Does the child have to detect this and die (self-supervision)?
That’s what linking does, the linking is bidirectional, so if the link is severed between the processes, both processes receive an exit signal. What makes supervisors special is that they trap_exit. In this scenario it will mean that the child will terminate if the supervisor is no longer reachable.The supervisor will also receive an exit signal if the child is no longer reachable, but instead of terminating it will attempt to start a new child to replace the old one. Now if that fails many times, the supervisor itself will fail.
This of course doesn’t have to span networks. It could be 2 separate nodes on the same machine, or separate machines connected directly with an Ethernet/serial cable, it really all depends on the topology of the system.
All in all, the most important is to define/understand the failure modes and possibilities of your system. Is there a network? Is the network a possible problem? Is hardware failure a possibility you want to handle? Can someone yank a cable?
Once understood, you can build the right mechanism to handle the failures. No system can handle every single possible failure.
