spark

Spark Streaming Source Kafka 0.8

2018-08-01  本文已影响4人  lioversky

描述

针对kafka0.8.2的API,Spark Streaming有两个版本的Source,Receiver和DirectAPI,其中Receiver模式使用HighLevel对应为KafkaInputDStream,继承自ReceiverInputDStream再继承InputDStream,DirectAPI使用SimpleConsumer对应DirectKafkaInputDStream直接继承自InputDStream。此二者的具体差别网上有更详细的描述,在此简单介绍一下:

正由于以上差别,在不同情景两种方式各有优劣,简单概括如下:

Receiver优势

  1. 由于两种方式kafka底层实现的不同,Receiver不需要自己管理offset而DirectAPI需要自己管理;
  2. 当消费的topic过多或者topic对应的partition过多时,DirectAPI会产生特别多的task,这会使spark对调度产生更大的开销;而Receiver可以通过spark参数控制;
  3. DirectAPI在生成Batch时,会首先查询topic的metadata数据,再根据topicAndPartition获取当前endOffset来决定Batch的起止offset,如果kafka的partition异常如topic对应的partition无leader,会无法生成新batch导致spark程序失败或者连接非partitionLeader,而Receiver不会出现此种错误;

DirectAPI优势

  1. 由于Receiver是预拉取,并且可以控制在处理完成batch后更新consumer的offset,所以在有数据堆积时程序异常终止会丢失数据,而DirectAPI不会丢失数据,可能会有数据重复消费情况;
  2. spark在2.0以前,在每个batch结束之后会调用cleanMetadata方法,此方法会清除当前batch及之前的所有数据,包括metadta和block,所以在开启并行(spark.streaming.concurrentJobs>1)如果后面的batch先执行完,会出现 block not found的异常,大部分都是由于此问题导致;
  3. Receiver模式在kafka不太稳定情况下,日志经常会出现reblance,并且有数据重复消费情况;

实现原理与使用

Receiver模式

Receiver模式实现类为KafkaInputDStream,继承自ReceiverInputDStream,实现其getReceiver方法。KafkaReceiver在onStart方法中创建指定线程数读取数据,再通过Receiver的store方法写到blockManager中;

  1. 下面代码是创建KafkaInputDStream并生成numStreams个Receiver:

    List<JavaPairDStream<String, String>> kafkaStreams = new ArrayList<JavaPairDStream<String, String>>(numStreams);
    for (int i = 0; i < numStreams; i++) {
      kafkaStreams.add(KafkaUtils.createStream(streamingContext,zookeeper, groupId, topicMap,storageLevel));
    }
    
  2. 下面代码负责为每个进程创建线程(源码):

    val topicMessageStreams = consumerConnector.createMessageStreams(
          topics, keyDecoder, valueDecoder)
    val executorPool =
          ThreadUtils.newDaemonFixedThreadPool(topics.values.sum, "KafkaMessageHandler")
    try {
      // Start the messages handler for each partition
      topicMessageStreams.values.foreach { streams =>
        streams.foreach { stream => executorPool.submit(new MessageHandler(stream)) }
      }
    } finally {
      executorPool.shutdown() // Just causes threads to terminate after work is done
    }
    
  3. 下面代码在MessageHandler负责读取处理消息数据(源码):

    val streamIterator = stream.iterator()
    while (streamIterator.hasNext()) {
      val msgAndMetadata = streamIterator.next()
      store((msgAndMetadata.key, msgAndMetadata.message))
    }
    
  4. ReceiverInputDStream的compute创建BlockRDD(源码):

    //ask the tracker for all the blocks that have been allocated to this stream
    // for this batch
    val blockInfos = receiverTracker.getBlocksOfBatch(validTime).getOrElse(id, Seq.empty)
    
    // Create the BlockRDD
    createBlockRDD(validTime, blockInfos)
    

DirectAPI模式

direct模式需要自己保存offset,可以通过checkpoint保存或者外部存储。使用KafkaUtils.createDirectStream创建InputDStream,此方法有多个重载,根据需要使用。

  1. 在每次启动时都要传入partitionOffset,offset从保存位置读取,创建stream代码如下:

    Map<TopicAndPartition, java.lang.Long> offsets = ....
    Class<Tuple2<String, String>> c = (Class<Tuple2<String, String>>) (Class) Tuple2.class;
    JavaInputDStream<Tuple2<String, String>> messages = KafkaUtils
        .createDirectStream(streamingContext, String.class,
            String.class, StringDecoder.class, StringDecoder.class, c, kafkaParams, offsets,
            new Function<MessageAndMetadata<String, String>, Tuple2<String, String>>() {
              public Tuple2<String, String> call(MessageAndMetadata<String, String> md)
                  throws Exception {
                return new Tuple2<String, String>(md.key(), md.message());
              }
            }
        );
    
  2. 在生成每个batch任务时,DirectKafkaInputDStream内保存当前已记录的offsets值,再通过获取latestLeaderOffsets,计算出本本次batch要处理的数据区间,此区间会被每个分区的配置最大消费数限制;

    val untilOffsets = clamp(latestLeaderOffsets(maxRetries))
    
  3. 正是由于每个batch都要获取当前各paritition最大的offset值,kc.getLatestLeaderOffsets(currentOffsets.keySet),所以在kafka的partition出现异常时会导致任务出错或者由于连接超时阻塞任务生成;

  4. 得到currentOffsets和untilOffsets后,创建KafkaRDD,rdd内部属性offsetRanges记录此rdd要处理的各partition的offset区间值,通过此属性生成对应数量的KafkaRDDPartition。

    val rdd = KafkaRDD[K, V, U, T, R](
      context.sparkContext, kafkaParams, currentOffsets, untilOffsets, messageHandler)
    
    val offsetRanges = fromOffsets.map { case (tp, fo) =>
        val uo = untilOffsets(tp)
        OffsetRange(tp.topic, tp.partition, fo, uo.offset)
    }.toArray
    
    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
    }
    
  5. 在执行每个batch job时,为每个partition生成KafkaRDDIterator实例,根据每个partition中的信息创建SimpleConsumer连接,并认为此节点为对应kafka Partition的Leader,如果在此之前切换Leader也会出现kafka异常。在KafkaRDDIterator中通过getNext方法即时获取数据。

上一篇下一篇

猜你喜欢

热点阅读