Github user dorx commented on a diff in the pull request:
https://github.com/apache/spark/pull/1520#discussion_r15265778
--- Diff: mllib/src/main/scala/org/apache/spark/mllib/rdd/RandomRDD.scala
---
@@ -0,0 +1,140 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.mllib.rdd
+
+import org.apache.spark.{Partition, SparkContext, TaskContext}
+import org.apache.spark.mllib.linalg.{DenseVector, Vector}
+import org.apache.spark.mllib.random.DistributionGenerator
+import org.apache.spark.rdd.RDD
+import org.apache.spark.util.Utils
+
+private[mllib] class RandomRDDPartition(val idx: Int,
+ val size: Long,
+ val rng: DistributionGenerator,
+ val seed: Long) extends Partition {
+
+ override val index: Int = idx
+
+}
+
+// These two classes are necessary since Range objects in Scala cannot
have size > Int.MaxValue
+private[mllib] class RandomRDD(@transient private var sc: SparkContext,
+ size: Long,
+ numSlices: Int,
+ @transient rng: DistributionGenerator,
+ @transient seed: Long = Utils.random.nextLong) extends RDD[Double](sc,
Nil) {
+
+ require(size > 0, "Positive RDD size required.")
+ require(numSlices > 0, "Positive number of partitions required")
+
+ override def compute(splitIn: Partition, context: TaskContext):
Iterator[Double] = {
+ val split = splitIn.asInstanceOf[RandomRDDPartition]
+ RandomRDD.getPointIterator(split)
+ }
+
+ override def getPartitions: Array[Partition] = {
+ RandomRDD.getPartitions(size, numSlices, rng, seed)
+ }
+}
+
+private[mllib] class RandomVectorRDD(@transient private var sc:
SparkContext,
+ size: Long,
+ vectorSize: Int,
+ numSlices: Int,
+ @transient rng: DistributionGenerator,
+ @transient seed: Long = Utils.random.nextLong) extends RDD[Vector](sc,
Nil) {
+
+ require(size > 0, "Positive RDD size required.")
+ require(numSlices > 0, "Positive number of partitions required")
+ require(vectorSize > 0, "Positive vector size required.")
+
+ override def compute(splitIn: Partition, context: TaskContext):
Iterator[Vector] = {
+ val split = splitIn.asInstanceOf[RandomRDDPartition]
+ RandomRDD.getVectorIterator(split, vectorSize)
+ }
+
+ override protected def getPartitions: Array[Partition] = {
+ RandomRDD.getPartitions(size, numSlices, rng, seed)
+ }
+}
+
+private[mllib] object RandomRDD {
+
+ private[mllib] class FixedSizePointIterator(val numElem: Long, val rng:
DistributionGenerator)
+ extends Iterator[Double] {
+
+ private var currentSize = 0
+
+ override def hasNext: Boolean = currentSize < numElem
+
+ override def next(): Double = {
+ currentSize += 1
+ rng.nextValue()
+ }
+ }
+
+ private[mllib] class FixedSizeVectorIterator(val numElem: Long,
+ val vectorSize: Int,
+ val rng: DistributionGenerator)
+ extends Iterator[Vector] {
+
+ private var currentSize = 0
+
+ override def hasNext: Boolean = currentSize < numElem
+
+ override def next(): Vector = {
+ currentSize += 1
+ new DenseVector((0 until vectorSize).map { _ => rng.nextValue()
}.toArray)
+ }
+ }
+
+ def getPartitions(size: Long,
+ numSlices: Int,
+ rng: DistributionGenerator,
+ seed: Long): Array[Partition] = {
+
+ val firstPartitionSize = size / numSlices + size % numSlices
+ val otherPartitionSize = size / numSlices
+
+ val partitions = new Array[RandomRDDPartition](numSlices)
+ var i = 0
+ while (i < numSlices) {
+ partitions(i) = if (i == 0) {
+ new RandomRDDPartition(i, firstPartitionSize, rng, seed)
+ } else {
+ new RandomRDDPartition(i, otherPartitionSize, rng.newInstance(),
seed)
+ }
+ i += 1
+ }
+ partitions.asInstanceOf[Array[Partition]]
+ }
+
+ // The RNG has to be reset every time the iterator is requested to
guarantee same data
+ // every time the content of the RDD is examined.
+ def getPointIterator(partition: RandomRDDPartition): Iterator[Double] = {
+ partition.rng.setSeed(partition.seed + partition.index)
--- End diff --
Thanks for the pointer to PartitionwiseSampledRDD. I've actually looked a
lot at it for the sampling PRs. There were two considerations for going with
this more deterministic seed assignment per partition. One is that this is a
lot easier to test to make sure that each partition has a different seed with
this deterministic seed assignment, and the other is that we're guaranteed to
have a unique seed per partition. Since most RNG implementations hash or
otherwise scramble the bits in the input seed before setting it anyway, I think
we can get away with the current implementation.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
