Predicates data is going to be split up between the 3 dgraph instances. When you shut one of them down, there won’t be enough data to complete queries anymore.
You can get around this by using replication. E.g. use a single dgraphzero, and 3 dgraph instances. When you start dgraphzero, set --replicas=3. That way, each instance will serve the same set of predicates and you will be able to survive one of the nodes going down.