FLUME-3027 has been backported to CDH5.11.0 and above, so if you are able to upgrade, it would prevent the issue of offsets bouncing back and forward. One thing you may want to consider, if you are getting rebalances, it may be because it is taking too long to deliver by your sink, before polling kafka again. You may want to consider lowering your sink batch size in order to deliver and ack the messages in a timely fashion. Additionally, if you upgrade to CDH5.14 or higher, the flume kafka client is 0.10.2, and you would be able to set max.poll.records to match the batchSize you are using for the flume sink. Additionally, you could increase the max.poll.interval.ms, which is decoupled from the session.timeout.ms in 0.10.0 and above. This would prevent the rebalancing from occurring since the client would still heartbeat without having to do a poll to pull more records before session.timeout.ms expires. -pd ... View more