I improved the docs to say:

The Supervisor module was designed to handle mostly static children
that are started in the given order when the supervisor starts. A
DynamicSupervisor starts with no children. Instead, children are
started on demand via start_child/2 and there is no ordering between
children. This allows the DynamicSupervisor to hold millions of
children by using efficient data structures and to execute certain
options, such as shutting down, concurrently.

Permanent == started on app startup and staying there until the app shuts down.

Thanks to this distinction, DynamicSupervisor allows you to start and stop children processes at any time. You also are not strictly mandated to stop them; they will be stopped when the supervisor itself is stopped (although it’s definitely a good idea to stop children proactively after they’ve done their job IMO).

To me, using DynamicSupervisor is appropriate in two scenarios:

1. When you would like to spawn quite a lot of (say, at least 200,000+) processes because I’d wager then a performance (or memory) impact would start becoming noticeable.
2. When you need to create processes in response to – or because of the specifics of – external systems whose resources you have to monitor e.g. a web crawling service to which you give a new domain to crawl. It would need to spawn processes to enumerate URLs and then maintain a small pool of workers to do the crawling (and apply rate limits if applicable).

I’d say especially the latter option is pretty valid; there are many real-world examples where you cannot neatly plan a supervision from the start because things do change at runtime. DynamicSupervisor is ideal for this.

I think the DynamicSupervisor is only a special case of Supervisor that handle only dynamic children. In Erlang there is only supervisor. It is a good habit to separate the static children and the dynamic children in different branches of the supervision tree.