Is it possible/advisable to have a different supervision strategy for different children of an Akka 2 actor?

Question

I'm playing with Akka and I have a design in which a supervisor actor has a child playing role A and several children playing role B. I want to define a supervision policy such as A failures are escalated (terminating the supervisor) and B ones produces the individual actors to be restarted.

Is this possible? Is it advisable?

Equally

Answer 1

Yes, by overriding supervisorStrategy. For example, from the docs:

import akka.actor.OneForOneStrategy
import akka.actor.SupervisorStrategy._
import scala.concurrent.duration._

override val supervisorStrategy =
  OneForOneStrategy(maxNrOfRetries = 10, withinTimeRange = 1 minute) {
    case _: ArithmeticException      => Resume
    case _: NullPointerException     => Restart
    case _: IllegalArgumentException => Stop
    case _: Exception                => Escalate
  }

Then reading the note:

If the strategy is declared inside the supervising actor (as opposed to within a companion object) its decider has access to all internal state of the actor in a thread-safe fashion, including obtaining a reference to the currently failed child (available as the sender of the failure message).

So rather than matching on the type of the exception, you would match on the sender (just typing here; it might not compile):

override val supervisorStrategy =
  OneForOneStrategy(maxNrOfRetries = 10, withinTimeRange = 1 minute) {
    if (sender == SomeSpecialChild) Escalate
    else Restart
  }

There's nothing wrong with this, but if you have groups of children playing different roles, you should probably follow the first rule of Akka (well, my first rule of Akka): when in doubt, you probably want another actor. Create one actor to supervise role A and another to supervise role B. You'll still need to override supervisorStrategy, but the test will be simpler. And it'll be easier to manage any other special differences between role A and B.

Straitlaced

I agree with the idea of using separate supervisors for A and B actors. It just seems to be the more manageable approach. I don't think mixing together supervisor concepts is a good idea. You can have one main supervisor at the top of this hierarchy and then one actor for an A supervisor and one for a B. If you want failures in A to terminate all A actors and all B actors then just escalate up to the single supervisor for this hierarchy and when that stops, it will stop the entire tree. – Collencollenchyma