have a 2 broker cluster test setup for 0.9. now i had a strange behavior the second time. topics.sh --describe told be all fine and in service. in zookeeper there was no broker in the broker list. i was not able to produce. i restarted the kafka service then it worked again.
some point in the night the brokers started to do crazy offset things like the following infinite times.

[2016-01-13 23:00:03,294] INFO Partition [__consumer_offsets,46] on broker 1: Shrinking ISR for partition [__consumer_offsets,46] from 1,2 to 1 (kafka.cluster.Partition)
[2016-01-13 23:00:03,304] INFO Partition [__consumer_offsets,46] on broker 1: Cached zkVersion [73] not equal to that in zookeeper, skip updating ISR (kafka.cluster.Partition)
[2016-01-13 23:00:03,304] INFO Partition [__consumer_offsets,21] on broker 1: Shrinking ISR for partition [__consumer_offsets,21] from 1,2 to 1 (kafka.cluster.Partition)
[2016-01-13 23:00:03,321] INFO Partition [__consumer_offsets,21] on broker 1: Cached zkVersion [72] not equal to that in zookeeper, skip updating ISR (kafka.cluster.Partition)
[2016-01-13 23:00:03,321] INFO Partition [__consumer_offsets,0] on broker 1: Shrinking ISR for partition [__consumer_offsets,0] from 1,2 to 1 (kafka.cluster.Partition)
[2016-01-13 23:00:03,331] INFO Partition [__consumer_offsets,0] on broker 1: Cached zkVersion [64] not equal to that in zookeeper, skip updating ISR (kafka.cluster.Partition)