[
https://issues.apache.org/jira/browse/KAFKA-6128?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16220293#comment-16220293
]
Alastair Munro edited comment on KAFKA-6128 at 10/26/17 10:45 AM:
------------------------------------------------------------------
It does it with 0.11.0.0. It would seem the broker id's are not being
replicated in zookeeper. So kafka connects to zookeeper:2181 which is a load
balancer. Then when it terminates, it removes the broker-id on the zookeeper it
connected to, but the change is not replicated to the other two nodes. As
zookeeper may scale up and down, we use zookeeper:2181 rather than a list of
hosts. Testing replication in zookeeper (eg creating /test using zkCli.sh),
replication works fine. So why doesn't the broker changes replicate?
I scaled this down to have only brokers 0 and 1, but the change is only made to
the node the departing kafka connected to:
{code}
oc rsh zoo-0 bin/zkCli.sh ls /brokers/ids
[0, 1]
oc rsh zoo-1 bin/zkCli.sh ls /brokers/ids
[0, 1, 2]
{code}
was (Author: amunro):
It does it with 0.11.0.0. It would seem the broker id's are not being
replicated in zookeeper. So kafka connects to zookeeper:2181 which is a load
balancer. Then when it terminates, it removes the broker-id on the zookeeper it
connected to, but the change is not replicated to the other two nodes. As
zookeeper may scale up and down, we use zookeeper:2181 rather than a list of
hosts. Testing replication in zookeeper (eg creating /test using zkCli.sh),
replication works fine. So why doesn't the broker changes replicate?
I scaled this down to have only brokers 0 and 1, but some of the zoo nodes
don't see the change:
{code}
oc rsh zoo-0 bin/zkCli.sh ls /brokers/ids
[0, 1]
oc rsh zoo-1 bin/zkCli.sh ls /brokers/ids
[0, 1, 2]
{code}
> Shutdown script does not do a clean shutdown
> --------------------------------------------
>
> Key: KAFKA-6128
> URL: https://issues.apache.org/jira/browse/KAFKA-6128
> Project: Kafka
> Issue Type: Bug
> Affects Versions: 0.11.0.1
> Reporter: Alastair Munro
> Priority: Minor
>
> Shutdown script (sending term signal) does not do a clean shutdown.
> We are running kafka in kubernetes/openshift 0.11.0.0. The statefulset kafka
> runs the shutdown script prior to stopping the pod kafka is running on:
> {code}
> lifecycle:
> preStop:
> exec:
> command:
> - ./bin/kafka-server-stop.sh
> {code}
> This worked perfectly in 0.11.0.0 but doesn't in 0.11.0.1. Also we see the
> same behaviour if we send a TERM signal to the kafka process (same as the
> shutdown script).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
