With multiple application instances, you would generally not use group management, and manually assign partitions. Or, as mentioned in the documentation, each instance must have a different group.id. In either case, there will be no rebalances.

When you configure with a single reply TopicPartitionOffset, you can use the same reply topic for multiple templates, as long as each instance listens on a different partition. When configuring with a single reply topic, each instance must use a different group.id. In this case, all instances receive each reply, but only the instance that sent the request finds the correlation ID. This may be useful for auto-scaling, but with the overhead of additional network traffic and the small cost of discarding each unwanted reply. When you use this setting, we recommend that you set the template’s sharedReplyTopic to true, which reduces the logging level of unexpected replies to DEBUG instead of the default ERROR.

... and cherry-picked to 2.9.x & 2.8.x