部分内容参考自:http://blog.csdn.net/lizhitao/article/details/41778193
下面以一个Kafka集群中4个Broker举例,创建1个topic包含4个Partition,2 Replication;数据Producer流动如图所示:
(1)
pic
(2)当集群中新增2节点,Partition增加到6个时分布情况如下:
副本分配逻辑规则如下:
在Kafka集群中,每个Broker都有均等分配Partition的Leader机会。
上述图Broker Partition中,箭头指向为副本,以Partition-0为例:broker1中parition-0为Leader,Broker2中Partition-0为副本。
上述图种每个Broker(按照BrokerId有序)依次分配主Partition,下一个Broker为副本,如此循环迭代分配,多副本都遵循此规则。
副本分配算法如下:
将所有N Broker和待分配的i个Partition排序.
将第i个Partition分配到第(i mod n)个Broker上.
将第i个Partition的第j个副本分配到第((i + j) mod n)个Broker上.
事实上以上的算法是有误的,因为很明显,每个topic的分区0都会被分配在broker 0上,第1个分区都分配到broker 1上,直到partition的id超过broker的数据才开始从头开始重复,这样会导致前面几台机器的压力比后面的机器压力更大。
因此,kafka是先随机挑选一个broker放置分区0,然后再按顺序放置其它分区。如下图的情况:
Topic:ljh_test3 PartitionCount:10 ReplicationFactor:2 Configs:
Topic: ljh_test3 Partition: 0 Leader: 5 Replicas: 5,6 Isr: 5,6
Topic: ljh_test3 Partition: 1 Leader: 6 Replicas: 6,7 Isr: 6,7
Topic: ljh_test3 Partition: 2 Leader: 7 Replicas: 7,2 Isr: 7,2
Topic: ljh_test3 Partition: 3 Leader: 2 Replicas: 2,3 Isr: 2,3
Topic: ljh_test3 Partition: 4 Leader: 3 Replicas: 3,4 Isr: 3,4
Topic: ljh_test3 Partition: 5 Leader: 4 Replicas: 4,5 Isr: 4,5
Topic: ljh_test3 Partition: 6 Leader: 5 Replicas: 5,7 Isr: 5,7
Topic: ljh_test3 Partition: 7 Leader: 6 Replicas: 6,2 Isr: 6,2
Topic: ljh_test3 Partition: 8 Leader: 7 Replicas: 7,3 Isr: 7,3
Topic: ljh_test3 Partition: 9 Leader: 2 Replicas: 2,4 Isr: 2,4
这里分区0放到了broker5中,分区1–broker6,分区2—broker7….
再看一个例子:
Topic:ljh_test2 PartitionCount:10 ReplicationFactor:2 Configs:
Topic: ljh_test2 Partition: 0 Leader: 2 Replicas: 2,7 Isr: 2,7
Topic: ljh_test2 Partition: 1 Leader: 3 Replicas: 3,2 Isr: 3,2
Topic: ljh_test2 Partition: 2 Leader: 4 Replicas: 4,3 Isr: 4,3
Topic: ljh_test2 Partition: 3 Leader: 5 Replicas: 5,4 Isr: 5,4
Topic: ljh_test2 Partition: 4 Leader: 6 Replicas: 6,5 Isr: 6,5
Topic: ljh_test2 Partition: 5 Leader: 7 Replicas: 7,6 Isr: 7,6
Topic: ljh_test2 Partition: 6 Leader: 2 Replicas: 2,3 Isr: 2,3
Topic: ljh_test2 Partition: 7 Leader: 3 Replicas: 3,4 Isr: 3,4
Topic: ljh_test2 Partition: 8 Leader: 4 Replicas: 4,5 Isr: 4,5
Topic: ljh_test2 Partition: 9 Leader: 5 Replicas: 5,6 Isr: 5,6
版权声明:本文为博主原创文章,转载请注明来自 http://www.lujinhong.com http://blog.csdn.net/lujinhong2/