Spark Streaming发行版笔记15：no receivers彻底思考

Posted 2020-07-17 永不服输2016

数据接入Spark Streaming的二种方式：Receiver和no receivers方式

建议企业级采用no receivers方式开发Spark Streaming应用程序，好处：

1、更优秀的自由度控制

2、语义一致性

no receivers更符合数据读取和数据操作，Spark 计算框架底层有数据来源，如果只有direct直接操作数据来源则更天然。操作数据来源封装其一定是rdd级别的。

所以Spark 推出了自定义的rdd即Kafkardd，只是数据来源不同。

进入源码区：

注释基于Batch消费数据，首先确定开始和结束的offSet，特别强调语义一致性。

关键是metaData.broker.list，通过bootstrap.servers直接操作Kafka集群，操作kafka数据是一个offset范围。

/**
 * A batch-oriented interface for consuming from Kafka.
 * Starting and ending offsets are specified in advance,
 * so that you can control exactly-once semantics.
 * @param kafkaParams Kafka <a href="http://kafka.apache.org/documentation.html#configuration">
 * configuration parameters</a>. Requires "metadata.broker.list" or "bootstrap.servers" to be set
 * with Kafka broker(s) specified in host1:port1,host2:port2 form.
 * @param offsetRanges offset ranges that define the Kafka data belonging to this RDD
 * @param messageHandler function for translating each message into the desired type
 */
private[kafka]
class KafkaRDD[
  K: ClassTag,
  V: ClassTag,
  U <: Decoder[_]: ClassTag,
  T <: Decoder[_]: ClassTag,
  R: ClassTag] private[spark] (
    sc: SparkContext,
    kafkaParams: Map[String, String],
    val offsetRanges: Array[OffsetRange],
    leaders: Map[TopicAndPartition, (String, Int)],
    messageHandler: MessageAndMetadata[K, V] => R
  ) extends RDD[R](sc, Nil) with Logging with HasOffsetRanges {

你要直接访问Kafka中的数据需要自定义一个KafkaRDD，如果读取hBase上的数据

也必须自定义一个hBaseRDD。有一点必须定义接口HasOffsetRange，RDD天然的是一个

A List Partitions，基于kafka直接访问RDD时必须是HasOffsetRange类型，代表了

来自kafka topicAndParttion，其实力被HasOffsetRange Create创建，从fromOffset到untilOffset ，

分布式传输Offset数据时必须序列化。

/**
 * Represents any object that has a collection of [[OffsetRange]]s. This can be used to access the
 * offset ranges in RDDs generated by the direct Kafka DStream (see
 * [[KafkaUtils.createDirectStream()]]).
 * {{{
 *   KafkaUtils.createDirectStream(...).foreachRDD { rdd =>
 *      val offsetRanges = rdd.asInstanceOf[HasOffsetRanges].offsetRanges
 *      ...
 *   }
 * }}}
 */
trait HasOffsetRanges {
  def offsetRanges: Array[OffsetRange]
}

/**
 * Represents a range of offsets from a single Kafka TopicAndPartition. Instances of this class
 * can be created with `OffsetRange.create()`.
 * @param topic Kafka topic name
 * @param partition Kafka partition id
 * @param fromOffset Inclusive starting offset
 * @param untilOffset Exclusive ending offset
 */
final class OffsetRange private(
    val topic: String,
    val partition: Int,
    val fromOffset: Long,
    val untilOffset: Long) extends Serializable {
  import OffsetRange.OffsetRangeTuple

  /** Kafka TopicAndPartition object, for convenience */
  def topicAndPartition(): TopicAndPartition = TopicAndPartition(topic, partition)

  /** Number of messages this OffsetRange refers to */
  def count(): Long = untilOffset - fromOffset

Offset是消息偏移量，假设untilOffset是10万，fromOffset是5万，第10万条消息

和5万条消息，一般处理数据规模大小是以数据条数为单位。

创建一个offSetrange实例时可以确定从kafka集群partition中读取哪些topic，从

foreachrdd中可以获得当前rdd访问的所有分区数据。Batch Duration中产生的rdd的分区数据，这个是对元数据的控制。

在看getPartitions方法，offsetRanges指定了每个offsetrange从什么位置开始到什么位置结束。

override def getPartitions: Array[Partition] = {
  offsetRanges.zipWithIndex.map { case (o, i) =>
      val (host, port) = leaders(TopicAndPartition(o.topic, o.partition))
      new KafkaRDDPartition(i, o.topic, o.partition, o.fromOffset, o.untilOffset, host, port)

 }.toArray
}

看KafkaRDDPartition类，会从传入的topic和partition及offset中获取kafka数据

/** @param topic kafka topic name
  * @param partition kafka partition id
  * @param fromOffset inclusive starting offset
  * @param untilOffset exclusive ending offset
  * @param host preferred kafka host, i.e. the leader at the time the rdd was created
  * @param port preferred kafka host\'s port
  */
private[kafka]
class KafkaRDDPartition(
  val index: Int,
  val topic: String,
  val partition: Int,
  val fromOffset: Long,
  val untilOffset: Long,
  val host: String,
  val port: Int
) extends Partition {
  /** Number of messages this partition refers to */
  def count(): Long = untilOffset - fromOffset
}

Host port指定读取数据来源的kfakf机器。

看kafka rdd的compute计算每个数据分片，和rdd理念是一样的，每次迭代操作获取计算的rdd一部分。

操作KafkaRDDIterator和操作rdd分片是一样的，需要迭代数据分片：

override def compute(thePart: Partition, context: TaskContext): Iterator[R] = {
  val part = thePart.asInstanceOf[KafkaRDDPartition]
  assert(part.fromOffset <= part.untilOffset, errBeginAfterEnd(part))
  if (part.fromOffset == part.untilOffset) {
    log.info(s"Beginning offset ${part.fromOffset} is the same as ending offset " +
      s"skipping ${part.topic} ${part.partition}")
    Iterator.empty
  } else {
    new KafkaRDDIterator(part, context)
  }
}

private class KafkaRDDIterator(
    part: KafkaRDDPartition,
    context: TaskContext) extends NextIterator[R] {

  context.addTaskCompletionListener{ context => closeIfNeeded() }

  log.info(s"Computing topic ${part.topic}, partition ${part.partition} " +
    s"offsets ${part.fromOffset} -> ${part.untilOffset}")

  val kc = new KafkaCluster(kafkaParams)
  val keyDecoder = classTag[U].runtimeClass.getConstructor(classOf[VerifiableProperties])
    .newInstance(kc.config.props)
    .asInstanceOf[Decoder[K]]
  val valueDecoder = classTag[T].runtimeClass.getConstructor(classOf[VerifiableProperties])
    .newInstance(kc.config.props)
    .asInstanceOf[Decoder[V]]
  val consumer = connectLeader
  var requestOffset = part.fromOffset
  var iter: Iterator[MessageAndOffset] = null

  // The idea is to use the provided preferred host, except on task retry attempts,
  // to minimize number of kafka metadata requests
  private def connectLeader: SimpleConsumer = {
    if (context.attemptNumber > 0) {
      kc.connectLeader(part.topic, part.partition).fold(
        errs => throw new SparkException(
          s"Couldn\'t connect to leader for topic ${part.topic} ${part.partition}: " +
            errs.mkString("\\n")),
        consumer => consumer
      )
    } else {
      kc.connect(part.host, part.port)
    }
  }

  private def handleFetchErr(resp: FetchResponse) {
    if (resp.hasError) {
      val err = resp.errorCode(part.topic, part.partition)
      if (err == ErrorMapping.LeaderNotAvailableCode ||
        err == ErrorMapping.NotLeaderForPartitionCode) {
        log.error(s"Lost leader for topic ${part.topic} partition ${part.partition}, " +
          s" sleeping for ${kc.config.refreshLeaderBackoffMs}ms")
        Thread.sleep(kc.config.refreshLeaderBackoffMs)
      }
      // Let normal rdd retry sort out reconnect attempts
      throw ErrorMapping.exceptionFor(err)
    }
  }

关键的地方kafkaCluster对象时在kafkaUtils中直接创建了directStream，看下之前操作kafka代码发现传入的参数是上下文、 broker.List.topic.list参数：

构建时传入topics为Set，当然可以直接指定ranges，他从kafka集群直接创建了kafkaCluster和集群进行交互，从fromOffset获取数据具体的偏移量：

/**
 * Create an input stream that directly pulls messages from Kafka Brokers
 * without using any receiver. This stream can guarantee that each message
 * from Kafka is included in transformations exactly once (see points below).
 *
 * Points to note:
 *  - No receivers: This stream does not use any receiver. It directly queries Kafka
 *  - Offsets: This does not use Zookeeper to store offsets. The consumed offsets are tracked
 *    by the stream itself. For interoperability with Kafka monitoring tools that depend on
 *    Zookeeper, you have to update Kafka/Zookeeper yourself from the streaming application.
 *    You can access the offsets used in each batch from the generated RDDs (see
 *    [[org.apache.spark.streaming.kafka.HasOffsetRanges]]).
 *  - Failure Recovery: To recover from driver failures, you have to enable checkpointing
 *    in the [[StreamingContext]]. The information on consumed offset can be
 *    recovered from the checkpoint. See the programming guide for details (constraints, etc.).
 *  - End-to-end semantics: This stream ensures that every records is effectively received and
 *    transformed exactly once, but gives no guarantees on whether the transformed data are
 *    outputted exactly once. For end-to-end exactly-once semantics, you have to either ensure
 *    that the output operation is idempotent, or use transactions to output records atomically.
 *    See the programming guide for more details.
 *
 * @param ssc StreamingContext object
 * @param kafkaParams Kafka <a href="http://kafka.apache.org/documentation.html#configuration">
 *   configuration parameters</a>. Requires "metadata.broker.list" or "bootstrap.servers"
 *   to be set with Kafka broker(s) (NOT zookeeper servers), specified in
 *   host1:port1,host2:port2 form.
 *   If not starting from a checkpoint, "auto.offset.reset" may be set to "largest" or "smallest"
 *   to determine where the stream starts (defaults to "largest")
 * @param topics Names of the topics to consume
 * @tparam K type of Kafka message key
 * @tparam V type of Kafka message value
 * @tparam KD type of Kafka message key decoder
 * @tparam VD type of Kafka message value decoder
 * @return DStream of (Kafka message key, Kafka message value)
 */
def createDirectStream[
  K: ClassTag,
  V: ClassTag,
  KD <: Decoder[K]: ClassTag,
  VD <: Decoder[V]: ClassTag] (
    ssc: StreamingContext,
    kafkaParams: Map[String, String],
    topics: Set[String]
): InputDStream[(K, V)] = {
  val messageHandler = (mmd: MessageAndMetadata[K, V]) => (mmd.key, mmd.message)
  val kc = new KafkaCluster(kafkaParams)
  val fromOffsets = getFromOffsets(kc, kafkaParams, topics)
  new DirectKafkaInputDStream[K, V, KD, VD, (K, V)](
    ssc, kafkaParams, fromOffsets, messageHandler)

看下getFromOffsets方法：

private[kafka] def getFromOffsets(
    kc: KafkaCluster,
    kafkaParams: Map[String, String],
    topics: Set[String]
  ): Map[TopicAndPartition, Long] = {
  val reset = kafkaParams.get("auto.offset.reset").map(_.toLowerCase)
  val result = for {
    topicPartitions <- kc.getPartitions(topics).right
    leaderOffsets <- (if (reset == Some("smallest")) {
      kc.getEarliestLeaderOffsets(topicPartitions)
    } else {
      kc.getLatestLeaderOffsets(topicPartitions)
    }).right
  } yield {
    leaderOffsets.map { case (tp, lo) =>
        (tp, lo.offset)
    }
  }
  KafkaCluster.checkErrors(result)
}

如果不知道fromOffsets的话直接从配置中获取fromOffsets，创建kafka DirectKafkaInputDStream的时候会从kafka集群进行交互获得partition、offset信息，通过DirectKafkaInputDStream无论什么情况最后都会创建DirectKafkaInputDStream：

/**
 *  A stream of {@link org.apache.spark.streaming.kafka.KafkaRDD} where
 * each given Kafka topic/partition corresponds to an RDD partition.
 * The spark configuration spark.streaming.kafka.maxRatePerPartition gives the maximum number
 *  of messages
 * per second that each \'\'\'partition\'\'\' will accept.
 * Starting offsets are specified in advance,
 * and this DStream is not responsible for committing offsets,
 * so that you can control exactly-once semantics.
 * For an easy interface to Kafka-managed offsets,
 *  see {@link org.apache.spark.streaming.kafka.KafkaCluster}
 * @param kafkaParams Kafka <a href="http://kafka.apache.org/documentation.html#configuration">
 * configuration parameters</a>.
 *   Requires "metadata.broker.list" or "bootstrap.servers" to be set with Kafka broker(s),
 *   NOT zookeeper servers, specified in host1:port1,host2:port2 form.
 * @param fromOffsets per-topic/partition Kafka offsets defining the (inclusive)
 *  starting point of the stream
 * @param messageHandler function for translating each message into the desired type
 */
private[streaming]
class DirectKafkaInputDStream[
  K: ClassTag,
  V: ClassTag,
  U <: Decoder[K]: ClassTag,
  T <: Decoder[V]: ClassTag,
  R: ClassTag](
    ssc_ : StreamingContext,
    val kafkaParams: Map[String, String],
    val fromOffsets: Map[TopicAndPartition, Long],
    messageHandler: MessageAndMetadata[K, V] => R
  ) extends InputDStream[R](ssc_) with Logging {
  val maxRetries = context.sparkContext.getConf.getInt(
    "spark.streaming.kafka.maxRetries", 1)

DirectKafkaInputDStream会产生kafkaRDD，不同的topic partitions生成对应的的kafkarddpartitions，控制消费读取速度。操作数据的时候是compute直接构建出kafka rdd，读取kafka 上的数据。确定获取读取数据的期间就知道需要读取多少条数据，然后构建kafkardd实例。Kafkardd的实例和DirectKafkaInputDStream是一一对应的，每次compute会产生一个kafkardd，其会包含很多partitions，有多少partition就是对应多少kafkapartition。

看下KafkaRDDPartition就是一个简单的数据结构：

/** @param topic kafka topic name
  * @param partition kafka partition id
  * @param fromOffset inclusive starting offset
  * @param untilOffset exclusive ending offset
  * @param host preferred kafka host, i.e. the leader at the time the rdd was created
  * @param port preferred kafka host\'s port
  */
private[kafka]
class KafkaRDDPartition(
  val index: Int,
  val topic: String,
  val partition: Int,
  val fromOffset: Long,
  val untilOffset: Long,
  val host: String,
  val port: Int
) extends Partition {
  /** Number of messages this partition refers to */
  def count(): Long = untilOffset - fromOffset
}

总结：

而且KafkaRDDPartition只能属于一个topic，不能让partition跨多个topic，直接消费一个kafkatopic，topic不断进来、数据不断偏移，Offset代表kafka数据偏移量指针。

数据不断流进kafka，batchDuration假如每十秒都会从配置的topic中消费数据，每次会消费一部分直到消费完，下一个batchDuration会再流进来的数据，又可以从头开始读或上一个数据的基础上读取数据。

思考直接抓取kafka数据和receiver读取数据：

好处一：

直接抓取fakfa数据的好处，没有缓存，不会出现内存溢出等之类的问题。但是如果kafka Receiver的方式读取会存在缓存的问题，需要设置读取的频率和block interval等信息。

好处二：

采用receiver方式的话receiver默认情况需要和worker的executor绑定，不方便做分布式，当然可以配置成分布式，采用direct方式默认情况下数据会存在多个worker上的executor。Kafkardd数据默认都是分布在多个executor上的，天然数据是分布式的存在多个executor，而receiver就不方便计算。

好处三：

数据消费的问题，在实际操作的时候采用receiver的方式有个弊端，消费数据来不及处理即操作数据有deLay多才时，Spark Streaming程序有可能奔溃。但如果是direct方式访问kafka数据不会存在此类情况。因为diect方式直接读取kafka数据，如果delay就不进行下一个batchDuration读取。

好处四：

完全的语义一致性，不会重复消费数据，而且保证数据一定被消费，跟kafka进行交互，只有数据真正执行成功之后才会记录下来。

生产环境下强烈建议采用direct方式读取kafka数据。

新浪微博：http://weibo.com/ilovepains

微信公众号：DT_Spark

博客：http://blog.sina.com.cn/ilovepains

手机：[phone]

QQ：1740415547

邮箱：[e-mail]

Spark Streaming发行版笔记15