When the coordinator dies or otherwise leaves the cluster, the new coordinator recovers the cluster state and handles any new changes.

However, if there was already a rebalance operation in progress, the new coordinator ignores the rebalance confirmations from the other members and doesn't install a balanced CH until there is another join/leave/etc. event for that cache.

Normally this doesn't appear in our tests because the state transfer is quick and so the old coordinator is able to finish the rebalance caused by its own leave before shutting down. I did see it however in Query's MultiNodeDistributedTest.