I just upgraded a three node cluster from 2.2.0 to 2.3.3 using the rolling restart method advised. The first two nodes were no problem, but after upgrading the third and restarting Elasticsearch the cluster turned red.
The node joins the cluster with name that's 22 apparently random characters and it says it has no shards.
Shutdown and restarted Elasticsearch on all nodes. All have rejoined the cluster but with random names. Kibi reports that the Elasticsearch version is still 2.2.0 rather than 2.3.3.