zachallaun
commented
1 year ago Ok, duh... Parents know their invoked children. When the interpreter stops, they need to clean up by sending an exit signal to each. There are still cases (I think with :brutal_kill?) where children can be "left behind", but I think this is sufficient for the moment.
311e880a4f543e7de4ece347ada77903eee85907
Protean's supervision strategy at the moment: The library asks that the user add
Proteanto their application supervision tree. That starts aDynamicSupervisorthat is then used for any invoked processes (regardless of "depth"). While it's possible to pass a different supervisor when starting a machine, any processes resulting from:invokeare started using that same supervisor.Whether this is the correct implementation depends on how we want to handle restarts & process dependencies. The problem is that the dependencies are cyclical -- invoked children can send messages to their parents. For instance: Let's say we have a parent and child process communicating with each other. While this is probably a bad design choice, these machines may be coordinating with each other such that one's in a state where it is expecting a certain event. Then, let's say the child dies -- if it's restarted, it may now be in a state that it will never send the event the parent expects.
I believe SCXML semantics account for this. An invoked process shouldn't ever be restarted -- the crash should be visible to the parent machine through the
errortransition, which the parent could use to e.g. self-transition and restart the invoked process itself.So: restarts bad, child processes should be started with
restart: :temporary.But what if that crashed child process had invoked processes itself? They need to be shut down. but all of those processes were started under the same top-level dynamic supervisor. Machine servers are given their parent's pid when they start. It could monitor that pid so that we receive a
:DOWNmessage. If the parent crashes, the child can terminate.I'm not yet sure what to do about misc processes we don't control. I'd like any OTP process to be invokable from within a machine, but I'd like to basically one-way-link children to their parents, so that the children are always terminated if the parent crashes.