[
https://issues.apache.org/jira/browse/KAFKA-1503?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14039417#comment-14039417
]
Jianwen Wang commented on KAFKA-1503:
-------------------------------------
Hi Guozhang,
That is not what I experienced in my real cluster environment:
Given a cluster contains three brokers, when all brokers are live and
running, and there is a topic with 10 partitions:
I just made a query on my cluster and got the topic info as below right now:
root@kafka-1:~/kafka-0.8.1.1-src# bin/kafka-topics.sh --zookeeper
localhost:2181 --topic ruby-p10 --describe
Topic:ruby-p10 PartitionCount:10 ReplicationFactor:3 Configs:
Topic: ruby-p10 Partition: 0 Leader: 2 Replicas: 1,2,3 Isr:
2,1,3
Topic: ruby-p10 Partition: 1 Leader: 2 Replicas: 2,3,1 Isr:
2,1,3
Topic: ruby-p10 Partition: 2 Leader: 3 Replicas: 3,1,2 Isr:
2,1,3
Topic: ruby-p10 Partition: 3 Leader: 1 Replicas: 1,3,2 Isr:
2,1,3
Topic: ruby-p10 Partition: 4 Leader: 2 Replicas: 2,1,3 Isr:
2,1,3
Topic: ruby-p10 Partition: 5 Leader: 3 Replicas: 3,2,1 Isr:
2,1,3
Topic: ruby-p10 Partition: 6 Leader: 2 Replicas: 1,2,3 Isr:
2,1,3
Topic: ruby-p10 Partition: 7 Leader: 2 Replicas: 2,3,1 Isr:
2,1,3
Topic: ruby-p10 Partition: 8 Leader: 3 Replicas: 3,1,2 Isr:
2,1,3
Topic: ruby-p10 Partition: 9 Leader: 1 Replicas: 1,3,2 Isr:
2,1,3
As you can see the ISR is not evenly distributed, and since current codes
always pick the first one. So if broker 1 is down, partition hosted by broker 1
will be changed to be broker 2 instead of evenly distributed to broker 2 and
broker 3.
As times goes on, all the partition will be only hosted on one broker.
> all partitions are using same broker as their leader after broker is down
> -------------------------------------------------------------------------
>
> Key: KAFKA-1503
> URL: https://issues.apache.org/jira/browse/KAFKA-1503
> Project: Kafka
> Issue Type: Bug
> Components: controller
> Affects Versions: 0.8.0, 0.8.1.1
> Environment: 0.8.1.1
> Reporter: Jianwen Wang
> Assignee: Neha Narkhede
>
> The current leader selection always pick the first live broker in ISR when
> the current leader broker is down. Since the list of liveBrokerInIsr is not
> evenly distributed. As time goes on, all the partitions will use only one
> broker as its leader.
> I figured out a fix which is to use the first live broker in replica list
> which is also in ISR list. Since the liveAssignedReplicas is evenly
> distributed across brokers, all the partitions will be evenly distributed in
> the live brokers in ISR.
> The fix is:
> kafka-0.8.1.1-src/core/src/main/scala/kafka/controller/PartitionLeaderSelector.scala
> 71 71
> case false =>
> 72
> - val newLeader = liveBrokersInIsr.head
> 72
> + val liveReplicasInIsr = liveAssignedReplicas.filter(r =>
> liveBrokersInIsr.contains(r))
> 73
> + val newLeader = liveReplicasInIsr.head
--
This message was sent by Atlassian JIRA
(v6.2#6252)
