spark不同结构Dataset合并

1.先将hdfs(或本地)存储的csv文件加载为Dataset
先在本地C盘准备两个csv文件
test.csv

client_id,behives,del,normal_status,cust_type,no_trd_days
7056,zl,1,hy,个人,2
7057,cf,1,hy,个人,12
7058,hs,2,hy,个人,1200
212121,0,sj,hy,个人,1100
212122,1,yx,hy,个人,100
212123,1,ls,hy,个人,100

test1.csv

cust_no,code,hight
sw7056,66661,125
es7057,66661,156
wq7058,66661,165

使用spark 加载，并为其创建视图

Dataset<Row> a = createRealView(session, "file:///C:\\\\test.csv");
Dataset<Row> b = createRealView(session, "file:///C:\\\\test1.csv");
a.createOrReplaceTempView("a");
b.createOrReplaceTempView("b");

由于这两个Dataset的结构和数据量均不一样，若要拼接为一个大的Dataset，可以把每个Dataset多生成一个自动增长编号的列，这里最快的方式是使用每行数据的索引号，则需要转换为RDD操作。

JavaPairRDD<Long, Row> aNewRDD = a.toJavaRDD().zipWithIndex().mapToPair(tuple -> {Long key = tuple._2;Row val = tuple._1;return new Tuple2<>(key, val);});JavaPairRDD<Long, Row> bNewRDD = b.toJavaRDD().zipWithIndex().mapToPair(tuple -> {Long key = tuple._2;Row val = tuple._1;return new Tuple2<>(key, val);});

原本zipWithIndex()之后的格式为JavaPairRDD<Row, Long> 但是我们需要转换一下结构为JavaPairRDD<Long, Row> 为后续的join操作做准备，因为join关联数据使用的是JavaPairRDD<T,M>中的T字段

JavaPairRDD<Long, Tuple2<Row, Row>> joinRDD = aNewRDD.join(bNewRDD);
joinRDD.collect().forEach(x -> System.out.println(x));

打印结果为

(0,([7056,zl,1,hy,个人,2],[sw7056,66661,125]))
(1,([7057,cf,1,hy,个人,12],[es7057,66661,156]))
(2,([7058,hs,2,hy,个人,1200],[wq7058,66661,165]))

那么接下来需要将Tuple2<Row, Row>中的两个Row合并为一个Row

JavaRDD<Row> rtRDD = joinRDD.map(tuple -> {List<Object> rowContent = new ArrayList<>();List<Object> tp1 = JavaConverters.seqAsJavaList(tuple._2._1.toSeq());List<Object> tp2 = JavaConverters.seqAsJavaList(tuple._2._2.toSeq());rowContent.addAll(tp1);rowContent.addAll(tp2);Seq<Object> rtSeq = JavaConverters.asScalaIteratorConverter(rowContent.iterator()).asScala().toSeq();return Row.fromSeq(rtSeq);});

这样就把JavaPairRDD<Long, Tuple2<Row, Row>>转换为JavaRDD<Row>，然后就需要转为Dataset<Row>
由于程序并不知道Dataset的数据结构，则需要建立结果的结构所需的schema

StructType schema = new StructType(new StructField[]{new StructField("client_id", DataTypes.StringType, false, Metadata.empty()),new StructField("behives", DataTypes.StringType, false, Metadata.empty()),new StructField("del", DataTypes.StringType, false, Metadata.empty()),new StructField("normal_status", DataTypes.StringType, false, Metadata.empty()),new StructField("cust_type", DataTypes.StringType, false, Metadata.empty()),new StructField("no_trd_days", DataTypes.StringType, false, Metadata.empty()),new StructField("cust_no", DataTypes.StringType, false, Metadata.empty()),new StructField("code", DataTypes.StringType, false, Metadata.empty()),new StructField("hight", DataTypes.StringType, false, Metadata.empty())});

然后创建出新的Dataset<Row>

Dataset<Row> dataView = session.createDataFrame(rtRDD, schema);
dataView.show();

最终结果是

+---------+-------+---+-------------+---------+-----------+-------+-----+-----+
|client_id|behives|del|normal_status|cust_type|no_trd_days|cust_no| code|hight|
+---------+-------+---+-------------+---------+-----------+-------+-----+-----+
|     7056|     zl|  1|           hy|     个人|          2| sw7056|66661|  125|
|     7057|     cf|  1|           hy|     个人|         12| es7057|66661|  156|
|     7058|     hs|  2|           hy|     个人|       1200| wq7058|66661|  165|
+---------+-------+---+-------------+---------+-----------+-------+-----+-----+

完整代码：

public void createView(SparkSession session, String portraitPath, String prodPath) {Dataset<Row> a = createRealView(session, "file:///C:\\\\test.csv");Dataset<Row> rowDataset = a.unionAll(a).unionAll(a).orderBy(functions.rand());a.createOrReplaceTempView("a");Dataset<Row> b = createRealView(session, "file:///C:\\\\test1.csv");b.createOrReplaceTempView("b");JavaPairRDD<Long, Row> aNewRDD = a.toJavaRDD().zipWithIndex().mapToPair(tuple -> {Long key = tuple._2;Row val = tuple._1;return new Tuple2<>(key, val);});JavaPairRDD<Long, Row> bNewRDD = b.toJavaRDD().zipWithIndex().mapToPair(tuple -> {Long key = tuple._2;Row val = tuple._1;return new Tuple2<>(key, val);});JavaPairRDD<Long, Tuple2<Row, Row>> joinRDD = aNewRDD.join(bNewRDD);joinRDD.collect().forEach(x -> System.out.println(x));StructType schema = new StructType(new StructField[]{new StructField("client_id", DataTypes.StringType, false, Metadata.empty()),new StructField("behives", DataTypes.StringType, false, Metadata.empty()),new StructField("del", DataTypes.StringType, false, Metadata.empty()),new StructField("normal_status", DataTypes.StringType, false, Metadata.empty()),new StructField("cust_type", DataTypes.StringType, false, Metadata.empty()),new StructField("no_trd_days", DataTypes.StringType, false, Metadata.empty()),new StructField("cust_no", DataTypes.StringType, false, Metadata.empty()),new StructField("code", DataTypes.StringType, false, Metadata.empty()),new StructField("hight", DataTypes.StringType, false, Metadata.empty())});JavaRDD<Row> rtRDD = joinRDD.map(tuple -> {List<Object> rowContent = new ArrayList<>();List<Object> tp1 = JavaConverters.seqAsJavaList(tuple._2._1.toSeq());List<Object> tp2 = JavaConverters.seqAsJavaList(tuple._2._2.toSeq());rowContent.addAll(tp1);rowContent.addAll(tp2);Seq<Object> rtSeq = JavaConverters.asScalaIteratorConverter(rowContent.iterator()).asScala().toSeq();return Row.fromSeq(rtSeq);});Dataset<Row> dataView = session.createDataFrame(rtRDD, schema);dataView.show();}private Dataset<Row> createRealView(SparkSession session, String hdfsPath) {Dataset<Row> wideTableDF = null;try {wideTableDF = session.read().format("csv").option("header", "true").load(hdfsPath);} catch (Exception e) {System.err.println("未找到有效的宽表数据，查找路径为：" + hdfsPath);}return wideTableDF;}