Spark Streaming整合Kafka指南(超详细)

kafka是做消息的缓存，数据和业务隔离操作的消息队列，而sparkstreaming是一款准实时流式计算框架，所以二者的整合，是大势所趋。二者的整合，有主要的两大版本。

在spark-stremaing-kafka-0-8的版本中又分为了两种方式：receiver的方式和direct的方式来读取kafka中的数据，主要区别就是是否依赖zookeeper来管理offset信息，以及是否拥有receiver。

API查询地址：http://spark.apache.org/docs/2.2.2/streaming-kafka-0-8-integration.html

导入下面要用到的Maven依赖

<properties>
	<spark.version>2.2.2</spark.version>
</properties>
<dependencies>
	<dependency>
		<groupId>org.apache.spark</groupId>
		<artifactId>spark-core_2.11</artifactId>
		<version>${spark.version}</version>
	</dependency>
	<dependency>
		<groupId>org.apache.spark</groupId>
		<artifactId>spark-streaming_2.11</artifactId>
		<version>${spark.version}</version>
	</dependency>
	<dependency>
		<groupId>org.apache.spark</groupId>
		<artifactId>spark-streaming-kafka-0-8_2.11</artifactId>
		<version>${spark.version}</version>
	</dependency>
	<dependency>
		<groupId>org.apache.hbase</groupId>
		<artifactId>hbase-server</artifactId>
		<version>1.1.5</version>
	</dependency>
	<dependency>
		<groupId>org.scalikejdbc</groupId>
		<artifactId>scalikejdbc_2.11</artifactId>
		<version>3.2.0</version>
	</dependency>
    <dependency>
         <groupId>mysql</groupId>
         <artifactId>mysql-connector-java</artifactId>
         <version>5.1.39</version>
    </dependency>    
	<!--        <dependency><groupId>org.apache.spark</groupId><artifactId>spark-sql_2.11</artifactId><version>${spark.version}</version></dependency>-->
</dependencies>

一、Receiver方式

1.kafka基于receiver方式一

这种方式使用Receiver来获取数据。Receiver是使用Kafka的高层次Consumer API来实现的。receiver从Kafka中获取的数据都是存储在Spark Executor的内存中的，然后Spark Streaming启动的job会去处理那些数据。

然而，在默认的配置下，这种方式可能会因为底层的失败而丢失数据。如果要启用高可靠机制，让数据零丢失，就必须启用Spark Streaming的预写日志机制（Write Ahead Log，WAL）。该机制会同步地将接收到的Kafka数据写入分布式文件系统（比如HDFS）上的预写日志中。所以，即使底层节点出现了失败，也可以使用预写日志中的数据进行恢复。

• 需要注意的地方

  • Kafka的topic分区和Spark Streaming中生成的RDD分区没有关系。 在KafkaUtils.createStream中增加分区数量只会增加单个receiver的线程数，不会增加Spark的并行度
  • 可以创建多个的Kafka的输入DStream， 使用不同的group和topic， 使用多个receiver并行接收数据。
  • 如果启用了HDFS等有容错的存储系统，并且启用了写入日志，则接收到的数据已经被复制到日志中。因此，输入流的存储级别设置StorageLevel.MEMORY_AND_DISK_SER（即使用KafkaUtils.createStream（…，StorageLevel.MEMORY_AND_DISK_SER））的存储级别。

  数据会丢失原因

具体操作：

首先创建一个topic

kafka-topics.sh --create \
--topic hadoop \
--zookeeper hadoop01:2181,hadoop02:2181,hadoop03:2181/kafka \
--partitions 3 \
--replication-factor 3

启动消费者

kafka-console-producer.sh --topic hadoop --broker-list hadoop01:9092,hadoop02:9092,hadoop03:9092

Direct.scala

package blog.kafka

import kafka.serializer.StringDecoder
import org.apache.spark.SparkConf
import org.apache.spark.streaming.dstream.InputDStream
import org.apache.spark.streaming.kafka.KafkaUtils
import org.apache.spark.streaming.{Seconds, StreamingContext}

/**
  * @Author Daniel
  * @Description SparkStreaming基于Direct整合Kafka方式一
  *
  **/
object Direct {
  def main(args: Array[String]): Unit = {
    if (args == null || args.length < 3) {
      println(
        """
          |Usage: <broker.list> <groupId> <topicStr>
        """.stripMargin)
      System.exit(-1)
    }
    val Array(brokerList, groupId, topicStr) = args
    //direct中的参数为一个set集合
    val topics = topicStr.split(",").toSet
    val conf = new SparkConf()
      .setAppName("Direct")
      //没有Receiver这里给一个就行
      .setMaster("local")
      .set("spark.streaming.receiver.writeAheadLog.enable", "true")

    val batchInterval = Seconds(2)
    val kafkaParams: Map[String, String] = Map[String, String](
      "metadata.broker.list" -> brokerList,
      "group.id" -> groupId,
      "auto.offset.reset" -> "smallest"
    )
    val ssc = new StreamingContext(conf, batchInterval)
    val input: InputDStream[(String, String)] = KafkaUtils
      .createDirectStream[String, String, StringDecoder, StringDecoder](ssc,
      kafkaParams, topics)

    val ret = input
      .flatMap(_._2.split("\\s+"))
      .map((_, 1))
      .reduceByKey(_ + _)

    ret.foreachRDD((rdd, time) => {
      if (!rdd.isEmpty()) {
        println(s"Time: $time")
        rdd.foreach(println)
      }
    })


    ssc.start()
    ssc.awaitTermination()
  }
}

输入参数

hadoop01:2181,hadoop02:2181,hadoop03:2181/kafka bdedev-group-1 hadoop

运行代码

输入信息

结果

2.kafka基于receiver方式二——使用checkpoint

DirectCheckPoint.scala

package blog.kafka

import kafka.serializer.StringDecoder
import org.apache.spark.SparkConf
import org.apache.spark.streaming.dstream.InputDStream
import org.apache.spark.streaming.kafka.KafkaUtils
import org.apache.spark.streaming.{Seconds, StreamingContext}

/**
  * @Author Daniel
  * @Description SparkStreaming基于Direct整合Kafka方式二——使用checkpoint
  *
  **/
object DirectCheckPoint {
  def main(args: Array[String]): Unit = {

    if (args == null || args.length < 3) {
      println(
        """
          |Usage: <broker.list> <groupId> <topicStr>
        """.stripMargin)
      System.exit(-1)
    }
    val Array(brokerList, groupId, topicStr) = args
    val topics = topicStr.split(",").toSet
    val conf = new SparkConf()
      .setAppName("DirectCheckPoint")
      .setMaster("local")

    val batchInterval = Seconds(2)
    val kafkaParams: Map[String, String] = Map[String, String](
      "metadata.broker.list" -> brokerList,
      "group.id" -> groupId,
      "auto.offset.reset" -> "smallest"
    )
    val checkpoint = "file:///F:/ssdata/checkpoint/ck2"

    def creatingFunc(): StreamingContext = {
      val ssc = new StreamingContext(conf, batchInterval)
      //使用checkpoint来存储offset信息 保证数据被依次消费
      ssc.checkpoint(checkpoint)
      val input: InputDStream[(String, String)] = KafkaUtils
        .createDirectStream[String, String, StringDecoder, StringDecoder](ssc,
        kafkaParams, topics)

      input.foreachRDD((rdd, time) => {
        if (!rdd.isEmpty()) {
          println(s"Time: $time")
          println("------------------------rdd's count: " + rdd.count())
        }
      })
      ssc
    }

    val ssc = StreamingContext.getOrCreate(checkpoint, creatingFunc)

    ssc.start()
    ssc.awaitTermination()
  }
}

输入参数

hadoop01:2181,hadoop02:2181,hadoop03:2181/kafka bde-dev-group-2 hadoop

输出

二、Direct方式

Direct方式的特点：

• 简化的并行性：不需要创建多个输入Kafka流并将其合并。 使用directStream，Spark Streaming将创建与使用Kafka分区一样多的RDD分区，这些分区将全部从Kafka并行读取数据。 所以在Kafka和RDD分区之间有一对一的映射关系。
• 效率：在第一种方法中实现零数据丢失需要将数据存储在预写日志中，这会进一步复制数据。这实际
  上是效率低下的，因为数据被有效地复制了两次:一次是Kafka，另一次是由预先写入日志（Write
  Ahead Log）复制。这个第二种方法消除了这个问题，因为没有接收器，因此不需要预先写入日志。
  只要Kafka数据保留时间足够长。
• 正好一次（Exactly-once）的语义：第一种方法使用Kafka的高级API来在Zookeeper中存储消耗的偏移量。传统上这是从Kafka消费数据的方式。虽然这种方法（结合提前写入日志）可以确保零数据丢失（即至少一次语义），但是在某些失败情况下，有一些记录可能会消费两次。发生这种情况是因为Spark Streaming可靠接收到的数据与Zookeeper跟踪的偏移之间的不一致。因此，在第二种方法中，我们使用不使用Zookeeper的简单Kafka API。在其检查点内，Spark Streaming跟踪偏移量。这消除了Spark Streaming和Zookeeper/Kafka之间的不一致，因此Spark Streaming每次记录都会在发生故障的情况下有效地收到一次。为了实现输出结果的一次语义，将数据保存到外部数据存储区的输出操作必须是幂等的，或者是保存结果和偏移量的原子事务。

1.kafka基于direct方式一

输入参数

hadoop01:9092,hadoop02:9092,hadoop03:9092 bde-dev-group-3 hadoop

Direct.scala

package blog.kafka

import kafka.serializer.StringDecoder
import org.apache.spark.SparkConf
import org.apache.spark.streaming.dstream.InputDStream
import org.apache.spark.streaming.kafka.KafkaUtils
import org.apache.spark.streaming.{Seconds, StreamingContext}

/**
  * @Author Daniel
  * @Description SparkStreaming基于Direct整合Kafka方式一
  *
  **/
object Direct {
  def main(args: Array[String]): Unit = {
    if (args == null || args.length < 3) {
      println(
        """
          |Usage: <broker.list> <groupId> <topicStr>
        """.stripMargin)
      System.exit(-1)
    }
    val Array(brokerList, groupId, topicStr) = args
    //direct中的参数为一个set集合
    val topics = topicStr.split(",").toSet
    val conf = new SparkConf()
      .setAppName("Direct")
      //没有Receiver这里给一个就行
      .setMaster("local")
      .set("spark.streaming.receiver.writeAheadLog.enable", "true")

    val batchInterval = Seconds(2)
    val kafkaParams: Map[String, String] = Map[String, String](
      "metadata.broker.list" -> brokerList,
      "group.id" -> groupId,
      "auto.offset.reset" -> "smallest"
    )
    val ssc = new StreamingContext(conf, batchInterval)
    val input: InputDStream[(String, String)] = KafkaUtils
      .createDirectStream[String, String, StringDecoder, StringDecoder](ssc,
      kafkaParams, topics)

    val ret = input
      .flatMap(_._2.split("\\s+"))
      .map((_, 1))
      .reduceByKey(_ + _)

    ret.foreachRDD((rdd, time) => {
      if (!rdd.isEmpty()) {
        println(s"Time: $time")
        rdd.foreach(println)
      }
    })


    ssc.start()
    ssc.awaitTermination()
  }
}

输出

2.kafka基于direct方式二——使用checkpoint

输入参数

hadoop01:9092,hadoop02:9092,hadoop03:9092 bde-dev-group-4 hadoop

DirectCheckPoint.scala

package blog.kafka

import kafka.serializer.StringDecoder
import org.apache.spark.SparkConf
import org.apache.spark.streaming.dstream.InputDStream
import org.apache.spark.streaming.kafka.KafkaUtils
import org.apache.spark.streaming.{Seconds, StreamingContext}

/**
  * @Author Daniel
  * @Description SparkStreaming基于Direct整合Kafka方式二——使用checkpoint
  *
  **/
object DirectCheckPoint {
  def main(args: Array[String]): Unit = {

    if (args == null || args.length < 3) {
      println(
        """
          |Usage: <broker.list> <groupId> <topicStr>
        """.stripMargin)
      System.exit(-1)
    }
    val Array(brokerList, groupId, topicStr) = args
    val topics = topicStr.split(",").toSet
    val conf = new SparkConf()
      .setAppName("DirectCheckPoint")
      .setMaster("local")

    val batchInterval = Seconds(2)
    val kafkaParams: Map[String, String] = Map[String, String](
      "metadata.broker.list" -> brokerList,
      "group.id" -> groupId,
      "auto.offset.reset" -> "smallest"
    )
    val checkpoint = "file:///F:/ssdata/checkpoint/ck2"

    def creatingFunc(): StreamingContext = {
      val ssc = new StreamingContext(conf, batchInterval)
      //使用checkpoint来存储offset信息 保证数据被依次消费
      ssc.checkpoint(checkpoint)
      val input: InputDStream[(String, String)] = KafkaUtils
        .createDirectStream[String, String, StringDecoder, StringDecoder](ssc,
        kafkaParams, topics)

      input.foreachRDD((rdd, time) => {
        if (!rdd.isEmpty()) {
          println(s"Time: $time")
          println("------------------------rdd's count: " + rdd.count())
        }
      })
      ssc
    }

    val ssc = StreamingContext.getOrCreate(checkpoint, creatingFunc)

    ssc.start()
    ssc.awaitTermination()
  }
}

输出

3.kafka基于direct方式三——使用HBase管理offset

确保集群HBase服务开启，拷贝hdfs-site.xml、core-site.xml、hbase-site.xml至当前目录

HBaseConnectionPool.java

package blog.util;

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.hbase.Cell;
import org.apache.hadoop.hbase.HBaseConfiguration;
import org.apache.hadoop.hbase.TableName;
import org.apache.hadoop.hbase.client.*;
import org.apache.hadoop.hbase.filter.BinaryComparator;
import org.apache.hadoop.hbase.filter.CompareFilter;
import org.apache.hadoop.hbase.filter.Filter;
import org.apache.hadoop.hbase.filter.RowFilter;

import java.io.IOException;
import java.util.HashMap;
import java.util.LinkedList;
import java.util.List;
import java.util.Map;

/**
 * @Author Daniel
 * @Description HBase Utils
 **/

public class HBaseConnectionPool {
    public static void main(String[] args) {
        System.out.println(getConnection());
    }

    //定义一个连接池
    private static LinkedList<Connection> pool = new LinkedList<Connection>();

    //初始化
    static {
        try {

            Configuration conf = HBaseConfiguration.create();
            conf.set("hbase.zookeeper.quorum", "hadoop01:2181,hadoop02:2181,hadoop03:2181");
            //尝试五次
            for (int i = 0; i < 5; i++) {
                //将对象添加到连接池
                pool.push(ConnectionFactory.createConnection(conf));
            }
        } catch (IOException e) {
            e.printStackTrace();
        }
    }

    public static Connection getConnection() {
        while (pool.isEmpty()) {
            try {
                System.out.println("Connection pool为空！！！");
                Thread.sleep(1000);
            } catch (InterruptedException e) {
                e.printStackTrace();
            }
        }
        return pool.poll();
    }

    //回收Connection对象
    public static void release(Connection connection) {
        pool.push(connection);
    }

    //获取字段的值
    public static Map<Integer, Long> getColValue(Connection connection, TableName tableName, byte[] rk, byte[] cf) {
        //定义一个Map存储分区和偏移量
        Map<Integer, Long> partition2Offset = new HashMap<>();
        try {
            Table table = connection.getTable(tableName);
            //获取一个扫描器
            Scan scan = new Scan();
            //创建一个过滤器
            Filter filter = new RowFilter(CompareFilter.CompareOp.EQUAL, new BinaryComparator(rk));
            //设置过滤规则
            scan.setFilter(filter);
            ResultScanner results = table.getScanner(scan);
            //遍历结果
            for (Result result : results) {
                List<Cell> cells = result.listCells();
                for (Cell cell : cells) {
                    //col
                    byte[] qArr = cell.getQualifierArray();
                    int qLen = cell.getQualifierLength();
                    int qOffset = cell.getQualifierOffset();
                    //value
                    byte[] vArr = cell.getValueArray();
                    int vLen = cell.getValueLength();
                    int vOffset = cell.getValueOffset();
                    int partition = Integer.valueOf(new String(qArr, qOffset, qLen));
                    long value = Long.valueOf(new String(vArr, vOffset, vLen));

                    partition2Offset.put(partition, value);
                }
            }
            table.close();
        } catch (IOException e) {
            e.printStackTrace();
        }
        return partition2Offset;
    }

    //保存偏移量到HBase
    public static void set(Connection connection, TableName tableName, byte[] rk, byte[] cf, byte[] col, byte[] value) {
        try {
            Table table = connection.getTable(tableName);
            //将行键封装到Put对象
            Put put = new Put(rk);
            //添加字段
            put.addColumn(cf, col, value);
            table.put(put);
            table.close();
        } catch (IOException e) {
            e.printStackTrace();
        }
    }
}

运行，测试连接，看到如下输入则连接成功

流程：

• 手动从hbase中读取上一次消费的offset信息

  • 有：从指定的offset位置开始消费
  • 无：从offset为0或者最小的位置开始消费

• 使用指定offset从kafka中拉取数据

• 拉取到数据之后进行业务处理

• 指定HBase进行offset的更新

• Table hadoop-topic-offset

字段	注释
topic-group	行键（Rowkey）
cf	列族（Columns）
partition	分区
offset	偏移量
connection	连接

建表语句

create 'hadoop-topic-offset', 'cf'

DirectWithHBase.scala

package blog.kafka

import blog.util.HBaseConnectionPool
import kafka.common.TopicAndPartition
import kafka.message.MessageAndMetadata
import kafka.serializer.StringDecoder
import org.apache.hadoop.hbase.TableName
import org.apache.spark.SparkConf
import org.apache.spark.streaming.dstream.InputDStream
import org.apache.spark.streaming.kafka.{HasOffsetRanges, KafkaUtils, OffsetRange}
import org.apache.spark.streaming.{Seconds, StreamingContext}

import scala.collection.JavaConversions._
import scala.collection.mutable

/**
  * @Author Daniel
  * @Description SparkStreaming基于Receiver整合Kafka方式三——使用HBase管理offset
  **/


object DirectWithHBase {
  def main(args: Array[String]): Unit = {
    if (args == null || args.length < 3) {
      println(
        """
          |Usage: <broker.list> <groupId> <topicStr>
        """.stripMargin)
      System.exit(-1)
    }
    val Array(brokerList, groupId, topicStr) = args
    val topics = topicStr.split(",").toSet
    val conf = new SparkConf()
      .setAppName("DirectWithHBase")
      .setMaster("local")
    val batchInterval = Seconds(2)
    val kafkaParams = Map[String, String](
      "metadata.broker.list" -> brokerList,
      "group.id" -> groupId,
      "auto.offset.reset" -> "smallest"
    )

    val ssc = new StreamingContext(conf, batchInterval)
    val input = createMsg(ssc, kafkaParams, topics)

    input.foreachRDD((rdd, time) => {
      if (!rdd.isEmpty()) {
        println(s"Time: $time")
        println("------------------------rdd's count: " + rdd.count())
        //更新偏移量
        storeOffsets(rdd.asInstanceOf[HasOffsetRanges].offsetRanges, kafkaParams("group.id"))
      }
    })

    ssc.start()
    ssc.awaitTermination()
  }


  //保存偏移量
  def storeOffsets(offsetRanges: Array[OffsetRange], group: String): Unit = {
    val connection = HBaseConnectionPool.getConnection
    val tableName = TableName.valueOf("hadoop-topic-offset")
    val cf = "cf".getBytes()
    for (offsetRange <- offsetRanges) {
      val rk = s"${offsetRange.topic}-${group}".getBytes()
      val partition = offsetRange.partition
      //偏移量
      val offset = offsetRange.untilOffset
      //将结果保存到Hbase中
      HBaseConnectionPool.set(connection, tableName, rk, cf, (partition + "").getBytes(), (offset + "").getBytes())
    }
    HBaseConnectionPool.release(connection)
  }

  def createMsg(ssc: StreamingContext, kafkaParams: Map[String, String], topics: Set[String]): InputDStream[(String, String)] = {
    //从zookeeper中读取偏移量
    val offsets = getOffsets(topics, kafkaParams("group.id"))
    var messages: InputDStream[(String, String)] = null
    if (offsets.isEmpty) { //如果为空就从0开始读取
      messages = KafkaUtils
        .createDirectStream[String, String, StringDecoder, StringDecoder](ssc,
        kafkaParams, topics)
    } else { //有值就从指定的offset位置开始读取
      val messageHandler = (msg: MessageAndMetadata[String, String]) => (msg.key(), msg.message())
      messages = KafkaUtils
        .createDirectStream[String, String, StringDecoder, StringDecoder, (String, String)](ssc,
        kafkaParams, offsets,
        messageHandler)
    }
    messages
  }

  /**
    * create 'hadoop-topic-offset', 'cf'
    * rowkey
    * topic-group
    *
    * @return offsets
    */
  def getOffsets(topics: Set[String], group: String): Map[TopicAndPartition, Long] = {
    val offsets = mutable.Map[TopicAndPartition, Long]()
    //获取连接
    val connection = HBaseConnectionPool.getConnection
    //拿到HBase中的表名
    val tableName = TableName.valueOf("hadoop-topic-offset")
    //数据库中的列族名
    val cf = "cf".getBytes()
    for (topic <- topics) {
      //行键
      val rk = s"${topic}-${group}".getBytes()
      //获取分区与偏移量信息
      val partition2Offsets = HBaseConnectionPool.getColValue(connection, tableName, rk, cf)
      partition2Offsets.foreach { case (partition, offset) => {
        offsets.put(TopicAndPartition(topic, partition), offset)
      }
      }
    }
    HBaseConnectionPool.release(connection)
    offsets.toMap
  }
}

启动，然后在生产者中发送消息！！

结果如下

三、Spark Streaming与Kafka整合的常见问题

这里把前面整合HBase中用到的一些方法提取出来

KafkaManager.scala

package blog.kafka

import kafka.common.TopicAndPartition
import kafka.message.MessageAndMetadata
import kafka.serializer.StringDecoder
import org.apache.curator.framework.CuratorFramework
import org.apache.spark.streaming.StreamingContext
import org.apache.spark.streaming.dstream.InputDStream
import org.apache.spark.streaming.kafka.{KafkaUtils, OffsetRange}

import scala.collection.JavaConversions._
import scala.collection.mutable

/**
  * @Author Daniel
  * @Description Utils
  *              将前面整合HBase的一些方法抽取出来
  **/

object KafkaManager {
  //保存偏移量
  def storeOffsets(offsetRanges: Array[OffsetRange], group: String, curator: CuratorFramework): Unit = {
    for (offsetRange <- offsetRanges) {
      val topic = offsetRange.topic
      val partition = offsetRange.partition
      val offset = offsetRange.untilOffset
      val path = s"${topic}/${group}/${partition}"
      checkExists(path, curator)
      curator.setData().forPath(path, new String(offset + "").getBytes())
    }
  }

  def createMsg(ssc: StreamingContext, kafkaParams: Map[String, String], topics: Set[String], curator: CuratorFramework): InputDStream[(String, String)] = {
    //从zookeeper中读取offset
    val offsets: Map[TopicAndPartition, Long] = getOffsets(topics, kafkaParams("group.id"), curator)
    var messages: InputDStream[(String, String)] = null
    if (offsets.isEmpty) { //为空则从0开始读取
      messages = KafkaUtils
        .createDirectStream[String, String, StringDecoder, StringDecoder](ssc,
        kafkaParams, topics)
    } else { //有就从指定位置开始读取
      val messageHandler = (msgH: MessageAndMetadata[String, String]) => (msgH.key(), msgH.message())
      messages = KafkaUtils
        .createDirectStream[String, String, StringDecoder, StringDecoder, (String, String)](ssc,
        kafkaParams, offsets,
        messageHandler)
    }
    messages
  }

  def getOffsets(topics: Set[String], group: String, curator: CuratorFramework): Map[TopicAndPartition, Long] = {
    val offsets = mutable.Map[TopicAndPartition, Long]()
    for (topic <- topics) {
      val parent = s"${topic}/${group}"
      checkExists(parent, curator)
      //此时目录一定存在
      for (partition <- curator.getChildren.forPath(parent)) {
        val path = s"${parent}/${partition}"
        val offset = new String(curator.getData.forPath(path)).toLong
        offsets.put(TopicAndPartition(topic, partition.toInt), offset)
      }
    }
    offsets.toMap
  }

  //检测是否存在
  def checkExists(path: String, curator: CuratorFramework): Unit = {
    if (curator.checkExists().forPath(path) == null) {
      //不存在则创建
      curator.create().creatingParentsIfNeeded()
        .forPath(path)
    }
  }
}

1.输出一致性语义的问题

以上是官网上对一致性语义的说明，大意是为了使获得的结果保持输出一致性语义，你用来保存结果和偏移量到外部数据存储的操作必须是幂等或者是原子事务。

• 幂等操作

  • 创建测试的mysql数据库

    create database db1;
    

  • 建表

    create table myorders(name varchar(20), orderid varchar(100) primary key);
    

  • 新建topic：mytopic

    kafka-topics.sh --zookeeper hadoop01:2181/kafka --create --topic mytopic --partitions 3 --replication-factor 1
    

  • 运行程序之后，向mytopic发送数据，数据格式为 “字符,数字” 比如 abc,3

    kafka-console-producer.sh --topic mytopic --broker-list hadoop01:9092,hadoop02:9092,hadoop03:9092
    

  KafkaOffsetIdempotent.scala

  package blog.kafka
  
  import java.sql.DriverManager
  
  import org.apache.curator.framework.CuratorFrameworkFactory
  import org.apache.curator.retry.ExponentialBackoffRetry
  import org.apache.spark.SparkConf
  import org.apache.spark.streaming.kafka.HasOffsetRanges
  import org.apache.spark.streaming.{Seconds, StreamingContext}
  
  /**
    * @Author Daniel
    * @Description 幂等处理
    **/
  
  object KafkaOffsetIdempotent {
    def main(args: Array[String]): Unit = {
      val sparkConf = new SparkConf()
        .setAppName("KafkaOffsetIdempotent")
        .setMaster("local[2]")
  
      val processingInterval = 2
      val brokers = "hadoop01:9092,hadoop02:9092,hadoop03:9092"
      val topic = "mytopic"
      // Create direct kafka stream with brokers and topics
      val topicsSet = topic.split(",").toSet
      val kafkaParams = Map[String, String](
        "metadata.broker.list" -> brokers,
        "auto.offset.reset" -> "smallest",
        "group.id" -> "myspark"
      )
  
      val ssc = new StreamingContext(sparkConf, Seconds(processingInterval))
      val messages = KafkaManager.createMsg(ssc, kafkaParams, topicsSet, client)
  
      val jdbcUrl = "jdbc:mysql://localhost:3306/db1"
      val jdbcUser = "root"
      val jdbcPassword = "root"
  
      messages.foreachRDD(rdd => {
        if (!rdd.isEmpty()) {
          rdd.map(x => x._2).foreachPartition(partition => {
            //拿到connection
            val dbConn = DriverManager.getConnection(jdbcUrl, jdbcUser, jdbcPassword)
            partition.foreach(msg => {
              //按照格式切割
              val name = msg.split(",")(0)
              val orderid = msg.split(",")(1)
              //幂等操作：如果主键相同，则覆盖这个结果
              val sql = s"insert into myorders(name, orderid) values ('$name', '$orderid') ON DUPLICATE KEY UPDATE name='${name}'"
              val pstmt = dbConn.prepareStatement(sql)
              //执行SQL
              pstmt.execute()
            })
            dbConn.close()
          })
          //将数据保存到偏移量
          KafkaManager.storeOffsets(rdd.asInstanceOf[HasOffsetRanges].offsetRanges, kafkaParams("group.id"), client)
        }
      })
  
      ssc.start()
      ssc.awaitTermination()
    }
  
    //构建一个Curator的Client
    val client = {
      val client = CuratorFrameworkFactory.builder()
        .connectString("hadoop01:2181,hadoop02:2181,hadoop03:2181")
        .retryPolicy(new ExponentialBackoffRetry(1000, 3))
        .namespace("kafka/consumers/offsets")
        .build()
      client.start()
      client
    }
  }
  

  测试结果如下：

• 原子性操作

  基于scala-jdbc的方式操作数据库

  • 建表

    create table mytopic(topic varchar(200), partid int, offset bigint);
    create table mydata(name varchar(200), id int);
    

  • 插入数据

    insert into mytopic(topic, partid, offset) values('mytopic',0,0);
    insert into mytopic(topic, partid, offset) values('mytopic',1,0);
    insert into mytopic(topic, partid, offset) values('mytopic',2,0);
    

  • 往mytopic发送数据， 数据格式为 “字符,数字” 比如 abc,3

  KafkaOffsetTransaction.scala

  package blog.kafka
  
  import kafka.common.TopicAndPartition
  import kafka.message.MessageAndMetadata
  import kafka.serializer.StringDecoder
  import org.apache.spark.streaming.kafka.{HasOffsetRanges, KafkaUtils}
  import org.apache.spark.streaming.{Seconds, StreamingContext}
  import org.apache.spark.{SparkConf, TaskContext}
  import scalikejdbc.{ConnectionPool, DB, _}
  
  /**
    * @Author Daniel
    * @Description 原子处理
    *              将偏移量存储在事务里面
    **/
  
  object KafkaOffsetTransaction {
    def main(args: Array[String]): Unit = {
      val sparkConf = new SparkConf().setAppName("test").setMaster("local[2]")
  
      val processingInterval = 2
      val brokers = "hadoop01:9092,hadoop02:9092,hadoop03:9092"
      val topic = "mytopic"
      // Create direct kafka stream with brokers and topics
      topic.split(",").toSet
      val kafkaParams = Map[String, String]("metadata.broker.list" -> brokers,
        "auto.offset.reset" -> "smallest"
      )
  
      val ssc = new StreamingContext(sparkConf, Seconds(processingInterval))
  
      //相关配置信息
      val driver = "com.mysql.jdbc.Driver"
      val jdbcUrl = "jdbc:mysql://localhost:3306/db1"
      val jdbcUser = "root"
      val jdbcPassword = "root"
  
      // 设置jdbc
      Class.forName(driver)
      // 设置连接池
      ConnectionPool.singleton(jdbcUrl, jdbcUser, jdbcPassword)
      //隐式转换参数
      val fromOffsets = DB.readOnly { implicit session =>
        sql"select topic, partid, offset from mytopic".
          //r就是ResultSet结果集
          map { r =>
          //topic partition offset
          TopicAndPartition(r.string(1), r.int(2)) -> r.long(3)
        }.list
          .apply()
          .toMap //转换结果类型
      }
  
      val messageHandler = (mmd: MessageAndMetadata[String, String]) => (mmd.topic, mmd.message())
      val messages = KafkaUtils.createDirectStream[String, String, StringDecoder, StringDecoder, (String, String)](ssc, kafkaParams, fromOffsets, messageHandler)
  
      messages.foreachRDD((rdd, time) => {
        if (!rdd.isEmpty()) {
          println(s"Time: $time")
          rdd.foreachPartition(partition => {
            //偏移量
            val offsetRanges = rdd.asInstanceOf[HasOffsetRanges].offsetRanges
            //某个分区所对应的偏移量
            val pOffsetRange = offsetRanges(TaskContext.get.partitionId)
  
            // scala中使用localTx 开启事务操作
            DB.localTx { implicit session =>
              //数据
              partition.foreach(msg => {
                // 或者使用scalike的batch 插入
                val name = msg._2.split(",")(0)
                val id = msg._2.split(",")(1)
                sql"""insert into  mydata(name,id) values (${name},${id})""".execute().apply()
              })
              //            val i = 1 / 0//测试，若发生错误，则更新失败，进行事务的回滚操作
              //偏移量
              sql"""update mytopic set offset = ${pOffsetRange.untilOffset} where topic =
                    ${pOffsetRange.topic} and partid = ${pOffsetRange.partition}""".update.apply()
            }
          })
        }
      })
      ssc.start()
      ssc.awaitTermination()
    }
  }
  

  测试结果如下：

初始的offset信息：

增加数据后可以看到，每个offset被增加了一

2.限流的处理

sparkstreaming要从kafka拉取数据，并进行处理；下一次再循环，如果批次的间隔时间为2s，但是数据的处理时间为3s，所以会有越来越多的没有被处理的数据进行累积，最后会拖垮程序，这不是我们所期望的。

解决思路，只能限制流量。非常简单，通过一个参数搞定：spark.streaming.kafka.maxRatePerPartition

spark.streaming.kafka.maxRatePerPartition: spark程序每秒中从每个partition分区读取的最大的数据条数。比如batchInterval为2s，topic的分区为3，该参数的值为100，请问，每个批次最多可以读取多少条数据？2×3×100=600条。

只用在Direct.scala的基础上改一个参数即可

输入参数

hadoop01:9092,hadoop02:9092,hadoop03:9092 bde-dev-group-5 hadoop

DirectLimit.scala

package blog.kafka

import kafka.serializer.StringDecoder
import org.apache.spark.SparkConf
import org.apache.spark.streaming.dstream.InputDStream
import org.apache.spark.streaming.kafka.KafkaUtils
import org.apache.spark.streaming.{Seconds, StreamingContext}

/**
  * @Author Daniel
  * @Description 限流处理
  *
  **/
object DirectLimit {
  def main(args: Array[String]): Unit = {
    if (args == null || args.length < 3) {
      println(
        """
          |Usage: <broker.list> <groupId> <topicStr>
        """.stripMargin)
      System.exit(-1)
    }
    val Array(brokerList, groupId, topicStr) = args
    //direct中的参数为一个set集合
    val topics = topicStr.split(",").toSet
    val conf = new SparkConf()
      .setAppName("DirectLimit")
      //没有Receiver这里给一个就行
      .setMaster("local")
      //设置spark程序每秒中从每个partition分区读取的最大的数据条数
      .set("spark.streaming.kafka.maxRatePerPartition", "100")

    val batchInterval = Seconds(2)
    val kafkaParams: Map[String, String] = Map[String, String](
      "metadata.broker.list" -> brokerList,
      "group.id" -> groupId,
      "auto.offset.reset" -> "smallest"
    )
    val ssc = new StreamingContext(conf, batchInterval)
    val input: InputDStream[(String, String)] = KafkaUtils
      .createDirectStream[String, String, StringDecoder, StringDecoder](ssc,
      kafkaParams, topics)

    val ret = input
      .flatMap(_._2.split("\\s+"))
      .map((_, 1))
      .reduceByKey(_ + _)

    ret.foreachRDD((rdd, time) => {
      if (!rdd.isEmpty()) {
        println(s"Time: $time")
        rdd.foreach(println)
      }
    })


    ssc.start()
    ssc.awaitTermination()
  }
}

可以看到结果(这里的参数大小适合有数据重复次数超过六百次的)：