KafkaController 分区Rebalance平衡机制

private def checkAndTriggerPartitionRebalance(): Unit = {
  if (isActive()) {
    trace("checking need to trigger partition rebalance")
    // 获取(存活的broker,AR副本集) => (2,Map([message,0]-> List(2, 0), [hadoop,0] -> List(2, 1)))
    var preferredReplicasForTopicsByBrokers: Map[Int, Map[TopicAndPartition, Seq[Int]]] = null
    inLock(controllerContext.controllerLock) {
      preferredReplicasForTopicsByBrokers=
        controllerContext.partitionReplicaAssignment.filterNot(p => deleteTopicManager.isTopicQueuedUpForDeletion(p._1.topic)).groupBy {
          case(topicAndPartition, assignedReplicas) => assignedReplicas.head
        }
    }
    debug("preferred replicas by broker " + preferredReplicasForTopicsByBrokers)
    // 过滤每一个存活的broker,检查是否需要一个preferredreplica 选举被触发
    preferredReplicasForTopicsByBrokers.foreach {
      case(leaderBroker, topicAndPartitionsForBroker) => {
        var imbalanceRatio: Double = 0
        var topicsNotInPreferredReplica: Map[TopicAndPartition, Seq[Int]] = null
        inLock(controllerContext.controllerLock) {
          // 我们知道，正常情况下，broker的AR副本集第一个副本(preferred replica )就是leader
          // 如果leader不是preferred replica，比如 Leader : 0 ISR[2,0]
          // 我们需要过滤出这种topicPartition，然后我们好进行在平衡
          topicsNotInPreferredReplica=
            topicAndPartitionsForBroker.filter {
              case(topicPartition, replicas) => {
                controllerContext.partitionLeadershipInfo.contains(topicPartition) &&
                controllerContext.partitionLeadershipInfo(topicPartition).leaderAndIsr.leader != leaderBroker
              }
            }
          debug("topics not in preferred replica " + topicsNotInPreferredReplica)
          // broker 的AR副本数量
          val totalTopicPartitionsForBroker = topicAndPartitionsForBroker.size
          // 过滤出的leader不是AR副本集的preferred replica的数量
          val totalTopicPartitionsNotLedByBroker = topicsNotInPreferredReplica.size
          // 计算(过滤出的leader不是AR副本集的preferred replica的数量)/(broker 的AR副本数量)不平衡比例
          imbalanceRatio= totalTopicPartitionsNotLedByBroker.toDouble / totalTopicPartitionsForBroker
          trace("leader imbalance ratio for broker %d is %f".format(leaderBroker, imbalanceRatio))
        }
        // 如果比例大于我们配置的leader.imbalance.per.broker.percentage参数，比如50%，就触发这个topic partitions的再平衡操作
        if (imbalanceRatio > (config.leaderImbalancePerBrokerPercentage.toDouble / 100)) {
          topicsNotInPreferredReplica.foreach {
            case(topicPartition, replicas) => {
              inLock(controllerContext.controllerLock) {
                // 首先确保broker存活，而且没有分区正在重新分配或者没有进行preferredreplica 选举，且没有分区将被删除
                if (controllerContext.liveBrokerIds.contains(leaderBroker) &&
                    controllerContext.partitionsBeingReassigned.isEmpty &&
                    controllerContext.partitionsUndergoingPreferredReplicaElection.isEmpty &&
                    !deleteTopicManager.isTopicQueuedUpForDeletion(topicPartition.topic) &&
                    controllerContext.allTopics.contains(topicPartition.topic)) {
                  // 然后真正触发Prederred Replica选举操作
                  onPreferredReplicaElection(Set(topicPartition), true)
                }
              }
            }
          }
        }
      }
    }
  }
}