如何在 Spark RDD （Java） 中按索引获取元素

这应该可以通过首先为RDD编制索引来实现。转换提供稳定的索引，按其原始顺序对每个元素进行编号。zipWithIndex

鉴于：rdd = (a,b,c)

val withIndex = rdd.zipWithIndex // ((a,0),(b,1),(c,2))

若要按索引查找元素，此窗体没有用处。首先，我们需要使用索引作为键：

val indexKey = withIndex.map{case (k,v) => (v,k)}  //((0,a),(1,b),(2,c))

现在，可以使用 PairRDD 中的操作按键查找元素：lookup

val b = indexKey.lookup(1) // Array(b)

如果您希望经常在同一RDD上使用，我建议缓存RDD以提高性能。lookupindexKey

如何使用 Java API 执行此操作是留给读者的练习。

我尝试这个类按索引获取项目。首先，当您构造时，它会计算RDD的每个分区中的元素数。然后，当您调用 时，它仅在包含该索引的分区上运行作业。new IndexedFetcher(rdd, itemClass)indexedFetcher.get(n)

请注意，我需要使用Java 1.7而不是1.8编译它;从 Spark 1.1.0 开始，com.esotericsoftware.reflectasm 中捆绑的 org.objectweb.asm 还不能读取 Java 1.8 类（当您尝试运行 Java 1.8 函数时，会抛出 IllegalStateException）。

import java.io.Serializable;

import org.apache.spark.SparkContext;
import org.apache.spark.TaskContext;
import org.apache.spark.rdd.RDD;

import scala.reflect.ClassTag;

public static class IndexedFetcher<E> implements Serializable {
    private static final long serialVersionUID = 1L;
    public final RDD<E> rdd;
    public Integer[] elementsPerPartitions;
    private Class<?> clazz;
    public IndexedFetcher(RDD<E> rdd, Class<?> clazz){
        this.rdd = rdd;
        this.clazz = clazz;
        SparkContext context = this.rdd.context();
        ClassTag<Integer> intClassTag = scala.reflect.ClassTag$.MODULE$.<Integer>apply(Integer.class);
        elementsPerPartitions = (Integer[]) context.<E, Integer>runJob(rdd, IndexedFetcher.<E>countFunction(), intClassTag);
    }
    public static class IteratorCountFunction<E> extends scala.runtime.AbstractFunction2<TaskContext, scala.collection.Iterator<E>, Integer> implements Serializable {
        private static final long serialVersionUID = 1L;
        @Override public Integer apply(TaskContext taskContext, scala.collection.Iterator<E> iterator) {
            int count = 0;
            while (iterator.hasNext()) {
                count++;
                iterator.next();
            }
            return count;
        }
    }
    static <E> scala.Function2<TaskContext, scala.collection.Iterator<E>, Integer> countFunction() {
        scala.Function2<TaskContext, scala.collection.Iterator<E>, Integer> function = new IteratorCountFunction<E>();
        return function;
    }
    public E get(long index) {
        long remaining = index;
        long totalCount = 0;
        for (int partition = 0; partition < elementsPerPartitions.length; partition++) {
            if (remaining < elementsPerPartitions[partition]) {
                return getWithinPartition(partition, remaining);
            }
            remaining -= elementsPerPartitions[partition];
            totalCount += elementsPerPartitions[partition];
        }
        throw new IllegalArgumentException(String.format("Get %d within RDD that has only %d elements", index, totalCount));
    }
    public static class FetchWithinPartitionFunction<E> extends scala.runtime.AbstractFunction2<TaskContext, scala.collection.Iterator<E>, E> implements Serializable {
        private static final long serialVersionUID = 1L;
        private final long indexWithinPartition;
        public FetchWithinPartitionFunction(long indexWithinPartition) {
            this.indexWithinPartition = indexWithinPartition;
        }
        @Override public E apply(TaskContext taskContext, scala.collection.Iterator<E> iterator) {
            int count = 0;
            while (iterator.hasNext()) {
                E element = iterator.next();
                if (count == indexWithinPartition)
                    return element;
                count++;
            }
            throw new IllegalArgumentException(String.format("Fetch %d within partition that has only %d elements", indexWithinPartition, count));
        }
    }
    public E getWithinPartition(int partition, long indexWithinPartition) {
        System.out.format("getWithinPartition(%d, %d)%n", partition, indexWithinPartition);
        SparkContext context = rdd.context();
        scala.Function2<TaskContext, scala.collection.Iterator<E>, E> function = new FetchWithinPartitionFunction<E>(indexWithinPartition);
        scala.collection.Seq<Object> partitions = new scala.collection.mutable.WrappedArray.ofInt(new int[] {partition});
        ClassTag<E> classTag = scala.reflect.ClassTag$.MODULE$.<E>apply(this.clazz);
        E[] result = (E[]) context.<E, E>runJob(rdd, function, partitions, true, classTag);
        return result[0];
    }
}