To add to the comment about subscribing to vintage_net events, quite a few libraries choose to repeatedly try to connect whenever they’re disconnected from a server. For example, MQTT client libraries like Tortoise and the Phoenix channel client libraries like PhoenixClient and phoenix_gen_socket_client take this approach.

Losing a connection to a server isn’t a Nerves-specific issue. One advantage to polling is that your code doesn’t carry a dependency on vintage_net and therefore can be developed and run outside of a Nerves device.

The disadvantage of polling is that it can take your code a long time to connect to the server after it boots. A way around this is to register for interface up notifications from vintage_net and have those events “kick” your code to retry connecting to the server immediately.

Regarding your question about using a dynamic supervisor, I feel like that there might be some nuance in your use case that pushes you to one solution or another. You’ll have a GenServer that tries to connect. Once it connects, though, it could either hand off connection session to another GenServer or manage it itself. Some people like the state machine aspect of doing this in one place, but there’s a point where I think the code gets too complicated.