一篇文章了解 Kafka 幂等性的原理及实践

01 幂等性如此重要

Kafka作为分布式MQ，大量用于分布式系统中，如消息推送系统、业务平台系统（如结算平台），就拿结算来说，业务方作为上游把数据打到结算平台，如果一份数据被计算、处理了多次，产生的后果将会特别严重。

02 哪些因素影响幂等性

使用Kafka时,需要保证exactly-once语义。要知道在分布式系统中，出现网络分区是不可避免的，如果kafka broker 在回复ack时，出现网络故障或者是full gc导致ack timeout，producer将会重发，如何保证producer重试时不造成重复or乱序？又或者producer 挂了，新的producer并没有old producer的状态数据，这个时候如何保证幂等？即使Kafka 发送消息满足了幂等，consumer拉取到消息后，把消息交给线程池workers，workers线程对message的处理可能包含异步操作，又会出现以下情况：

• 先commit，再执行业务逻辑：提交成功，处理失败 。造成丢失

• 先执行业务逻辑，再commit：提交失败，执行成功。造成重复执行

• 先执行业务逻辑，再commit：提交成功，异步执行fail。造成丢失

本文将针对以上问题作出讨论

03 Kafka保证发送幂等性

       针对以上的问题，kafka在0.11版新增了幂等型producer和事务型producer。前者解决了单会话幂等性等问题，后者解决了多会话幂等性。

单会话幂等性

为解决producer重试引起的乱序和重复。Kafka增加了pid和seq。Producer中每个RecordBatch都有一个单调递增的seq; Broker上每个tp也会维护pid-seq的映射，并且每Commit都会更新lastSeq。这样recordBatch到来时，broker会先检查RecordBatch再保存数据：如果batch中 baseSeq(第一条消息的seq)比Broker维护的序号(lastSeq)大1，则保存数据，否则不保存(inSequence方法)。

ProducerStateManager.scala

private def maybeValidateAppend(producerEpoch: Short, firstSeq: Int, offset: Long): Unit = {	
    validationType match {	
      case ValidationType.None =>	
      case ValidationType.EpochOnly =>	
        checkProducerEpoch(producerEpoch, offset)	
      case ValidationType.Full =>	
        checkProducerEpoch(producerEpoch, offset)	
        checkSequence(producerEpoch, firstSeq, offset)	
    }	
}	
private def checkSequence(producerEpoch: Short, appendFirstSeq: Int, offset: Long): Unit = {	
  if (producerEpoch != updatedEntry.producerEpoch) {	
    if (appendFirstSeq != 0) {	
      if (updatedEntry.producerEpoch != RecordBatch.NO_PRODUCER_EPOCH) {	
        throw new OutOfOrderSequenceException(s"Invalid sequence number for new epoch at offset $offset in " +	
          s"partition $topicPartition: $producerEpoch (request epoch), $appendFirstSeq (seq. number)")	
      } else {	
        throw new UnknownProducerIdException(s"Found no record of producerId=$producerId on the broker at offset $offset" +	
          s"in partition $topicPartition. It is possible that the last message with the producerId=$producerId has " +	
          "been removed due to hitting the retention limit.")	
      }	
    }	
  } else {	
    val currentLastSeq = if (!updatedEntry.isEmpty)	
      updatedEntry.lastSeq	
    else if (producerEpoch == currentEntry.producerEpoch)	
      currentEntry.lastSeq	
    else	
      RecordBatch.NO_SEQUENCE	
    if (currentLastSeq == RecordBatch.NO_SEQUENCE && appendFirstSeq != 0) {	

	
ne throw mew UnknownProducerIdException(s"Local producer state matches expected epoch $producerEpoch " +	
        s"for producerId=$producerId at offset $offset in partition $topicPartition, but the next expected " +	
        "sequence number is not known.")	
    } else if (!inSequence(currentLastSeq, appendFirstSeq)) {	
      throw new OutOfOrderSequenceException(s"Out of order sequence number for producerId $producerId at " +	
        s"offset $offset in partition $topicPartition: $appendFirstSeq (incoming seq. number), " +	
        s"$currentLastSeq (current end sequence number)")	
    }	
  }	
}	
  private def inSequence(lastSeq: Int, nextSeq: Int): Boolean = {	
    nextSeq == lastSeq + 1L || (nextSeq == 0 && lastSeq == Int.MaxValue)	
  }

引申：Kafka producer 对有序性做了哪些处理

假设我们有5个请求，batch1、batch2、batch3、batch4、batch5；如果只有batch2 ack failed，3、4、5都保存了，那2将会随下次batch重发而造成重复。我们可以设置max.in.flight.requests.per.connection=1（客户端在单个连接上能够发送的未响应请求的个数）来解决乱序，但降低了系统吞吐。

新版本kafka设置enable.idempotence=true后能够动态调整max-in-flight-request。正常情况下max.in.flight.requests.per.connection大于1。当重试请求到来且时，batch 会根据 seq重新添加到队列的合适位置，并把max.in.flight.requests.per.connection设为1，这样它 前面的 batch序号都比它小，只有前面的都发完了，它才能发。

    private void insertInSequenceOrder(Deque<ProducerBatch> deque, ProducerBatch batch) {	
        // When we are requeing and have enabled idempotence, the reenqueued batch must always have a sequence.	
        if (batch.baseSequence() == RecordBatch.NO_SEQUENCE)	
            throw new IllegalStateException("Trying to re-enqueue a batch which doesn't have a sequence even " +	
                "though idempotency is enabled.");	
        if (transactionManager.nextBatchBySequence(batch.topicPartition) == null)	
            throw new IllegalStateException("We are re-enqueueing a batch which is not tracked as part of the in flight " +	
                "requests. batch.topicPartition: " + batch.topicPartition + "; batch.baseSequence: " + batch.baseSequence());	
        ProducerBatch firstBatchInQueue = deque.peekFirst();	
        if (firstBatchInQueue != null && firstBatchInQueue.hasSequence() && firstBatchInQueue.baseSequence() < batch.baseSequence()) {	
            List<ProducerBatch> orderedBatches = new ArrayList<>();	
            while (deque.peekFirst() != null && deque.peekFirst().hasSequence() && deque.peekFirst().baseSequence() < batch.baseSequence())	
                orderedBatches.add(deque.pollFirst());	
            log.debug("Reordered incoming batch with sequence {} for partition {}. It was placed in the queue at " +	
                "position {}", batch.baseSequence(), batch.topicPartition, orderedBatches.size())	

	
            deque.addFirst(batch);	
            // Now we have to re insert the previously queued batches in the right order.	
            for (int i = orderedBatches.size() - 1; i >= 0; --i) {	
                deque.addFirst(orderedBatches.get(i));	
            }	
            // At this point, the incoming batch has been queued in the correct place according to its sequence.	
        } else {	
            deque.addFirst(batch);	
        }	
    }

多会话幂等性

在单会话幂等性中介绍，kafka通过引入pid和seq来实现单会话幂等性，但正是引入了pid，当应用重启时，新的producer并没有old producer的状态数据。可能重复保存。

Kafka事务通过隔离机制来实现多会话幂等性

kafka事务引入了transactionId 和Epoch，设置transactional.id后，一个transactionId只对应一个pid, 且Server 端会记录最新的 Epoch 值。这样有新的producer初始化时，会向TransactionCoordinator发送InitPIDRequest请求， TransactionCoordinator 已经有了这个 transactionId对应的 meta，会返回之前分配的 PID，并把 Epoch 自增 1 返回，这样当old producer恢复过来请求操作时，将被认为是无效producer抛出异常。     如果没有开启事务，TransactionCoordinator会为新的producer返回new pid，这样就起不到隔离效果，因此无法实现多会话幂等。

private def maybeValidateAppend(producerEpoch: Short, firstSeq: Int, offset: Long): Unit = {	
    validationType match {	
      case ValidationType.None =>	
      case ValidationType.EpochOnly =>	
        checkProducerEpoch(producerEpoch, offset)	
      case ValidationType.Full => //开始事务，执行这个判断	
        checkProducerEpoch(producerEpoch, offset)	
        checkSequence(producerEpoch, firstSeq, offset)	
    }	
}	
private def checkProducerEpoch(producerEpoch: Short, offset: Long): Unit = {	
    if (producerEpoch < updatedEntry.producerEpoch) {	
      throw new ProducerFencedException(s"Producer's epoch at offset $offset is no longer valid in " +	
        s"partition $topicPartition: $producerEpoch (request epoch), ${updatedEntry.producerEpoch} (current epoch)")	
    }	
  }

04 Consumer端幂等性

如上所述，consumer拉取到消息后，把消息交给线程池workers，workers对message的handle可能包含异步操作，又会出现以下情况：

• 先commit，再执行业务逻辑：提交成功，处理失败 。造成丢失

• 先执行业务逻辑，再commit：提交失败，执行成功。造成重复执行

• 先执行业务逻辑，再commit：提交成功，异步执行fail。造成丢失

对此我们常用的方法时，works取到消息后先执行如下code：

if(cache.contain(msgId)){	
  // cache中包含msgId，已经处理过	
        continue;	
}else {	
  lock.lock();	
  cache.put(msgId,timeout);	
  commitSync();	
  lock.unLock();	
}	
// 后续完成所有操作后，删除cache中的msgId，只要msgId存在cache中，就认为已经处理过。Note：需要给cache设置有消息

猜你喜欢

1、看完这篇还不会kafka，我跪榴莲！

2、Kylin 在满帮集团千亿级用户访问行为分析中的应用

3、如何在 Kylin 中优雅地使用 Spark

4、流系统Spark/Flink/Kafka/DataFlow端到端一致性实现对比

过往记忆大数据技术交流群，请添加个人微信：fangzhen0219,备注进群。