java - 转换 JavaRDD> 进入 Spark 数据集在爪哇

Question

在 Java（不是 Scala！）中，Spark 3.0.1 有一个JavaRDD实例对象neighborIdsRDD，其类型为JavaRDD<Tuple2<Object, long[]>>.

我与生成 JavaRDD 相关的部分代码如下：

GraphOps<String, String> graphOps = new GraphOps<>(graph, stringTag, stringTag);
JavaRDD<Tuple2<Object, long[]>> neighborIdsRDD = graphOps.collectNeighborIds(EdgeDirection.Either()).toJavaRDD();

我不得不使用 JavaRDD，toJavaRDD()因为collectNeighborIds它返回一个org.apache.spark.graphx.VertexRDD<long[]>对象（VertexRDD doc）。

但是，在我的应用程序的其余部分中，我需要从对象Dataset<Row>构建一个 Spark。collectNeighborIds

将JavaRDD<Tuple2<Object, long[]>>转换为Dataset<Row>的可能性和最佳方法是什么？

---

评论更新：

我根据评论调整了代码：

        GraphOps<String, String> graphOps = new GraphOps<>(graph, stringTag, stringTag);
        JavaRDD<Tuple2<Object, long[]>> neighborIdsRDD = graphOps.collectNeighborIds(EdgeDirection.Either()).toJavaRDD();
        System.out.println("VertexRDD neighborIdsRDD is:");
        for (int i = 0; i < neighborIdsRDD.collect().size(); i++) {
            System.out.println(
                    ((Tuple2<Object, long[]>) neighborIdsRDD.collect().get(i))._1() + " -- " +
                            Arrays.toString(((Tuple2<Object, long[]>) neighborIdsRDD.collect().get(i))._2())
            );
        }

        Dataset<Row> dr = spark_session.createDataFrame(neighborIdsRDD.rdd(), Tuple2.class);
        System.out.println("converted Dataset<Row> is:");
        dr.show();

但我得到一个空数据集，如下所示：

VertexRDD neighborIdsRDD is:
4 -- [3]
1 -- [2, 3]
5 -- [3, 2]
2 -- [1, 3, 5]
3 -- [1, 2, 5, 4]
converted Dataset<Row> is:
++
||
++
||
||
||
||
||
++

标签： javaapache-sparkdatasetspark-graphx

Answer 1

我遇到了同样的情况，但幸运的是我找到了找回Dataframe的解决方案。

解决方案代码在步骤[1]和[2]注释[3]。

GraphOps<String, String> graphOps = new GraphOps<>(graph, stringTag, stringTag);
System.out.println("VertexRDD neighborIdsRDD is:");
JavaRDD<Tuple2<Object, long[]>> neighborIdsRDD = graphOps.collectNeighborIds(EdgeDirection.Either()).toJavaRDD();
for (int i = 0; i < neighborIdsRDD.collect().size(); i++) {
    System.out.println(
            ((Tuple2<Object, long[]>) neighborIdsRDD.collect().get(i))._1() + " -- " +
                    Arrays.toString(((Tuple2<Object, long[]>) neighborIdsRDD.collect().get(i))._2())
    );
}

// [1] Define encoding schema
StructType graphStruct =  new StructType(new StructField[]{
        new StructField("father", DataTypes.LongType, false, Metadata.empty()),
        new StructField("children", DataTypes.createArrayType(DataTypes.LongType), false, Metadata.empty()),
});

// [2] Build a JavaRDD<Row> from a JavaRDD<Tuple2<Object,long[]>>
JavaRDD<Row> dr = neighborIdsRDD.map(tupla -> RowFactory.create(tupla._1(), tupla._2()));
        
// [3] Finally build the reqired Dataframe<Row>
Dataset<Row> dsr = spark_session.createDataFrame(dr.rdd(), graphStruct);

System.out.println("DATASET IS:");
dsr.show();

打印输出：

VertexRDD neighborIdsRDD is:
4 -- [3]
1 -- [2, 3]
5 -- [3, 2]
2 -- [1, 3, 5]
3 -- [1, 2, 5, 4]
DATASET IS:
+------+------------+
|father|    children|
+------+------------+
|     4|         [3]|
|     1|      [2, 3]|
|     5|      [3, 2]|
|     2|   [1, 3, 5]|
|     3|[1, 2, 5, 4]|
+------+------------+