Spark2.X-自定义累加器AccumulatorV2

本是古典何须时尚 2022-04-10 04:52 131阅读 0赞

*  累加器作为spark的一个共享变量的实现，在用于累加计数计算计算指标的时候可以有效的减少网络的消耗

1.  spark中有一个节点的角色是Master，根据配置文件进行分配，Master节点的职责主要是参与worker节点之间的资源调度。
2.  参与spark作业计算的是worker节点上的excutor，在最开始会将原始RDD复制到excutor的各个task进程上以供计算。这时候如果task过多，或者原始RDD过大，则会耗费更多的时间在资源复制上。
    
    --------------------
    
    累加器可以实现将资源文件复制到每个excutor上，供excutor中的task进程计算使用，减少网络的占用

现在要实现一个累加的效果是，比如有很多个字符串，比如：

ONE:1    TWO:2    THREE:3    ONE:9

想要计算出ONE,TWO,THREE的count和，那么现在可以先自定义一个Accumulator2

import org.apache.spark.util.AccumulatorV2;
    
    import java.util.HashMap;
    import java.util.Map;
    import java.util.Objects;
    
    public class UserDefinedAccumulator extends AccumulatorV2<String, String> {
    	private String one = "ONE";
    	private String two = "TWO";
    	private String three = "THREE";
    	//将想要计算的字符串拼接起来，并赋初始值，后续针对data进行累加，并返回
    	private String data = one + ":0;" + two + ":0;" + three + ":0;";
    	//原始状态
    	private String zero = data;
    
    	//判断是否是初始状态，直接与原始状态的字符串进行对比
    	@Override
    	public boolean isZero() {
    		return data.equals(zero);
    	}
    
    	//复制一个新的累加器
    	@Override
    	public AccumulatorV2<String, String> copy() {
    		return new UserDefinedAccumulator();
    	}
    
    	//重置，恢复原始状态
    	@Override
    	public void reset() {
    		data = zero;
    	}
    
    	//针对传入的字符串，与当前累加器现有的值进行累加
    	@Override
    	public void add(String v) {
    		data = mergeData(v, data, ";");
    	}
    	
    	//将两个累加器的计算结果进行合并
    	@Override
    	public void merge(AccumulatorV2<String, String> other) {
    		data = mergeData(other.value(), data, ";");
    	}
    
    	//将此累加器的计算值返回
    	@Override
    	public String value() {
    		return data;
    	}
    
    	/**
    	 * 合并两个字符串
    	 * @param data_1 字符串1
    	 * @param data_2 字符串2
    	 * @param delimit 分隔符
    	 * @return 结果
    	 */
    	private String mergeData(String data_1, String data_2, String delimit) {
    		StringBuffer res = new StringBuffer();
    		String[] infos_1 = data_1.split(delimit);
    		String[] infos_2 = data_2.split(delimit);
    		Map<String, Integer> map_1 = new HashMap<>();
    		Map<String, Integer> map_2 = new HashMap<>();
    		for (String info : infos_1) {
    			String[] kv = info.split(":");
    			if (kv.length == 2) {
    				String k = kv[0].toUpperCase();
    				Integer v = Integer.valueOf(kv[1]);
    				map_1.put(k, v);
    				continue;
    			}
    		}
    		for (String info : infos_2) {
    			String[] kv = info.split(":");
    			if (kv.length == 2) {
    				String k = kv[0].toUpperCase();
    				Integer v = Integer.valueOf(kv[1]);
    				map_2.put(k, v);
    				continue;
    			}
    		}
    		for (Map.Entry<String, Integer> entry : map_1.entrySet()) {
    			String key = entry.getKey();
    			Integer value = entry.getValue();
    			if (map_2.containsKey(key)) {
    				value = value + map_2.get(key);
    				map_2.remove(key);
    			}
    			res.append(key + ":" + value + delimit);
    		}
    		for (Map.Entry<String, Integer> entry : map_1.entrySet()) {
    			String key = entry.getKey();
    			Integer value = entry.getValue();
    			if (res.toString().contains(key)) {
    				continue;
    			}
    			res.append(key + ":" + value + delimit);
    		}
    		if (!map_2.isEmpty()) {
    			for (Map.Entry<String, Integer> entry : map_2.entrySet()) {
    				String key = entry.getKey();
    				Integer value = entry.getValue();
    				res.append(key + ":" + value + delimit);
    			}
    		}
    		return res.toString().substring(0, res.toString().length() - 1);
    	}
    }

新建一个测试类：

import org.apache.spark.api.java.JavaPairRDD;
    import org.apache.spark.api.java.JavaRDD;
    import org.apache.spark.api.java.JavaSparkContext;
    import org.apache.spark.api.java.function.PairFunction;
    import org.apache.spark.api.java.function.VoidFunction;
    import org.apache.spark.sql.SparkSession;
    import scala.Tuple2;
    
    import java.util.Arrays;
    import java.util.Random;
    
    public class CustomerAccumulator {
    	public static void main(String[] args) {
    		SparkSession sparkSession = SparkSession.builder()
    				.master("local[4]")
    				.appName("CustomerBroadcast")
    				.getOrCreate();
    		JavaSparkContext sc = JavaSparkContext.fromSparkContext(sparkSession.sparkContext());
    		sc.setLogLevel("ERROR");
    		JavaRDD<String> rdd = sc.parallelize(Arrays.asList("ONE", "TWO", "THREE","ONE"));
    		UserDefinedAccumulator count = new UserDefinedAccumulator();
    		//将累加器进行注册
    		sc.sc().register(count, "user_count");
    		//随机设置值
    		JavaPairRDD<String, String> pairRDD = rdd.mapToPair((PairFunction<String, String, String>) s -> {
    			int num = new Random().nextInt(10);
    			return new Tuple2<>(s, s + ":" + num);
    		});
    		//foreach中进行累加
    		pairRDD.foreach((VoidFunction<Tuple2<String, String>>) tuple2 -> {
    			System.out.println(tuple2._2);
    			count.add(tuple2._2);
    		});
    		System.out.println("the value of accumulator is:"+count.value());
    	}
    }

查看运行结果：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2hscDQyMDc_size_16_color_FFFFFF_t_70][]

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2hscDQyMDc_size_16_color_FFFFFF_t_70]: /images/20220402/94b5815d2bbb4190a4202427f171e9b4.png