Uncover the mystery, will flow will be big data

If you know the flow of any language, it makes no sense not to develop big data.

As the saying goes, men chase women across the mountain, and women chase men through layers of gauze. If you learn big data with zero foundation, you feel that there is a mountain ahead, then as long as you know java or the flow of any language, big data is only separated by a layer of veil.
This article takes java stream computing as an example to explain some basic spark operations. The same goes for flink, another popular big data framework.

Preparation

For test data, the following columns represent name, age, department, and position respectively.

张三,20,研发部,普通员工
李四,31,研发部,普通员工
李丽,36,财务部,普通员工
张伟,38,研发部,经理
杜航,25,人事部,普通员工
周歌,28,研发部,普通员工

Create a Employeeclass.

    @Getter
    @Setter
    @AllArgsConstructor
    @NoArgsConstructor
    @ToString
    static
    class Employee implements Serializable {
        private String name;
        private Integer age;
        private String department;
        private String level;
    }
}

Version: jdk: 1.8 spark: 3.2.0 scala: 2.12.15
The above scala version is just that the spark framework itself needs to depend on scala.
Because scala is indeed a relatively small language, this article still uses java to demonstrate spark code.

1. map class

1.1 java stream map

map represents a one-to-one operation. Perform any operation on a row of upstream data, and finally get a piece of data after operation.
This kind of thinking is consistent in java, spark, and flink.

We first use java stream to demonstrate reading files, and then use the map operation to map each line of data into Employeeobjects.

List<String> list = FileUtils.readLines(new File("f:/test.txt"), "utf-8");
        List<Employee> employeeList = list.stream().map(word -> {
            List<String> words = Arrays.stream(word.split(",")).collect(Collectors.toList());
            Employee employee = new Employee(words.get(0), Integer.parseInt(words.get(1)), words.get(2), words.get(3));
            return employee;
        }).collect(Collectors.toList());

        employeeList.forEach(System.out::println);

Transformed data:

JavaStreamDemo.Employee(name=张三, age=20, department=研发部, level=普通员工)
JavaStreamDemo.Employee(name=李四, age=31, department=研发部, level=普通员工)
JavaStreamDemo.Employee(name=李丽, age=36, department=财务部, level=普通员工)
JavaStreamDemo.Employee(name=张伟, age=38, department=研发部, level=经理)
JavaStreamDemo.Employee(name=杜航, age=25, department=人事部, level=普通员工)
JavaStreamDemo.Employee(name=周歌, age=28, department=研发部, level=普通员工)

1.2 spark map

First get a SparkSession object, read the file, and get a DataSet elastic data set object.

SparkSession session = SparkSession.builder().master("local[*]").getOrCreate();
Dataset<Row> reader = session.read().text("F:/test.txt");
reader.show();

The show() here is to print out the current data set, which is an operator of the action class.
got the answer:

+-----------------------+
|                  value|
+-----------------------+
|张三,20,研发部,普通员工|
|李四,31,研发部,普通员工|
|李丽,36,财务部,普通员工|
|    张伟,38,研发部,经理|
|杜航,25,人事部,普通员工|
|周歌,28,研发部,普通员工|
+-----------------------+

现在我们拿到了基础数据，我们使用map一对一操作，将一行行数据转换为Employee对象。
我们这里不使用lamda表达式，让大家看得更加清晰。
这里实现了MapFunction接口里的call方法，每次拿到一行数据，我们这里进行切分，再转换为对象。

需要特别指出的一点是，与后端WEB应用有一个统一异常处理不同的是，大数据应用，特别是流式计算，要保证7*24在线，需要对每个算子进行异常捕获。
因为你不知道上游数据清洗到底怎么样，很可能拿到一条脏数据，处理的时候抛出异常，如果没有捕获处理，那么整个应用就会挂掉。
spark的算子分为Transformation和Action两种类型。Transformation会开成一个DAG图，具有lazy延迟性，它只会从一个dataset(rdd/df)转换成另一个dataset(rdd/df)，只有当遇到action类的算子才会真正执行。我们今天会演示的算子都是Transformation类的算子。
典型的Action算子包括show,collect,save之类的。比如在本地进行show查看结果，或者完成运行后save到数据库，或者HDFS。
spark执行时分为driver和executor。但不是本文的重点，不会展开讲。只需要注意driver端会将代码分发到各个分布式系统的节点executor上，它本身不会参与计算。一般来说，算子外部，如以下示例代码的a处会在driver端执行，b处算子内部会不同服务器上的executor端执行。所以在算子外部定义的变量，在算子内部使用的时候要特别注意！！不要想当然地以为都是一个main方法里写的代码，就一定会在同一个JVM里。
这里涉及到序列化的问题，同时它们分处不同的JVM，使用"=="比较的时候也可能会出问题！！
这是一个后端WEB开发转向大数据开发时，这个思想一定要转变过来。
简言之，后端WEB服务的分布式是我们自己实现的，大数据的分布式是框架天生帮我们实现的。

1.2.1 MapFunction

// a 算子外部，driver端
Dataset<Employee> employeeDataset = reader.map(new MapFunction<Row, Employee>() {
            @Override
            public Employee call(Row row) throws Exception {
                // b 算子内部，executor端
                Employee employee = null;
                try {
                    // gson.fromJson(); 这里使用gson涉及到序列化问题
                    List<String> list = Arrays.stream(row.mkString().split(",")).collect(Collectors.toList());
                    employee = new Employee(list.get(0), Integer.parseInt(list.get(1)), list.get(2), list.get(3));
                } catch (Exception exception) {
                    // 日志记录
                    // 流式计算中要做到7*24小时不间断，任意一条上流脏数据都可能导致失败，从而导致任务退出，所以这里要做好异常的抓取
                    exception.printStackTrace();
                }
                return employee;
            }
        }, Encoders.bean(Employee.class));

        employeeDataset.show();

输出

+---+----------+--------+----+
|age|department|   level|name|
+---+----------+--------+----+
| 20|    研发部|普通员工|张三|
| 31|    研发部|普通员工|李四|
| 36|    财务部|普通员工|李丽|
| 38|    研发部|    经理|张伟|
| 25|    人事部|普通员工|杜航|
| 28|    研发部|普通员工|周歌|

1.2.2 MapPartitionsFunction

spark中 map和mapPartitions有啥区别？
map是1条1条处理数据
mapPartitions是一个分区一个分区处理数据

后者一定比前者效率高吗？
不一定，看具体情况。

这里使用前面 map 一样的逻辑处理。可以看到在call方法里得到的是一个Iterator迭代器，是一批数据。
得到一批数据，然后再一对一映射为对象，再以Iterator的形式返回这批数据。

Dataset<Employee> employeeDataset2 = reader.mapPartitions(new MapPartitionsFunction<Row, Employee>() {
            @Override
            public Iterator<Employee> call(Iterator<Row> iterator) throws Exception {
                List<Employee> employeeList = new ArrayList<>();
                while (iterator.hasNext()){
                    Row row = iterator.next();
                    try {
                        List<String> list = Arrays.stream(row.mkString().split(",")).collect(Collectors.toList());
                        Employee employee = new Employee(list.get(0), Integer.parseInt(list.get(1)), list.get(2), list.get(3));
                        employeeList.add(employee);
                    } catch (Exception exception) {
                        // 日志记录
                        // 流式计算中要做到7*24小时不间断，任意一条上流脏数据都可能导致失败，从而导致任务退出，所以这里要做好异常的抓取
                        exception.printStackTrace();
                    }
                }
                return employeeList.iterator();
            }
        }, Encoders.bean(Employee.class));

        employeeDataset2.show();

输出结果跟map一样，这里就不贴出来了。

2.flatMap类

map和flatMap有什么区别？
map是一对一，flatMap是一对多。当然在java stream中，flatMap叫法叫做扁平化。

这种思想，在java和spark,flink都是一致的。

2.1 java stream flatMap

以下代码将1条原始数据映射到2个对象上并返回。

List<Employee> employeeList2 = list.stream().flatMap(word -> {
            List<String> words = Arrays.stream(word.split(",")).collect(Collectors.toList());
            List<Employee> lists = new ArrayList<>();
            Employee employee = new Employee(words.get(0), Integer.parseInt(words.get(1)), words.get(2), words.get(3));
            lists.add(employee);
            Employee employee2 = new Employee(words.get(0)+"_2", Integer.parseInt(words.get(1)), words.get(2), words.get(3));
            lists.add(employee2);
            return lists.stream();
        }).collect(Collectors.toList());
        employeeList2.forEach(System.out::println);

输出

JavaStreamDemo.Employee(name=张三, age=20, department=研发部, level=普通员工)
JavaStreamDemo.Employee(name=张三_2, age=20, department=研发部, level=普通员工)
JavaStreamDemo.Employee(name=李四, age=31, department=研发部, level=普通员工)
JavaStreamDemo.Employee(name=李四_2, age=31, department=研发部, level=普通员工)
JavaStreamDemo.Employee(name=李丽, age=36, department=财务部, level=普通员工)
JavaStreamDemo.Employee(name=李丽_2, age=36, department=财务部, level=普通员工)
JavaStreamDemo.Employee(name=张伟, age=38, department=研发部, level=经理)
JavaStreamDemo.Employee(name=张伟_2, age=38, department=研发部, level=经理)
JavaStreamDemo.Employee(name=杜航, age=25, department=人事部, level=普通员工)
JavaStreamDemo.Employee(name=杜航_2, age=25, department=人事部, level=普通员工)
JavaStreamDemo.Employee(name=周歌, age=28, department=研发部, level=普通员工)
JavaStreamDemo.Employee(name=周歌_2, age=28, department=研发部, level=普通员工)

2.2 spark flatMap

这里实现FlatMapFunction的call方法，一次拿到1条数据，然后返回值是Iterator，所以可以返回多条。

Dataset<Employee> employeeDatasetFlatmap = reader.flatMap(new FlatMapFunction<Row, Employee>() {
            @Override
            public Iterator<Employee> call(Row row) throws Exception {
                List<Employee> employeeList = new ArrayList<>();
                try {
                    List<String> list = Arrays.stream(row.mkString().split(",")).collect(Collectors.toList());
                    Employee employee = new Employee(list.get(0), Integer.parseInt(list.get(1)), list.get(2), list.get(3));
                    employeeList.add(employee);

                    Employee employee2 = new Employee(list.get(0)+"_2", Integer.parseInt(list.get(1)), list.get(2), list.get(3));
                    employeeList.add(employee2);
                } catch (Exception exception) {
                    exception.printStackTrace();
                }
                return employeeList.iterator();
            }
        }, Encoders.bean(Employee.class));
        employeeDatasetFlatmap.show();

输出

+---+----------+--------+------+
|age|department|   level|  name|
+---+----------+--------+------+
| 20|    研发部|普通员工|  张三|
| 20|    研发部|普通员工|张三_2|
| 31|    研发部|普通员工|  李四|
| 31|    研发部|普通员工|李四_2|
| 36|    财务部|普通员工|  李丽|
| 36|    财务部|普通员工|李丽_2|
| 38|    研发部|    经理|  张伟|
| 38|    研发部|    经理|张伟_2|
| 25|    人事部|普通员工|  杜航|
| 25|    人事部|普通员工|杜航_2|
| 28|    研发部|普通员工|  周歌|
| 28|    研发部|普通员工|周歌_2|
+---+----------+--------+------+

3 groupby类

与SQL类似，java stream流和spark一样，groupby对数据集进行分组并在此基础上可以进行聚合函数操作。也可以分组直接得到一组子数据集。

3.1 java stream groupBy

按部门分组统计部门人数：

Map<String, Long> map = employeeList.stream().collect(Collectors.groupingBy(Employee::getDepartment, Collectors.counting()));
        System.out.println(map);

输出

{财务部=1, 人事部=1, 研发部=4}

3.2 spark groupBy

将映射为对象的数据集按部门分组，在此基础上统计部门员工数和平均年龄。

RelationalGroupedDataset datasetGroupBy = employeeDataset.groupBy("department");
// 统计每个部门有多少员工
datasetGroupBy.count().show(); 
/**
 * 每个部门的平均年龄
 */
datasetGroupBy.avg("age").withColumnRenamed("avg(age)","avgAge").show();

输出分别为

+----------+-----+
|department|count|
+----------+-----+
|    财务部|    1|
|    人事部|    1|
|    研发部|    4|
+----------+-----+

+----------+------+
|department|avgAge|
+----------+------+
|    财务部|  36.0|
|    人事部|  25.0|
|    研发部| 29.25|
+----------+------+

3.3 spark groupByKey

spark的groupBy和groupByKey的区别,前者在此基础上使用聚合函数得到一个聚合值，后者只是进行分组，不进行任何计算。
类似于java stream的：

Map<String, List<Employee>> map2 = employeeList.stream().collect(Collectors.groupingBy(Employee::getDepartment));
System.out.println(map2);

输出

{财务部=[JavaStreamDemo.Employee(name=李丽, age=36, department=财务部, level=普通员工)], 
人事部=[JavaStreamDemo.Employee(name=杜航, age=25, department=人事部, level=普通员工)], 
研发部=[JavaStreamDemo.Employee(name=张三, age=20, department=研发部, level=普通员工), JavaStreamDemo.Employee(name=李四, age=31, department=研发部, level=普通员工), JavaStreamDemo.Employee(name=张伟, age=38, department=研发部, level=经理), JavaStreamDemo.Employee(name=周歌, age=28, department=研发部, level=普通员工)]}

使用spark groupByKey。
先得到一个key-value的一对多的一个集合数据集。这里的call()方法返回的是key,即分组的key。

KeyValueGroupedDataset keyValueGroupedDataset = employeeDataset.groupByKey(new MapFunction<Employee, String>() {
            @Override
            public String call(Employee employee) throws Exception {
                // 返回分组的key,这里表示根据部门进行分组
                return employee.getDepartment();
            }
        }, Encoders.STRING());

再在keyValueGroupedDataset 的基础上进行mapGroups，在call()方法里就可以拿到每个key的所有原始数据。

keyValueGroupedDataset.mapGroups(new MapGroupsFunction() {
            @Override
            public Object call(Object key, Iterator iterator) throws Exception {
                System.out.println("key = " + key);
                while (iterator.hasNext()){
                    System.out.println(iterator.next());
                }
                return iterator; 
            }
        }, Encoders.bean(Iterator.class))
                .show(); // 这里的show()没有意义，只是触发计算而已

输出

key = 人事部
SparkDemo.Employee(name=杜航, age=25, department=人事部, level=普通员工)
key = 研发部
SparkDemo.Employee(name=张三, age=20, department=研发部, level=普通员工)
SparkDemo.Employee(name=李四, age=31, department=研发部, level=普通员工)
SparkDemo.Employee(name=张伟, age=38, department=研发部, level=经理)
SparkDemo.Employee(name=周歌, age=28, department=研发部, level=普通员工)
key = 财务部
SparkDemo.Employee(name=李丽, age=36, department=财务部, level=普通员工)

4 reduce类

reduce的字面意思是：减少；减小；降低；缩小。
又叫归约。

它将数据集进行循环，让当前对象和前一对象两两进行计算，每次计算得到的结果作为下一次计算的前一对象，并最终得到一个对象。
假设有5个数据【1，2，3，4，5】，使用reduce进行求和计算，分别是

企业微信截图_16825817781997.png
比如上面的测试数据集，我要计算各部门年龄总数。使用聚合函数得到的是一个int类型的数字。

4.1 java stream reduce

int age = employeeList.stream().mapToInt(e -> e.age).sum();
System.out.println(age);//178

使用reduce也可进行上面的计算

int age1 = employeeList.stream().mapToInt(e -> e.getAge()).reduce(0,(a,b) -> a+b);
System.out.println(age1);// 178

但是我将年龄求和，同时得到一个完整的对象呢？

JavaStreamDemo.Employee(name=周歌, age=178, department=研发部, level=普通员工)

可以使用reduce将数据集两两循环，将年龄相加，同时返回最后一个遍历的对象。
下面代码的pre 代表前一个对象,current 代表当前对象。

 /**
 * pre 代表前一个对象
 * current 代表当前对象
 */
Employee reduceEmployee = employeeList.stream().reduce(new Employee(), (pre,current) -> {
     // 当第一次循环时前一个对象为null
    if (pre.getAge() == null) {
        current.setAge(current.getAge());
    } else {
        current.setAge(pre.getAge() + current.getAge());
    }
    return current;
});
System.out.println(reduceEmployee);

4.2 spark reduce

spark reduce的基本思想跟java stream是一样的。
直接看代码：

 Employee datasetReduce = employeeDataset.reduce(new ReduceFunction<Employee>() {
            @Override
            public Employee call(Employee t1, Employee t2) throws Exception {
                // 不同的版本看是否需要判断t1 == null
                t2.setAge(t1.getAge() + t2.getAge());
                return t2;
            }
        });

        System.out.println(datasetReduce);

输出

SparkDemo.Employee(name=周歌, age=178, department=研发部, level=普通员工)

其它常见操作类

Employee employee = employeeDataset.filter("age > 30").limit(3).sort("age").first();
System.out.println(employee);
// SparkDemo.Employee(name=李四, age=31, department=研发部, level=普通员工)

同时可以将dataset注册成table，使用更为强大的SQL来进行各种强大的运算。
现在SQL是flink的一等公民，spark也不遑多让。
这里举一个非常简单的例子。

employeeDataset.registerTempTable("table");
session.sql("select * from table where age > 30 order by age desc limit 3").show();

输出

+---+----------+--------+----+
|age|department|   level|name|
+---+----------+--------+----+
| 38|    研发部|    经理|张伟|
| 36|    财务部|普通员工|李丽|
| 31|    研发部|普通员工|李四|
+---+----------+--------+----+

employeeDataset.registerTempTable("table");
session.sql("select 
            concat_ws(',',collect_set(name)) as names, // group_concat
            avg(age) as age,
            department from table 
            where age > 30  
            group by department 
            order by age desc 
            limit 3").show();

输出

+---------+----+----------+
|    names| age|department|
+---------+----+----------+
|     李丽|36.0|    财务部|
|张伟,李四|34.5|    研发部|
+---------+----+----------+

小结

本文依据java stream的相似性，介绍了spark里面一些常见的算子操作。
本文只是做一个非常简单的入门介绍。
如果感兴趣的话，后端的同学可以尝试着操作一下，非常简单，本地不需要搭建环境，只要引入spark 的 maven依赖即可。
我把本文的所有代码全部贴在最后面。

java stream 源码：

点击查看代码

import lombok.*;
import org.apache.commons.io.FileUtils;

import java.io.File;
import java.io.IOException;
import java.io.Serializable;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.List;
import java.util.Map;
import java.util.stream.Collectors;

public class JavaStreamDemo {
    public static void main(String[] args) throws IOException {
        /**
         * 张三,20,研发部,普通员工
         * 李四,31,研发部,普通员工
         * 李丽,36,财务部,普通员工
         * 张伟,38,研发部,经理
         * 杜航,25,人事部,普通员工
         * 周歌,28,研发部,普通员工
         */
        List<String> list = FileUtils.readLines(new File("f:/test.txt"), "utf-8");
        List<Employee> employeeList = list.stream().map(word -> {
            List<String> words = Arrays.stream(word.split(",")).collect(Collectors.toList());
            Employee employee = new Employee(words.get(0), Integer.parseInt(words.get(1)), words.get(2), words.get(3));
            return employee;
        }).collect(Collectors.toList());

        // employeeList.forEach(System.out::println);

        List<Employee> employeeList2 = list.stream().flatMap(word -> {
            List<String> words = Arrays.stream(word.split(",")).collect(Collectors.toList());
            List<Employee> lists = new ArrayList<>();
            Employee employee = new Employee(words.get(0), Integer.parseInt(words.get(1)), words.get(2), words.get(3));
            lists.add(employee);
            Employee employee2 = new Employee(words.get(0)+"_2", Integer.parseInt(words.get(1)), words.get(2), words.get(3));
            lists.add(employee2);
            return lists.stream();
        }).collect(Collectors.toList());
        // employeeList2.forEach(System.out::println);

        Map<String, Long> map = employeeList.stream().collect(Collectors.groupingBy(Employee::getDepartment, Collectors.counting()));
        System.out.println(map);
        Map<String, List<Employee>> map2 = employeeList.stream().collect(Collectors.groupingBy(Employee::getDepartment));
        System.out.println(map2);

        int age = employeeList.stream().mapToInt(e -> e.age).sum();
        System.out.println(age);// 178

        int age1 = employeeList.stream().mapToInt(e -> e.getAge()).reduce(0,(a,b) -> a+b);
        System.out.println(age1);// 178

        /**
         * pre 代表前一个对象
         * current 代表当前对象
         */
        Employee reduceEmployee = employeeList.stream().reduce(new Employee(), (pre,current) -> {
            if (pre.getAge() == null) {
                current.setAge(current.getAge());
            } else {
                current.setAge(pre.getAge() + current.getAge());
            }
            return current;
        });
        System.out.println(reduceEmployee);




    }

    @Getter
    @Setter
    @AllArgsConstructor
    @NoArgsConstructor
    @ToString
    static
    class Employee implements Serializable {
        private String name;
        private Integer age;
        private String department;
        private String level;
    }
}

spark的源码：

点击查看代码

import com.google.gson.Gson;
import lombok.*;
import org.apache.spark.api.java.function.*;
import org.apache.spark.sql.*;

import java.io.Serializable;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Iterator;
import java.util.List;
import java.util.stream.Collectors;

public class SparkDemo {
    public static void main(String[] args) {
        SparkSession session = SparkSession.builder().master("local[*]").getOrCreate();
        Dataset<Row> reader = session.read().text("F:/test.txt");
        // reader.show();
        /**
         * +-----------------------+
         * |                  value|
         * +-----------------------+
         * |张三,20,研发部,普通员工|
         * |李四,31,研发部,普通员工|
         * |李丽,36,财务部,普通员工|
         * |张伟,38,研发部,经理|
         * |杜航,25,人事部,普通员工|
         * |周歌,28,研发部,普通员工|
         * +-----------------------+
         */

        // 本地演示而已，实际分布式环境，这里的gson涉及到序列化问题
        // 算子以外的代码都在driver端运行
        // 任何算子以内的代码都在executor端运行,即会在不同的服务器节点上执行
        Gson gson = new Gson();
        // a 算子外部，driver端
        Dataset<Employee> employeeDataset = reader.map(new MapFunction<Row, Employee>() {
            @Override
            public Employee call(Row row) throws Exception {
                // b 算子内部，executor端
                Employee employee = null;
                try {
                    // gson.fromJson(); 这里使用gson涉及到序列化问题
                    List<String> list = Arrays.stream(row.mkString().split(",")).collect(Collectors.toList());
                    employee = new Employee(list.get(0), Integer.parseInt(list.get(1)), list.get(2), list.get(3));
                } catch (Exception exception) {
                    // 日志记录
                    // 流式计算中要做到7*24小时不间断，任意一条上流脏数据都可能导致失败，从而导致任务退出，所以这里要做好异常的抓取
                    exception.printStackTrace();
                }
                return employee;
            }
        }, Encoders.bean(Employee.class));

        // employeeDataset.show();
        /**
         * +---+----------+--------+----+
         * |age|department|   level|name|
         * +---+----------+--------+----+
         * | 20|    研发部|普通员工|张三|
         * | 31|    研发部|普通员工|李四|
         * | 36|    财务部|普通员工|李丽|
         * | 38|    研发部|    经理|张伟|
         * | 25|    人事部|普通员工|杜航|
         * | 28|    研发部|普通员工|周歌|
         */

        Dataset<Employee> employeeDataset2 = reader.mapPartitions(new MapPartitionsFunction<Row, Employee>() {
            @Override
            public Iterator<Employee> call(Iterator<Row> iterator) throws Exception {
                List<Employee> employeeList = new ArrayList<>();
                while (iterator.hasNext()){
                    Row row = iterator.next();
                    try {
                        List<String> list = Arrays.stream(row.mkString().split(",")).collect(Collectors.toList());
                        Employee employee = new Employee(list.get(0), Integer.parseInt(list.get(1)), list.get(2), list.get(3));
                        employeeList.add(employee);
                    } catch (Exception exception) {
                        // 日志记录
                        // 流式计算中要做到7*24小时不间断，任意一条上流脏数据都可能导致失败，从而导致任务退出，所以这里要做好异常的抓取
                        exception.printStackTrace();
                    }
                }
                return employeeList.iterator();
            }
        }, Encoders.bean(Employee.class));

        // employeeDataset2.show();
        /**
         * +---+----------+--------+----+
         * |age|department|   level|name|
         * +---+----------+--------+----+
         * | 20|    研发部|普通员工|张三|
         * | 31|    研发部|普通员工|李四|
         * | 36|    财务部|普通员工|李丽|
         * | 38|    研发部|    经理|张伟|
         * | 25|    人事部|普通员工|杜航|
         * | 28|    研发部|普通员工|周歌|
         * +---+----------+--------+----+
         */

        Dataset<Employee> employeeDatasetFlatmap = reader.flatMap(new FlatMapFunction<Row, Employee>() {
            @Override
            public Iterator<Employee> call(Row row) throws Exception {
                List<Employee> employeeList = new ArrayList<>();
                try {
                    List<String> list = Arrays.stream(row.mkString().split(",")).collect(Collectors.toList());
                    Employee employee = new Employee(list.get(0), Integer.parseInt(list.get(1)), list.get(2), list.get(3));
                    employeeList.add(employee);

                    Employee employee2 = new Employee(list.get(0)+"_2", Integer.parseInt(list.get(1)), list.get(2), list.get(3));
                    employeeList.add(employee2);
                } catch (Exception exception) {
                    exception.printStackTrace();
                }
                return employeeList.iterator();
            }
        }, Encoders.bean(Employee.class));
//        employeeDatasetFlatmap.show();
        /**
         * +---+----------+--------+------+
         * |age|department|   level|  name|
         * +---+----------+--------+------+
         * | 20|    研发部|普通员工|  张三|
         * | 20|    研发部|普通员工|张三_2|
         * | 31|    研发部|普通员工|  李四|
         * | 31|    研发部|普通员工|李四_2|
         * | 36|    财务部|普通员工|  李丽|
         * | 36|    财务部|普通员工|李丽_2|
         * | 38|    研发部|    经理|  张伟|
         * | 38|    研发部|    经理|张伟_2|
         * | 25|    人事部|普通员工|  杜航|
         * | 25|    人事部|普通员工|杜航_2|
         * | 28|    研发部|普通员工|  周歌|
         * | 28|    研发部|普通员工|周歌_2|
         * +---+----------+--------+------+
         */

        RelationalGroupedDataset datasetGroupBy = employeeDataset.groupBy("department");
        // 统计每个部门有多少员工
        // datasetGroupBy.count().show();
        /**
         * +----------+-----+
         * |department|count|
         * +----------+-----+
         * |    财务部|    1|
         * |    人事部|    1|
         * |    研发部|    4|
         * +----------+-----+
         */
        /**
         * 每个部门的平均年龄
         */
        // datasetGroupBy.avg("age").withColumnRenamed("avg(age)","avgAge").show();
        /**
         * +----------+--------+
         * |department|avg(age)|
         * +----------+--------+
         * |    财务部|    36.0|
         * |    人事部|    25.0|
         * |    研发部|   29.25|
         * +----------+--------+
         */

        KeyValueGroupedDataset keyValueGroupedDataset = employeeDataset.groupByKey(new MapFunction<Employee, String>() {
            @Override
            public String call(Employee employee) throws Exception {
                // 返回分组的key,这里表示根据部门进行分组
                return employee.getDepartment();
            }
        }, Encoders.STRING());

        keyValueGroupedDataset.mapGroups(new MapGroupsFunction() {
            @Override
            public Object call(Object key, Iterator iterator) throws Exception {
                System.out.println("key = " + key);
                while (iterator.hasNext()){
                    System.out.println(iterator.next());
                }
                return iterator;
                /**
                 * key = 人事部
                 * SparkDemo.Employee(name=杜航, age=25, department=人事部, level=普通员工)
                 * key = 研发部
                 * SparkDemo.Employee(name=张三, age=20, department=研发部, level=普通员工)
                 * SparkDemo.Employee(name=李四, age=31, department=研发部, level=普通员工)
                 * SparkDemo.Employee(name=张伟, age=38, department=研发部, level=经理)
                 * SparkDemo.Employee(name=周歌, age=28, department=研发部, level=普通员工)
                 * key = 财务部
                 * SparkDemo.Employee(name=李丽, age=36, department=财务部, level=普通员工)
                 */
            }
        }, Encoders.bean(Iterator.class))
                .show(); // 这里的show()没有意义，只是触发计算而已


        Employee datasetReduce = employeeDataset.reduce(new ReduceFunction<Employee>() {
            @Override
            public Employee call(Employee t1, Employee t2) throws Exception {
                // 不同的版本看是否需要判断t1 == null
                t2.setAge(t1.getAge() + t2.getAge());
                return t2;
            }
        });

        System.out.println(datasetReduce);


        Employee employee = employeeDataset.filter("age > 30").limit(3).sort("age").first();
        System.out.println(employee);
        // SparkDemo.Employee(name=李四, age=31, department=研发部, level=普通员工)

        employeeDataset.registerTempTable("table");
        session.sql("select * from table where age > 30 order by age desc limit 3").show();

        /**
         * +---+----------+--------+----+
         * |age|department|   level|name|
         * +---+----------+--------+----+
         * | 38|    研发部|    经理|张伟|
         * | 36|    财务部|普通员工|李丽|
         * | 31|    研发部|普通员工|李四|
         * +---+----------+--------+----+
         */


    }

    @Getter
    @Setter
    @AllArgsConstructor
    @NoArgsConstructor
    @ToString
    public static class Employee implements Serializable {
        private String name;
        private Integer age;
        private String department;
        private String level;
    }
}

spark maven依赖,自行不需要的spark-streaming,kafka依赖去掉。

点击查看代码

<properties>
        <maven.compiler.source>8</maven.compiler.source>
        <maven.compiler.target>8</maven.compiler.target>
        <maven.compiler.source>1.8</maven.compiler.source>
        <maven.compiler.target>1.8</maven.compiler.target>
        <scala.version>2.12.15</scala.version>
        <spark.version>3.2.0</spark.version>
        <encoding>UTF-8</encoding>
    </properties>
    <dependencies>
        <!-- scala依赖-->
        <dependency>
            <groupId>org.scala-lang</groupId>
            <artifactId>scala-library</artifactId>
            <version>${scala.version}</version>
        </dependency>
        <!-- spark依赖-->
        <dependency>
            <groupId>org.apache.spark</groupId>
            <artifactId>spark-sql_2.12</artifactId>
            <version>${spark.version}</version>
        </dependency>

        <dependency>
            <groupId>org.projectlombok</groupId>
            <artifactId>lombok</artifactId>
            <version>1.18.2</version>
            <scope>provided</scope>
        </dependency>

        <!--<dependency>
            <groupId>org.apache.spark</groupId>
            <artifactId>spark-streaming-kafka-0-10_2.12</artifactId>
            <version>${spark.version}</version>
        </dependency>-->

        <dependency>
            <groupId>org.apache.spark</groupId>
            <artifactId>spark-sql-kafka-0-10_2.12</artifactId>
            <version>${spark.version}</version>
        </dependency>

        <dependency>
            <groupId>com.google.code.gson</groupId>
            <artifactId>gson</artifactId>
            <version>2.7</version>
        </dependency>

        <dependency>
            <groupId>mysql</groupId>
            <artifactId>mysql-connector-java</artifactId>
            <version>5.1.34</version>
        </dependency>

    </dependencies>