Spark Streaming No Receivers 方式的createDirectStream 方法不使用接收器，而是创建输入流直接从Kafka 集群节点拉取消息。输入流保证每个消息从Kafka 集群拉取以后只完全转换一次，保证语义一致性。但是当作业发生故障或重启时，要保障从当前的消费位点去处理数据(即Exactly Once语义)，单纯的依靠SparkStreaming本身的机制是不太理想的，生产环境中通常借助手动管理offset的方式来维护kafka的消费位点。本文分享将介绍如何手动管理Kafka的Offset，希望对你有所帮助。本文主要包括以下内容：

• 如何使用MySQL管理Kafka的Offset
• 如何使用Redis管理Kafka的OffSet

如何使用MySQL管理Kafka的Offset

我们可以从Spark Streaming 应用程序中编写代码来手动管理Kafka偏移量，偏移量可以从每一批流处理中生成的RDDS偏移量来获取，获取方式为：

KafkaUtils.createDirectStream(...).foreachRDD { rdd =>
// 获取偏移量
val offsetRanges = rdd.asInstanceOf[HasOffsetRanges].offsetRanges
 ...
  }

当获取到偏移量之后，可以将将其保存到外部存储设备中(MySQL、Redis、Zookeeper、HBase等)。

使用案例代码

• MySQL中用于保存偏移量的表

CREATE TABLE `topic_par_group_offset` (
  `topic` varchar(255) NOT NULL,
  `partition` int(11) NOT NULL,
  `groupid` varchar(255) NOT NULL,
  `offset` bigint(20) DEFAULT NULL,
  PRIMARY KEY (`topic`,`partition`,`groupid`)
) ENGINE=InnoDB DEFAULT CHARSET=utf8 ;

• 常量配置类:ConfigConstants

object ConfigConstants {
  // Kafka配置
  val kafkaBrokers = "kms-2:9092,kms-3:9092,kms-4:9092"
  val groupId = "group_test"
  val kafkaTopics = "test"
  val batchInterval = Seconds(5)
  val streamingStorageLevel = StorageLevel.MEMORY_AND_DISK_SER_2
  val kafkaKeySer = "org.apache.kafka.common.serialization.StringSerializer"
  val kafkaValueSer = "org.apache.kafka.common.serialization.StringSerializer"
  val sparkSerializer = "org.apache.spark.serializer.KryoSerializer"
  val batchSize = 16384
  val lingerMs = 1
  val bufferMemory = 33554432
  // MySQL配置
  val user = "root"
  val password = "123qwe"
  val url = "jdbc:mysql://localhost:3306/kafka_offset"
  val driver = "com.mysql.jdbc.Driver"
  // 检查点配置
  val checkpointDir = "file:///e:/checkpoint"
  val checkpointInterval = Seconds(10)
  // Redis配置
  val redisAddress = "192.168.10.203"
  val redisPort = 6379
  val redisAuth = "123qwe"
  val redisTimeout = 3000
}

• JDBC连接工具类:JDBCConnPool

object JDBCConnPool {

  val log: Logger = Logger.getLogger(JDBCConnPool.getClass)
  var dataSource: BasicDataSource = null
  /**
    * 创建数据源
    *
    * @return
    */
  def getDataSource(): BasicDataSource = {
    if (dataSource == null) {
      dataSource = new BasicDataSource()
      dataSource.setDriverClassName(ConfigConstants.driver)
      dataSource.setUrl(ConfigConstants.url)
      dataSource.setUsername(ConfigConstants.user)
      dataSource.setPassword(ConfigConstants.password)
      dataSource.setMaxTotal(50)
      dataSource.setInitialSize(3)
      dataSource.setMinIdle(3)
      dataSource.setMaxIdle(10)
      dataSource.setMaxWaitMillis(2 * 10000)
      dataSource.setRemoveAbandonedTimeout(180)
      dataSource.setRemoveAbandonedOnBorrow(true)
      dataSource.setRemoveAbandonedOnMaintenance(true)
      dataSource.setTestOnReturn(true)
      dataSource.setTestOnBorrow(true)
    }
    return dataSource
  }
  /**
    * 释放数据源
    */
  def closeDataSource() = {
    if (dataSource != null) {
      dataSource.close()
    }
  }
  /**
    * 获取数据库连接
    *
    * @return
    */
  def getConnection(): Connection = {
    var conn: Connection = null
    try {
      if (dataSource != null) {
        conn = dataSource.getConnection()
      } else {
        conn = getDataSource().getConnection()
      }
    } catch {
      case e: Exception =>
        log.error(e.getMessage(), e)
    }
    conn
  }

  /**
    * 关闭连接
    */
 def closeConnection (ps:PreparedStatement , conn:Connection ) {
    if (ps != null) {
      try {
        ps.close();
      } catch  {
        case e:Exception =>
          log.error("预编译SQL语句对象PreparedStatement关闭异常！" + e.getMessage(), e);
      }
    }
    if (conn != null) {
      try {
        conn.close();
      } catch  {
        case e:Exception =>
        log.error("关闭连接对象Connection异常！" + e.getMessage(), e);
      }
    }
  }
}

• Kafka生产者：KafkaProducerTest

object KafkaProducerTest {
  def main(args: Array[String]): Unit = {
    val  props : Properties = new Properties()
    props.put("bootstrap.servers", ConfigConstants.kafkaBrokers)
    props.put("batch.size", ConfigConstants.batchSize.asInstanceOf[Integer])
    props.put("linger.ms", ConfigConstants.lingerMs.asInstanceOf[Integer])
    props.put("buffer.memory", ConfigConstants.bufferMemory.asInstanceOf[Integer])
    props.put("key.serializer",ConfigConstants.kafkaKeySer)
    props.put("value.serializer", ConfigConstants.kafkaValueSer)
   val  producer :  Producer[String, String] = new KafkaProducer[String, String](props)
    val startTime : Long  = System.currentTimeMillis()
    for ( i <- 1 to 100) {
      producer.send(new ProducerRecord[String, String](ConfigConstants.kafkaTopics, "Spark", Integer.toString(i)))
    }
  println("消耗时间：" + (System.currentTimeMillis() - startTime))
    producer.close()
  }
}

• 读取和保存Offset：

该对象的作用是从外部设备中读取和写入Offset，包括MySQL和Redis

object OffsetReadAndSave {

  /**
    * 从MySQL中获取偏移量
    *
    * @param groupid
    * @param topic
    * @return
    */

  def getOffsetMap(groupid: String, topic: String): mutable.Map[TopicPartition, Long] = {

    val conn = JDBCConnPool.getConnection()
    val selectSql = "select * from topic_par_group_offset where groupid = ? and topic = ?"
    val ppst = conn.prepareStatement(selectSql)
    ppst.setString(1, groupid)
    ppst.setString(2, topic)

    val result: ResultSet = ppst.executeQuery()

    // 主题分区偏移量
    val topicPartitionOffset = mutable.Map[TopicPartition, Long]()

    while (result.next()) {

      val topicPartition: TopicPartition = new TopicPartition(result.getString("topic"), result.getInt("partition"))

      topicPartitionOffset += (topicPartition -> result.getLong("offset"))
    }

    JDBCConnPool.closeConnection(ppst, conn)
    topicPartitionOffset
  }

  /**
    * 从Redis中获取偏移量
    *
    * @param groupid
    * @param topic
    * @return
    */
  def getOffsetFromRedis(groupid: String, topic: String): Map[TopicPartition, Long] = {
    val jedis: Jedis = JedisConnPool.getConnection()
    var offsets = mutable.Map[TopicPartition, Long]()

    val key = s"${topic}_${groupid}"
    val fields : java.util.Map[String, String] = jedis.hgetAll(key)
    for (partition <- JavaConversions.mapAsScalaMap(fields)) {

      offsets.put(new TopicPartition(topic, partition._1.toInt), partition._2.toLong)
    }

    offsets.toMap

  }
  /**
    * 将偏移量写入MySQL
    *
    * @param groupid     消费者组ID
    * @param offsetRange 消息偏移量范围
    */

  def saveOffsetRanges(groupid: String, offsetRange: Array[OffsetRange]) = {

    val conn = JDBCConnPool.getConnection()
    val insertSql = "replace into topic_par_group_offset(`topic`, `partition`, `groupid`, `offset`) values(?,?,?,?)"
    val ppst = conn.prepareStatement(insertSql)

    for (offset <- offsetRange) {

      ppst.setString(1, offset.topic)
      ppst.setInt(2, offset.partition)
      ppst.setString(3, groupid)
      ppst.setLong(4, offset.untilOffset)
      ppst.executeUpdate()
    }
    JDBCConnPool.closeConnection(ppst, conn)

  }
  /**
    * 将偏移量保存到Redis中
    * @param groupid
    * @param offsetRange
    */
  def saveOffsetToRedis(groupid: String, offsetRange: Array[OffsetRange]) = {
    val jedis :Jedis = JedisConnPool.getConnection()
    for(offsetRange<-offsetRange){
      val topic=offsetRange.topic
      val partition=offsetRange.partition
      val offset=offsetRange.untilOffset
      // key为topic_groupid,field为partition，value为offset
      jedis.hset(s"${topic}_${groupid}",partition.toString,offset.toString)
    }
  }
}

• 业务处理类

该对象是业务处理逻辑，主要是消费Kafka数据，再处理之后进行手动将偏移量保存到MySQL中。在启动程序时，会判断外部存储设备中是否存在偏移量，如果是首次启动则从最初的消费位点消费，如果存在Offset，则从当前的Offset去消费。

观察现象：当首次启动时会从头消费数据，手动停止程序，然后再次启动，会发现会从当前提交的偏移量消费数据。

object ManualCommitOffset {
  
  def main(args: Array[String]): Unit = {

    val brokers = ConfigConstants.kafkaBrokers
    val groupId = ConfigConstants.groupId
    val topics = ConfigConstants.kafkaTopics
    val batchInterval = ConfigConstants.batchInterval

    val conf = new SparkConf()
      .setAppName(ManualCommitOffset.getClass.getSimpleName)
      .setMaster("local[1]")
      .set("spark.serializer",ConfigConstants.sparkSerializer)

    val ssc = new StreamingContext(conf, batchInterval)
    // 必须开启checkpoint,否则会报错
    ssc.checkpoint(ConfigConstants.checkpointDir)

    ssc.sparkContext.setLogLevel("OFF")
    //使用broker和topic创建direct kafka stream
    val topicSet = topics.split(" ").toSet

    // kafka连接参数
    val kafkaParams = Map[String, Object](
      ConsumerConfig.BOOTSTRAP_SERVERS_CONFIG -> brokers,
      ConsumerConfig.GROUP_ID_CONFIG -> groupId,
      ConsumerConfig.KEY_DESERIALIZER_CLASS_CONFIG -> classOf[StringDeserializer],
      ConsumerConfig.VALUE_DESERIALIZER_CLASS_CONFIG -> classOf[StringDeserializer],
      ConsumerConfig.ENABLE_AUTO_COMMIT_CONFIG -> (false: java.lang.Boolean),
      ConsumerConfig.AUTO_OFFSET_RESET_CONFIG -> "earliest"
    )


    // 从MySQL中读取该主题对应的消费者组的分区偏移量
    val offsetMap = OffsetReadAndSave.getOffsetMap(groupId, topics)
    var inputDStream: InputDStream[ConsumerRecord[String, String]] = null

    //如果MySQL中已经存在了偏移量,则应该从该偏移量处开始消费
    if (offsetMap.size > 0) {
      println("存在偏移量，从该偏移量处进行消费！！")

      inputDStream = KafkaUtils.createDirectStream[String, String](
        ssc,
        LocationStrategies.PreferConsistent,
        ConsumerStrategies.Subscribe[String, String](topicSet, kafkaParams, offsetMap))

    } else {
      //如果MySQL中没有存在了偏移量，从最早开始消费
      inputDStream = KafkaUtils.createDirectStream[String, String](
        ssc,
        LocationStrategies.PreferConsistent,
        ConsumerStrategies.Subscribe[String, String](topicSet, kafkaParams))

    }
    // checkpoint时间间隔，必须是batchInterval的整数倍
    inputDStream.checkpoint(ConfigConstants.checkpointInterval)

    // 保存batch的offset
    var offsetRanges = Array[OffsetRange]()
    // 获取当前DS的消息偏移量
    val transformDS = inputDStream.transform { rdd =>
      // 获取offset
      offsetRanges = rdd.asInstanceOf[HasOffsetRanges].offsetRanges
      rdd
    }


    /**
      * 状态更新函数
      * @param newValues:新的value值
      * @param stateValue：状态值
      * @return
      */
    def updateFunc(newValues: Seq[Int], stateValue: Option[Int]): Option[Int] = {
      var oldvalue = stateValue.getOrElse(0) // 获取状态值
      // 遍历当前数据，并更新状态
      for (newValue <- newValues) {
        oldvalue += newValue
      }
      // 返回最新的状态
      Option(oldvalue)
    }
    // 业务逻辑处理
    // 该示例统计消息key的个数，用于查看是否是从已经提交的偏移量消费数据
    transformDS.map(meg => ("spark", meg.value().toInt)).updateStateByKey(updateFunc).print()

    // 打印偏移量和数据信息，观察输出的结果
    transformDS.foreachRDD { (rdd, time) =>
      // 遍历打印该RDD数据
      rdd.foreach { record =>
        println(s"key=${record.key()},value=${record.value()},partition=${record.partition()},offset=${record.offset()}")
      }
      // 打印消费偏移量信息
      for (o <- offsetRanges) {
        println(s"topic=${o.topic},partition=${o.partition},fromOffset=${o.fromOffset},untilOffset=${o.untilOffset},time=${time}")

      }

      //将偏移量保存到到MySQL中
      OffsetReadAndSave.saveOffsetRanges(groupId, offsetRanges)
    }
    ssc.start()
    ssc.awaitTermination()
  }
}

如何使用Redis管理Kafka的OffSet

• Redis连接类

object JedisConnPool {
  val config = new JedisPoolConfig
  //最大连接数
  config.setMaxTotal(60)
  //最大空闲连接数
  config.setMaxIdle(10)
  config.setTestOnBorrow(true)

  //服务器ip
  val redisAddress :String = ConfigConstants.redisAddress.toString
  // 端口号
  val redisPort:Int = ConfigConstants.redisPort.toInt
  //访问密码
  val redisAuth :String = ConfigConstants.redisAuth.toString
  //等待可用连接的最大时间
  val redisTimeout:Int = ConfigConstants.redisTimeout.toInt

  val pool = new JedisPool(config,redisAddress,redisPort,redisTimeout,redisAuth)

  def getConnection():Jedis = {
    pool.getResource
  }

}

• 业务逻辑处理

该对象与上面的基本类似，只不过使用的是Redis来进行存储Offset，存储到Redis的数据类型是Hash，基本格式为：[key field value] -> [ topic_groupid partition offset]，即 key为topic_groupid,field为partition，value为offset。

object ManualCommitOffsetToRedis {

  def main(args: Array[String]): Unit = {

    val brokers = ConfigConstants.kafkaBrokers
    val groupId = ConfigConstants.groupId
    val topics = ConfigConstants.kafkaTopics
    val batchInterval = ConfigConstants.batchInterval

    val conf = new SparkConf()
      .setAppName(ManualCommitOffset.getClass.getSimpleName)
      .setMaster("local[1]")
      .set("spark.serializer", ConfigConstants.sparkSerializer)


    val ssc = new StreamingContext(conf, batchInterval)
    // 必须开启checkpoint,否则会报错
    ssc.checkpoint(ConfigConstants.checkpointDir)

    ssc.sparkContext.setLogLevel("OFF")
    //使用broker和topic创建direct kafka stream
    val topicSet = topics.split(" ").toSet

    // kafka连接参数
    val kafkaParams = Map[String, Object](
      ConsumerConfig.BOOTSTRAP_SERVERS_CONFIG -> brokers,
      ConsumerConfig.GROUP_ID_CONFIG -> groupId,
      ConsumerConfig.KEY_DESERIALIZER_CLASS_CONFIG -> classOf[StringDeserializer],
      ConsumerConfig.VALUE_DESERIALIZER_CLASS_CONFIG -> classOf[StringDeserializer],
      ConsumerConfig.ENABLE_AUTO_COMMIT_CONFIG -> (false: java.lang.Boolean),
      ConsumerConfig.AUTO_OFFSET_RESET_CONFIG -> "earliest"
    )


    // 从Redis中读取该主题对应的消费者组的分区偏移量
    val offsetMap = OffsetReadAndSave.getOffsetFromRedis(groupId, topics)
    var inputDStream: InputDStream[ConsumerRecord[String, String]] = null

    //如果Redis中已经存在了偏移量,则应该从该偏移量处开始消费
    if (offsetMap.size > 0) {
      println("存在偏移量，从该偏移量处进行消费！！")

      inputDStream = KafkaUtils.createDirectStream[String, String](
        ssc,
        LocationStrategies.PreferConsistent,
        ConsumerStrategies.Subscribe[String, String](topicSet, kafkaParams, offsetMap))

    } else {
      //如果Redis中没有存在了偏移量，从最早开始消费
      inputDStream = KafkaUtils.createDirectStream[String, String](
        ssc,
        LocationStrategies.PreferConsistent,
        ConsumerStrategies.Subscribe[String, String](topicSet, kafkaParams))

    }
    // checkpoint时间间隔，必须是batchInterval的整数倍
    inputDStream.checkpoint(ConfigConstants.checkpointInterval)

    // 保存batch的offset
    var offsetRanges = Array[OffsetRange]()
    // 获取当前DS的消息偏移量
    val transformDS = inputDStream.transform { rdd =>
      // 获取offset
      offsetRanges = rdd.asInstanceOf[HasOffsetRanges].offsetRanges
      rdd
    }


    /**
      * 状态更新函数
      *
      * @param newValues  :新的value值
      * @param stateValue ：状态值
      * @return
      */
    def updateFunc(newValues: Seq[Int], stateValue: Option[Int]): Option[Int] = {
      var oldvalue = stateValue.getOrElse(0) // 获取状态值
      // 遍历当前数据，并更新状态
      for (newValue <- newValues) {
        oldvalue += newValue
      }
      // 返回最新的状态
      Option(oldvalue)
    }
    // 业务逻辑处理
    // 该示例统计消息key的个数，用于查看是否是从已经提交的偏移量消费数据
    transformDS.map(meg => ("spark", meg.value().toInt)).updateStateByKey(updateFunc).print()

    // 打印偏移量和数据信息，观察输出的结果
    transformDS.foreachRDD { (rdd, time) =>
      // 遍历打印该RDD数据
      rdd.foreach { record =>
        println(s"key=${record.key()},value=${record.value()},partition=${record.partition()},offset=${record.offset()}")
      }
      // 打印消费偏移量信息
      for (o <- offsetRanges) {
        println(s"topic=${o.topic},partition=${o.partition},fromOffset=${o.fromOffset},untilOffset=${o.untilOffset},time=${time}")

      }

      //将偏移量保存到到Redis中
      OffsetReadAndSave.saveOffsetToRedis(groupId, offsetRanges)
    }
    ssc.start()
    ssc.awaitTermination()
  }

}

总结

本文介绍了如何使用外部存储设备来保存Kafka的消费位点，通过详细的代码示例说明了使用MySQL和Redis管理消费位点的方式。当然，外部存储设备很多，用户也可以使用其他的存储设备进行管理Offset，比如Zookeeper和HBase等，其基本处理思路都十分相似。

大数据技术与数仓