A shutdown of the cluster (bin/stop-all.sh) leaves the ephemeral nodes (of master and nodes) behind. On newstart, the ephemera's are still there (but not connected to any owner). The master and all nodes delete unconnected ephemeral's with their address and create new one. This leed's to a lot of node connected / node disconnected events in the master log on startup and it looks like there is something wrong. Especially on large cluster the startup log looks very confusing. But redeployments shouldn't happen, thanks to the safe mode. Think there are several solutions. I is possible to re-own unconnected ephemerals if seesion id and password are known. So if f.e. a node would persist it zookeeper sessionId and password it could reown it ephemeral on startup if still existent. What is already there is a shutdown hook on node side. This hook tries, among other stuff, to delete the node's ephemeral. The problem with this is the stop-all script, since it stops the master with the zk process first and then the node. So if the hook executes there is no zk system to communicate anymore. I think it would be a good move to decouple master and zookeeper process and stop the zookeeper process a last in the stop-all.sh script. This would also help people who may have their own zookeeper system already running!