When the user stops a cache without stopping the cache manager on that node, subsequent PrepareCommands sent to that node will return a SuccessfulResponse.
If that node used to the primary owner of the command's modified key, the originator will proceed with the transaction as if it had acquired a lock on that key. It is thus possible for multiple transactions to think they have acquired the key lock at the same time.
On the other hand, in replicated caches is is quite possible that a cache is not running on all the cluster node and yet PrepareCommands are broadcasted to everyone in parallel. So the solution should not involve sending exceptions (which have huge stack traces), and the originator should be able to ignore failures responses from nodes that were not targeted in the first place.