转载请注明原创地址 http://www.cnblogs.com/dongxiao-yang/p/6238029.html
最近需要详细研究下kafka reblance过程中分区计算的算法细节，网上搜了部分说法，感觉比较晦涩且不太易懂，还是自己抠源码比较简便一点。
kafka reblance计算部分代码如下：
class RangeAssignor() extends PartitionAssignor with Logging {
def assign(ctx: AssignmentContext) = {
val valueFactory = (topic: String) => new mutable.HashMap[TopicAndPartition, ConsumerThreadId]
val partitionAssignment =
new Pool[String, mutable.Map[TopicAndPartition, ConsumerThreadId]](Some(valueFactory))
for (topic <- ctx.myTopicThreadIds.keySet) {
val curConsumers = ctx.consumersForTopic(topic)
val curPartitions: Seq[Int] = ctx.partitionsForTopic(topic)
val nPartsPerConsumer = curPartitions.size / curConsumers.size
val nConsumersWithExtraPart = curPartitions.size % curConsumers.size
info("Consumer " + ctx.consumerId + " rebalancing the following partitions: " + curPartitions +
" for topic " + topic + " with consumers: " + curConsumers)
for (consumerThreadId <- curConsumers) {
val myConsumerPosition = curConsumers.indexOf(consumerThreadId)
assert(myConsumerPosition >= 0)
val startPart = nPartsPerConsumer * myConsumerPosition + myConsumerPosition.min(nConsumersWithExtraPart)
val nParts = nPartsPerConsumer + (if (myConsumerPosition + 1 > nConsumersWithExtraPart) 0 else 1)
/**
* Range-partition the sorted partitions to consumers for better locality.
* The first few consumers pick up an extra partition, if any.
*/
if (nParts <= 0)
warn("No broker partitions consumed by consumer thread " + consumerThreadId + " for topic " + topic)
else {
for (i <- startPart until startPart + nParts) {
val partition = curPartitions(i)
info(consumerThreadId + " attempting to claim partition " + partition)
// record the partition ownership decision
val assignmentForConsumer = partitionAssignment.getAndMaybePut(consumerThreadId.consumer)
assignmentForConsumer += (TopicAndPartition(topic, partition) -> consumerThreadId)
}
}
}
}
def getPartitionsForTopics(topics: Seq[String]): mutable.Map[String, Seq[Int]] = {
getPartitionAssignmentForTopics(topics).map { topicAndPartitionMap =>
val topic = topicAndPartitionMap._1
val partitionMap = topicAndPartitionMap._2
debug("partition assignment of /brokers/topics/%s is %s".format(topic, partitionMap))
(topic -> partitionMap.keys.toSeq.sortWith((s,t) => s < t))
}
}
def getConsumersPerTopic(group: String, excludeInternalTopics: Boolean) : mutable.Map[String, List[ConsumerThreadId]] = {
val dirs = new ZKGroupDirs(group)
val consumers = getChildrenParentMayNotExist(dirs.consumerRegistryDir)
val consumersPerTopicMap = new mutable.HashMap[String, List[ConsumerThreadId]]
for (consumer <- consumers) {
val topicCount = TopicCount.constructTopicCount(group, consumer, this, excludeInternalTopics)
for ((topic, consumerThreadIdSet) <- topicCount.getConsumerThreadIdsPerTopic) {
for (consumerThreadId <- consumerThreadIdSet)
consumersPerTopicMap.get(topic) match {
case Some(curConsumers) => consumersPerTopicMap.put(topic, consumerThreadId :: curConsumers)
case _ => consumersPerTopicMap.put(topic, List(consumerThreadId))
}
}
}
for ( (topic, consumerList) <- consumersPerTopicMap )
consumersPerTopicMap.put(topic, consumerList.sortWith((s,t) => s < t))
consumersPerTopicMap
}
计算过程主要由上述高亮代码部分实现，举例说明，一个拥有十个分区的topic，相同group拥有三个consumerid为aaa,ccc,bbb的消费者
1 由后两段代码可知，获取consumerid列表和partition分区列表都是已经排好序的，所以
curConsumers=(aaa,bbb,ccc)
curPartitions=(0,1,2,3,4,5,6,7,8,9)
2
nPartsPerConsumer=10/3 =3
nConsumersWithExtraPart=10%3 =1
3 假设当前客户端id为aaa
myConsumerPosition= curConsumers.indexof(aaa) =0
4 计算分区范围
startPart= 3*0+0.min(1) = 0
nParts = 3+(if (0 + 1 > 1) 0 else 1)=3+1=4
所以aaa对应的分区号为[0,4),即0，1，2，3前面四个分区
同理可得bbb对应myConsumerPosition=1，对应分区4，5，6中间三个分区
ccc对应myConsumerPosition=2，对应7，8，9最后三个分区。