首页 > 代码库 > 从外部重置一个运行中consumer group的消费进度

从外部重置一个运行中consumer group的消费进度

2024-11-03 18:09:02   209人阅读

对于0.10.1以上版本的kafka, 如何从外部重置一个运行中的consumer group的进度呢？比如有一个控制台，可以主动重置任意消费组的消费进度重置到12小时之前。

需要这么几个步骤：

1. 加入这个group

2. 踢掉所有其它group memeber

3. try assign all TopicPartition to this client

4. commit offsets

5. leave group

 

其中第二步是为了让自己当上leader，当然有可能不需要踢掉其它所有成员就能当上leader(因为谁能当leader实际上是按hashmap的迭代次序来的)。

当上consumer group的leader以后，需要把所有partition assign给自己，这个需要一个特殊的PartitionAssignor。由于这个assignor的协议跟其它consumer group协议不同(但是也可以搞一个表面上协议相同，实际上逻辑不同的assignor)，而cooridnator会阻止与当前leader使用的协议不同的成员加入，所以还是需要踢掉其它成员。

public class ExclusiveAssignor extends AbstractPartitionAssignor {

    public interface Callback {
        void onSuccess();
    }


    private static Logger LOGGER = LoggerFactory.getLogger(ExclusiveAssignor.class);

    public static String NAME = "exclusive";


    private String leaderId = null;
    private Callback callback = null;

    public void setLeaderId(String leaderId) {
        this.leaderId = leaderId;
    }
    public void setCallBack(Callback callBack){this.callback = callBack;}


    @Override
    public String name() {
        return NAME;
    }

    private Map<String, List<String>> consumersPerTopic(Map<String, List<String>> consumerMetadata) {
        Map<String, List<String>> res = new HashMap<>();
        for (Map.Entry<String, List<String>> subscriptionEntry : consumerMetadata.entrySet()) {
            String consumerId = subscriptionEntry.getKey();
            for (String topic : subscriptionEntry.getValue())
                put(res, topic, consumerId);
        }
        return res;
    }


    @Override
    public Map<String, List<TopicPartition>> assign(Map<String, Integer> partitionsPerTopic,
                                                    Map<String, List<String>> subscriptions) {
        LOGGER.info("perform exclusive assign");
        if(leaderId == null)
            throw new IllegalArgumentException("leaderId should already been set before assign is called");
        if(callback == null)
            throw new IllegalArgumentException("callback should already been set before assign is called");

        List<TopicPartition> allPartitions = new ArrayList<TopicPartition>();
        partitionsPerTopic.forEach((topic, partitionNumber) -> {
            for(int i=0; i < partitionNumber; i++)
                allPartitions.add(new TopicPartition(topic, i));
        });
        Map<String, List<TopicPartition>> assignment = new HashMap<>();
        for (String memberId : subscriptions.keySet()) {
            assignment.put(memberId, new ArrayList<TopicPartition>());
            if(memberId.equals(leaderId)){
                assignment.get(memberId).addAll(allPartitions);
            }
        }
        callback.onSuccess();
        return assignment;
    }

}

这个assignor需要知道leaderId是哪个，而leaderId可以在KafkaConsumer的

 protected Map<String, ByteBuffer> performAssignment(String leaderId,
                                                        String assignmentStrategy,
                                                        Map<String, ByteBuffer> allSubscriptions)

中获取，所以还需要修改一下KafkaConsumer的代码，以确保这个KafkaConsumer的poll并不实际拉取消息，而只是执行commit。

 

驱逐其它member，可以使用AdminClient完成

  def forceLeave(coordinator: Node, memberId: String, groupId: String) = {
    logger.info(s"forcing group member: $memberId to leave group: $groupId ")
    send(coordinator, ApiKeys.LEAVE_GROUP, new LeaveGroupRequest(groupId, memberId))
  }

 

最终的逻辑就是

  private def forceCommit(consumer: SimpleKafkaConsumer[_, _], groupId: String, topics: Seq[String], maxRetries: Int, toCommit: Map[TopicPartition, OffsetAndMetadata], coordinatorOpt: Option[Node] = None) = {
    consumer.subscribe(JavaConversions.seqAsJavaList(topics))
    val assignedAll = new AtomicBoolean(false)
    consumer.setExclusiveAssignorCallback(new Callback {
      override def onSuccess(): Unit = assignedAll.set(true)
    })
    var currentRetries = 0
    val coordinatorNode = coordinatorOpt.getOrElse(adminClient.findCoordinator(groupId))
    while (!assignedAll.get() && currentRetries < maxRetries) {
      logger.info(s"trying to reset offset for $groupId, retry count $currentRetries  ....")
      clearCurrentMembers(coordinatorNode, groupId, Some(ConsumerGroupManager.magicConsumerId))
      consumer.poll(5000)
      printCurrentAssignment(consumer)
      currentRetries = currentRetries + 1
    }
    if (currentRetries >= maxRetries)
      throw new RuntimeException(s"retry exhausted when getting leadership of $groupId")
    val javaOffsetToCommit = JavaConversions.mapAsJavaMap(toCommit)
    consumer.commitSync(javaOffsetToCommit)
    logger.info(s"successfully committed offset for $groupId: $toCommit")
    consumer.unsubscribe()
  }

 

  def forceReset(offsetLookupActor: ActorRef, groupId: String, ts: Long, maxRetries: Int)(implicit executionContext: ExecutionContext): Boolean = {
    logger.info(s"resetting offset for $groupId to $ts")
    val groupSummary = adminClient.describeConsumerGroup(groupId)
    val topics = groupSummary.subscribedTopics
    if (topics.isEmpty)
      throw new IllegalStateException(s"group $groupId currently subscribed no topic")
    val offsetToCommit = getOffsetsBehindTs(offsetLookupActor, topics, ts, 10000)
    val consumer = createConsumer(groupId)
    try {
      forceCommit(consumer, groupId, topics, maxRetries, offsetToCommit)
      true
    } finally {
      consumer.close()
    }
  }

 

需要注意的是，发送LeaveGroupRequest可能会使得某些成员到broker的连接断掉，发生这种情况的原因是：当一个consumer发送JoinGroupRequest以后，外部的client再发送一个LeaveGroupRequest把这个consumer踢掉，会使得它个consumer无法收到JoinGroupResponse，从而使得NetworkClient以为连接挂掉。不过client以后会重新连接。而且，在外部client踢掉其它成员并且重新commit offset的过程中，其它consumer不一定有机会加入到group中，因而可能不受这个问题的影响。

 

从外部重置一个运行中consumer group的消费进度


联系
我们