 @@ import spark.rdd.MapPartitionsRDD
 import spark.rdd.MapPartitionsWithIndexRDD
 import spark.rdd.PipedRDD
 import spark.rdd.SampledRDD
+import spark.rdd.SubtractedRDD
 import spark.rdd.UnionRDD
 import spark.rdd.ZippedRDD
 import spark.storage.StorageLevel
@@ -393,6 +394,26 @@ abstract class RDD[T: ClassManifest](
     filter(f.isDefinedAt).map(f)
   }
 
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   * 
+   * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting
+   * RDD will be <= us.
+   */
+  def subtract(other: RDD[T]): RDD[T] =
+    subtract(other, partitioner.getOrElse(new HashPartitioner(partitions.size)))
+
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   */
+  def subtract(other: RDD[T], numPartitions: Int): RDD[T] =
+    subtract(other, new HashPartitioner(numPartitions))
+
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   */
+  def subtract(other: RDD[T], p: Partitioner): RDD[T] = new SubtractedRDD[T](this, other, p)
+
   /**
    * Reduces the elements of this RDD using the specified commutative and associative binary operator.
    */

core/src/main/scala/spark/api/java/JavaDoubleRDD.scala

@@ -6,8 +6,8 @@ import spark.api.java.function.{Function => JFunction}
 import spark.util.StatCounter
 import spark.partial.{BoundedDouble, PartialResult}
 import spark.storage.StorageLevel
-
 import java.lang.Double
+import spark.Partitioner
 
 class JavaDoubleRDD(val srdd: RDD[scala.Double]) extends JavaRDDLike[Double, JavaDoubleRDD] {
 
@@ -57,6 +57,27 @@ class JavaDoubleRDD(val srdd: RDD[scala.Double]) extends JavaRDDLike[Double, Jav
    */
   def coalesce(numPartitions: Int): JavaDoubleRDD = fromRDD(srdd.coalesce(numPartitions))
 
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   * 
+   * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting
+   * RDD will be <= us.
+   */
+  def subtract(other: JavaDoubleRDD): JavaDoubleRDD =
+    fromRDD(srdd.subtract(other))
+
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   */
+  def subtract(other: JavaDoubleRDD, numPartitions: Int): JavaDoubleRDD =
+    fromRDD(srdd.subtract(other, numPartitions))
+
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   */
+  def subtract(other: JavaDoubleRDD, p: Partitioner): JavaDoubleRDD =
+    fromRDD(srdd.subtract(other, p))
+
   /**
    * Return a sampled subset of this RDD.
    */

core/src/main/scala/spark/api/java/JavaPairRDD.scala

@@ -181,6 +181,27 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif
   def groupByKey(numPartitions: Int): JavaPairRDD[K, JList[V]] =
     fromRDD(groupByResultToJava(rdd.groupByKey(numPartitions)))
 
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   * 
+   * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting
+   * RDD will be <= us.
+   */
+  def subtract(other: JavaPairRDD[K, V]): JavaPairRDD[K, V] =
+    fromRDD(rdd.subtract(other))
+
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   */
+  def subtract(other: JavaPairRDD[K, V], numPartitions: Int): JavaPairRDD[K, V] =
+    fromRDD(rdd.subtract(other, numPartitions))
+
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   */
+  def subtract(other: JavaPairRDD[K, V], p: Partitioner): JavaPairRDD[K, V] =
+    fromRDD(rdd.subtract(other, p))
+
   /**
    * Return a copy of the RDD partitioned using the specified partitioner. If `mapSideCombine`
    * is true, Spark will group values of the same key together on the map side before the

core/src/main/scala/spark/api/java/JavaRDD.scala

@@ -55,6 +55,26 @@ JavaRDDLike[T, JavaRDD[T]] {
    */
   def union(other: JavaRDD[T]): JavaRDD[T] = wrapRDD(rdd.union(other.rdd))
 
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   * 
+   * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting
+   * RDD will be <= us.
+   */
+  def subtract(other: JavaRDD[T]): JavaRDD[T] = wrapRDD(rdd.subtract(other))
+
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   */
+  def subtract(other: JavaRDD[T], numPartitions: Int): JavaRDD[T] =
+    wrapRDD(rdd.subtract(other, numPartitions))
+
+  /**
+   * Return an RDD with the elements from `this` that are not in `other`.
+   */
+  def subtract(other: JavaRDD[T], p: Partitioner): JavaRDD[T] =
+    wrapRDD(rdd.subtract(other, p))
+
 }
 
 object JavaRDD {

core/src/main/scala/spark/rdd/SubtractedRDD.scala

+package spark.rdd
+
+import java.util.{HashSet => JHashSet}
+import scala.collection.JavaConversions._
+import spark.RDD
+import spark.Partitioner
+import spark.Dependency
+import spark.TaskContext
+import spark.Partition
+import spark.SparkEnv
+import spark.ShuffleDependency
+import spark.OneToOneDependency
+
+/**
+ * An optimized version of cogroup for set difference/subtraction.
+ *
+ * It is possible to implement this operation with just `cogroup`, but
+ * that is less efficient because all of the entries from `rdd2`, for
+ * both matching and non-matching values in `rdd1`, are kept in the
+ * JHashMap until the end.
+ *
+ * With this implementation, only the entries from `rdd1` are kept in-memory,
+ * and the entries from `rdd2` are essentially streamed, as we only need to
+ * touch each once to decide if the value needs to be removed.
+ *
+ * This is particularly helpful when `rdd1` is much smaller than `rdd2`, as
+ * you can use `rdd1`'s partitioner/partition size and not worry about running
+ * out of memory because of the size of `rdd2`.
+ */
+private[spark] class SubtractedRDD[T: ClassManifest](
+    @transient var rdd1: RDD[T],
+    @transient var rdd2: RDD[T],
+    part: Partitioner) extends RDD[T](rdd1.context, Nil) {
+
+  override def getDependencies: Seq[Dependency[_]] = {
+    Seq(rdd1, rdd2).map { rdd =>
+      if (rdd.partitioner == Some(part)) {
+        logInfo("Adding one-to-one dependency with " + rdd)
+        new OneToOneDependency(rdd)
+      } else {
+        logInfo("Adding shuffle dependency with " + rdd)
+        val mapSideCombinedRDD = rdd.mapPartitions(i => {
+          val set = new JHashSet[T]()
+          while (i.hasNext) {
+            set.add(i.next)
+          }
+          set.iterator
+        }, true)
+        // ShuffleDependency requires a tuple (k, v), which it will partition by k.
+        // We need this to partition to map to the same place as the k for
+        // OneToOneDependency, which means:
+        // - for already-tupled RDD[(A, B)], into getPartition(a)
+        // - for non-tupled RDD[C], into getPartition(c)
+        val part2 = new Partitioner() {
+          def numPartitions = part.numPartitions
+          def getPartition(key: Any) = key match {
+            case (k, v) => part.getPartition(k)
+            case k => part.getPartition(k)
+          }
+        }
+        new ShuffleDependency(mapSideCombinedRDD.map((_, null)), part2)
+      }
+    }
+  }
+
+  override def getPartitions: Array[Partition] = {
+    val array = new Array[Partition](part.numPartitions)
+    for (i <- 0 until array.size) {
+      // Each CoGroupPartition will depend on rdd1 and rdd2
+      array(i) = new CoGroupPartition(i, Seq(rdd1, rdd2).zipWithIndex.map { case (rdd, j) =>
+        dependencies(j) match {
+          case s: ShuffleDependency[_, _] =>
+            new ShuffleCoGroupSplitDep(s.shuffleId)
+          case _ =>
+            new NarrowCoGroupSplitDep(rdd, i, rdd.partitions(i))
+        }
+      }.toList)
+    }
+    array
+  }
+
+  override val partitioner = Some(part)
+
+  override def compute(p: Partition, context: TaskContext): Iterator[T] = {
+    val partition = p.asInstanceOf[CoGroupPartition]
+    val set = new JHashSet[T]
+    def integrate(dep: CoGroupSplitDep, op: T => Unit) = dep match {
+      case NarrowCoGroupSplitDep(rdd, _, itsSplit) =>
+        for (k <- rdd.iterator(itsSplit, context))
+          op(k.asInstanceOf[T])
+      case ShuffleCoGroupSplitDep(shuffleId) =>
+        for ((k, _) <- SparkEnv.get.shuffleFetcher.fetch(shuffleId, partition.index))
+          op(k.asInstanceOf[T])
+    }
+    // the first dep is rdd1; add all keys to the set
+    integrate(partition.deps(0), set.add)
+    // the second dep is rdd2; remove all of its keys from the set
+    integrate(partition.deps(1), set.remove)
+    set.iterator
+  }
+
+  override def clearDependencies() {
+    super.clearDependencies()
+    rdd1 = null
+    rdd2 = null
+  }
+
+}
\ No newline at end of file

core/src/test/scala/spark/ShuffleSuite.scala

@@ -234,6 +234,32 @@ class ShuffleSuite extends FunSuite with ShouldMatchers with LocalSparkContext {
     assert(rdd.keys.collect().toList === List(1, 2))
     assert(rdd.values.collect().toList === List("a", "b"))
   }
+
+  test("subtract") {
+    sc = new SparkContext("local", "test")
+    val a = sc.parallelize(Array(1, 2, 3), 2)
+    val b = sc.parallelize(Array(2, 3, 4), 4)
+    val c = a.subtract(b)
+    assert(c.collect().toSet === Set(1))
+    assert(c.partitions.size === a.partitions.size)
+  }
+
+  test("subtract with narrow dependency") {
+    sc = new SparkContext("local", "test")
+    // use a deterministic partitioner
+    val p = new Partitioner() {
+      def numPartitions = 5
+      def getPartition(key: Any) = key.asInstanceOf[Int]
+    }
+    // partitionBy so we have a narrow dependency
+    val a = sc.parallelize(Array((1, "a"), (2, "b"), (3, "c"))).partitionBy(p)
+    println(sc.runJob(a, (i: Iterator[(Int, String)]) => i.toList).toList)
+    // more splits/no partitioner so a shuffle dependency 
+    val b = sc.parallelize(Array((2, "b"), (3, "cc"), (4, "d")), 4)
+    val c = a.subtract(b)
+    assert(c.collect().toSet === Set((1, "a"), (3, "c")))
+    assert(c.partitioner.get === p)
+  }
 }
 
 object ShuffleSuite {