Varun Vijayaraghavan 2012-11-14, 16:52
Neha Narkhede 2012-11-14, 17:42
-Re: On SIGTERM and SIGKILL consumer ids in zookeeper do not always get cleared
BTW, this is a feature of the ephemeral nodes in ZooKeeper, not a feature of Kafka. If Kafka doesn't close the ZK connection explicitly, ZK waits for the session timeout before deleting ephemeral nodes.
On Nov 14, 2012, at 12:42 PM, Neha Narkhede wrote:
> Since you kill the consumers, it takes roughly upto
> zookeeper.sessiontimeout.ms time for zookeeper to detect that the
> consumer is gone.
> If you want to restart the consumers immediately, try shutting them
> down cleanly. (kill -15 not kill -9)
> On Wed, Nov 14, 2012 at 8:52 AM, Varun Vijayaraghavan
> <[EMAIL PROTECTED]> wrote:
>> I am a zookeeper based consumer using the brod library (which is a python
>> library for Kafka). I stop and start consumers using a simple shell script.
>> I have noticed that when I kill and immediately start running ~50 consumers
>> in a topic, some of the consumer ids in zookeeper do not get cleared. This
>> messes up rebalancing, and partitions get assigned to zombie consumers.
>> Right now, I have taken the hacky approach, and wait for ~15 seconds after
>> killing the existing consumers, before starting them.
>> Have you seen this issue, and have you implemented anything to handle this
>> in the JVM library?
>> Additional info:
>> I am using kafka 0.7.2 and zookeeper 3.4.4
>> - varun :)