The simple answer: supervision trees. And fault-tolerance usually means that the failing process would be restarted. It would not handle stuff like netsplits or node going down though.

Check code of cowboy, ejabberd, MongooseIM, RabbitMQ for examples. There are many factors on decision when to make a new process. Data locality, the pattern of interaction with other processes, performance considerations. Good idea is to have one process per TCP connection, but not one process per each routed message. And be careful with blocking gen_server calls - these could block or fail.