【赵强老师】Flink的DataSet算子_JAVA_编程开发_程序员俱乐部

中国优秀的程序员网站程序员频道CXYCLUB技术地图
热搜:
更多>>
 
您所在的位置: 程序员俱乐部 > 编程开发 > JAVA > 【赵强老师】Flink的DataSet算子

【赵强老师】Flink的DataSet算子

 2020/10/9 16:06:08  collen7788  程序员俱乐部  我要评论(0)
  • 摘要:Flink为了能够处理有边界的数据集和无边界的数据集,提供了对应的DataSetAPI和DataStreamAPI。我们可以开发对应的Java程序或者Scala程序来完成相应的功能。下面举例了一些DataSetAPI中的基本的算子。下面我们通过具体的代码来为大家演示每个算子的作用。1、Map、FlatMap与MapPartition//获取运行环境ExecutionEnvironmentenv=ExecutionEnvironment.getExecutionEnvironment()
  • 标签:DataSet 老师 ASE

Flink为了能够处理有边界的数据集和无边界的数据集,提供了对应的DataSet API和DataStream API。我们可以开发对应的Java程序或者Scala程序来完成相应的功能。下面举例了一些DataSet API中的基本的算子。

下面我们通过具体的代码来为大家演示每个算子的作用。

1、Map、FlatMap与MapPartitionclass="language-java">

//获取运行环境
ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();

ArrayList<String> data = new ArrayList<String>();
data.add("I love Beijing");
data.add("I love China");
data.add("Beijing is the capital of China");
DataSource<String> text = env.fromCollection(data);

DataSet<List<String>> mapData = text.map(new MapFunction<String, List<String>>() {

	public List<String> map(String data) throws Exception {
		String[] words = data.split(" ");
		
		//创建一个List
		List<String> result = new ArrayList<String>();
		for(String w:words){
			result.add(w);
		}
		return result;
	}
});
mapData.print();
System.out.println("*****************************************");

DataSet<String> flatMapData = text.flatMap(new FlatMapFunction<String, String>() {

	public void flatMap(String data, Collector<String> collection) throws Exception {
		String[] words = data.split(" ");
		for(String w:words){
			collection.collect(w);
		}
	}
});
flatMapData.print();

System.out.println("*****************************************");
/*	new MapPartitionFunction<String, String>
	第一个String:表示分区中的数据元素类型
	第二个String:表示处理后的数据元素类型*/
DataSet<String> mapPartitionData = text.mapPartition(new MapPartitionFunction<String, String>() {

	public void mapPartition(Iterable<String> values, Collector<String> out) throws Exception {
		//针对分区进行操作的好处是:比如要进行数据库的操作,一个分区只需要创建一个Connection
		//values中保存了一个分区的数据
		 Iterator<String> it = values.iterator();
		while (it.hasNext()) {
			String next = it.next();
			String[] split = next.split(" ");
			for (String word : split) {
				out.collect(word);
			}
		}
		//关闭链接
	}
});
mapPartitionData.print();

?

2、Filter与Distinct

//获取运行环境
ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();

ArrayList<String> data = new ArrayList<String>();
data.add("I love Beijing");
data.add("I love China");
data.add("Beijing is the capital of China");
DataSource<String> text = env.fromCollection(data);

DataSet<String> flatMapData = text.flatMap(new FlatMapFunction<String, String>() {

	public void flatMap(String data, Collector<String> collection) throws Exception {
		String[] words = data.split(" ");
		for(String w:words){
			collection.collect(w);
		}
	}
});

//去掉重复的单词
flatMapData.distinct().print();
System.out.println("*********************");

//选出长度大于3的单词
flatMapData.filter(new FilterFunction<String>() {
	
	public boolean filter(String word) throws Exception {
		int length = word.length();
		return length>3?true:false;
	}
}).print();

?

3、Join操作

//获取运行的环境
ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();

//创建第一张表:用户ID  姓名
ArrayList<Tuple2<Integer, String>> data1 = new ArrayList<Tuple2<Integer,String>>();
data1.add(new Tuple2(1,"Tom"));
data1.add(new Tuple2(2,"Mike"));
data1.add(new Tuple2(3,"Mary"));
data1.add(new Tuple2(4,"Jone"));
//创建第二张表:用户ID 所在的城市
ArrayList<Tuple2<Integer, String>> data2 = new ArrayList<Tuple2<Integer,String>>();
data2.add(new Tuple2(1,"北京"));
data2.add(new Tuple2(2,"上海"));
data2.add(new Tuple2(3,"广州"));
data2.add(new Tuple2(4,"重庆"));

//实现join的多表查询:用户ID  姓名  所在的程序
DataSet<Tuple2<Integer, String>> table1 = env.fromCollection(data1);
DataSet<Tuple2<Integer, String>> table2 = env.fromCollection(data2);

table1.join(table2).where(0).equalTo(0)
/*第一个Tuple2<Integer,String>:表示第一张表
 * 第二个Tuple2<Integer,String>:表示第二张表
 * Tuple3<Integer,String, String>:多表join连接查询后的返回结果   */		                   
.with(new JoinFunction<Tuple2<Integer,String>, Tuple2<Integer,String>, Tuple3<Integer,String, String>>() {
	public Tuple3<Integer, String, String> join(Tuple2<Integer, String> table1,
			Tuple2<Integer, String> table2) throws Exception {
		return new Tuple3<Integer, String, String>(table1.f0,table1.f1,table2.f1);
	} }).print();

?

4、笛卡尔积

ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();

//创建第一张表:用户ID  姓名
ArrayList<Tuple2<Integer, String>> data1 = new ArrayList<Tuple2<Integer,String>>();
data1.add(new Tuple2(1,"Tom"));
data1.add(new Tuple2(2,"Mike"));
data1.add(new Tuple2(3,"Mary"));
data1.add(new Tuple2(4,"Jone"));

//创建第二张表:用户ID 所在的城市
ArrayList<Tuple2<Integer, String>> data2 = new ArrayList<Tuple2<Integer,String>>();
data2.add(new Tuple2(1,"北京"));
data2.add(new Tuple2(2,"上海"));
data2.add(new Tuple2(3,"广州"));
data2.add(new Tuple2(4,"重庆"));

//实现join的多表查询:用户ID  姓名  所在的程序
DataSet<Tuple2<Integer, String>> table1 = env.fromCollection(data1);
DataSet<Tuple2<Integer, String>> table2 = env.fromCollection(data2);

//生成笛卡尔积
table1.cross(table2).print();

?

5、First-N

ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();

//这里的数据是:员工姓名、薪水、部门号
DataSet<Tuple3<String, Integer,Integer>> grade = 
		env.fromElements(new Tuple3<String, Integer,Integer>("Tom",1000,10),
						 new Tuple3<String, Integer,Integer>("Mary",1500,20),
						 new Tuple3<String, Integer,Integer>("Mike",1200,30),
						 new Tuple3<String, Integer,Integer>("Jerry",2000,10));

//按照插入顺序取前三条记录
grade.first(3).print();
System.out.println("**********************");

//先按照部门号排序,在按照薪水排序
grade.sortPartition(2, Order.ASCENDING).sortPartition(1, Order.ASCENDING).print();
System.out.println("**********************");

//按照部门号分组,求每组的第一条记录
grade.groupBy(2).first(1).print();

?

6、外链接操作

ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();

//创建第一张表:用户ID  姓名
ArrayList<Tuple2<Integer, String>> data1 = new ArrayList<Tuple2<Integer,String>>();
data1.add(new Tuple2(1,"Tom"));
data1.add(new Tuple2(3,"Mary"));
data1.add(new Tuple2(4,"Jone"));

//创建第二张表:用户ID 所在的城市
ArrayList<Tuple2<Integer, String>> data2 = new ArrayList<Tuple2<Integer,String>>();
data2.add(new Tuple2(1,"北京"));
data2.add(new Tuple2(2,"上海"));
data2.add(new Tuple2(4,"重庆"));

//实现join的多表查询:用户ID  姓名  所在的程序
DataSet<Tuple2<Integer, String>> table1 = env.fromCollection(data1);
DataSet<Tuple2<Integer, String>> table2 = env.fromCollection(data2);

//左外连接
table1.leftOuterJoin(table2).where(0).equalTo(0)
	  .with(new JoinFunction<Tuple2<Integer,String>, Tuple2<Integer,String>, Tuple3<Integer,String,String>>() {

		public Tuple3<Integer, String, String> join(Tuple2<Integer, String> table1,
				Tuple2<Integer, String> table2) throws Exception {
			// 左外连接表示等号左边的信息会被包含
			if(table2 == null){
				return new Tuple3<Integer, String, String>(table1.f0,table1.f1,null);
			}else{
				return new Tuple3<Integer, String, String>(table1.f0,table1.f1,table2.f1);
			}
		}
	}).print();

System.out.println("***********************************");
//右外连接
table1.rightOuterJoin(table2).where(0).equalTo(0)
	  .with(new JoinFunction<Tuple2<Integer,String>, Tuple2<Integer,String>, Tuple3<Integer,String,String>>() {

		public Tuple3<Integer, String, String> join(Tuple2<Integer, String> table1,
				Tuple2<Integer, String> table2) throws Exception {
			//右外链接表示等号右边的表的信息会被包含
			if(table1 == null){
				return new Tuple3<Integer, String, String>(table2.f0,null,table2.f1);
			}else{
				return new Tuple3<Integer, String, String>(table2.f0,table1.f1,table2.f1);
			}
		}
	}).print();

System.out.println("***********************************");

//全外连接
table1.fullOuterJoin(table2).where(0).equalTo(0)
.with(new JoinFunction<Tuple2<Integer,String>, Tuple2<Integer,String>, Tuple3<Integer,String,String>>() {

	public Tuple3<Integer, String, String> join(Tuple2<Integer, String> table1, Tuple2<Integer, String> table2)
			throws Exception {
		if(table1 == null){
			return new Tuple3<Integer, String, String>(table2.f0,null,table2.f1);
		}else if(table2 == null){
			return new Tuple3<Integer, String, String>(table1.f0,table1.f1,null);
		}else{
			return new Tuple3<Integer, String, String>(table1.f0,table1.f1,table2.f1);
		}
	}
	
}).print();

?

发表评论
用户名: 匿名