actors

actors

actors is a library that enables the use of the Actors Model in Dart.

It is a thin wrapper around Dart's Isolate (on Flutter and Dart VM) and Web Workers (on the Web - TODO) that makes them much easier to use.

Actor

To start an Actor is very easy. You create a Handler implementing the logic to handle messages within the Actor's Isolate, then create an Actor using it:

class Accumulator with Handler<int, int> {
  int _value;
  
  Accumulator([int initialValue = 0]): _value = initialValue;
  
  int handle(int n) => _value += n;
}

main() async {
  final actor = Actor.create(() => Accumulator(6));
  print(await actor.send(5)); // 11
  await actor.close();
}

If your actor does not maintain internal state, it can also be created from a function, or even a lambda:

int two(int n) => n * 2;

main() async {
  final actor = Actor.of(two);
  print(await actor.send(5)); // 10
  await actor.close();
}

As you can see, an Actor can send a message back to the caller asynchronously.

They can also send more than one message back by returning a Stream:

// A Handler that returns a Stream must use a StreamActor, not an Actor.
class StreamGenerator with Handler<int, Stream<int>> {
  @override
  Stream<int> handle(int message) {
    return Stream.fromIterable(Iterable.generate(message, (i) => i));
  }
}

main() async {
  // Create an StreamActor from a Handler that returns Stream.
  final actor = StreamActor.create(StreamGenerator.new);
  final stream = actor.send(2);
  await for (final item in stream) {
    print(item); // 0, 1
  }
  await actor.close();
}

Actor state

An actor can safely maintain internal state which cannot be reached by any other actors (as it resides in its own Dart Isolate in the DartVM - please notice that this does not hold on the web).

The actor's state can include anything, even Streams and open sockets, for example, except if the Handler is instantiated locally by using the Actor(Handler handler) constructor. For this reason, prefer to use either the Actor.of constructors to wrap functions, or the Actor.create(Handler Function()) constructor that ensures the Handler is not instantiated locally.

Initial state that requires asynchronous calls can be initialized in the Handler's init method. actors guarantees that an Actor will never be called to handle a message before its init method has successfully returned.

For example, an Actor which wraps a HttpServer could be initialized as shown below:

class HttpServerActor with Handler<Message, Answer> {
  late final HttpServer _server;
  final int port;

  HttpServerActor(this.port);

  // this method will only run in the Actor's own Isolate, so we can
  // create non-sendable state.
  @override
  Future<void> init() async {
    // binding a server with "shared: true" means that requests will be handled
    // by multiple Isolates if HttpServerActor is started on many Isolates (see ActorGroup). 
    _server = await HttpServer.bind(InternetAddress.loopbackIPv4, port, shared: true);
    unawaited(_serveRequests());
  }
  
  // ...
}

You can see the full example code at example/stateful_actor_example.dart.

Sending an Actor to another Actor

To send an Actor to another Actor is not possible directly, but you can send its Sendable object, which can be obtained by calling toSendable() (unfortunately, this does not currently work for StreamActor).

See the inter_actor_test test for an example where an Actor's sender function is given to another Actor via its constructor.

This enables a common pattern where many actors are given a reference to another actor which can aggregate their results in one place.

Notice that using a Sendable to send messages has a small performance overhead when compared to a plain Actor. For the big_messages benchmark, Actor messages' round trip is measured at ~21us VS 28us for Sendable (using a Macbook Air from 2019).

Actors themselves perform identically to using plain Isolates directly.

ActorGroup

ActorGroup allows several Actor instances to be grouped together, all based on the same Handler implementation, but executed according to one of the available strategies:

• RoundRobin - send message to a single Actor, alternating which member of the group receives the message.
• MultiHandler - send message to m Actors, wait for at least n successful answers.

RoundRobing is appropriate for cases where messages are CPU intensive to handle and there may be many of them.

MultiHandler is a way to achieve high reliability by duplicating effort, as not all Actors in the group may be healthy at all times. Having a few "backups" doing the same work on each message may be a good idea in case one or more of the expected receivers are likely to fail, as the system will still continue to work without issues as long as n actors remain healthy... Also, by sending the same message to several actors, the message might be received in different locations, making it much harder for it to be lost.

// create a group of 4 actors
final group = ActorGroup(Two(), size: 4);
print(await group.send(5)); // 10
group.close();

Messenger

The Messenger mixin is implemented by Actor, ActorGroup, and also LocalMessenger, which runs its Handler in the local Isolate.

Messenger<int, int> messenger;

// a Messenger can be local
messenger = LocalMessenger(Two());
print(await messenger.send(2)); // 4

// or it can be an Actor
messenger = Actor(Two());
print(await messenger.send(3)); // 6
messenger.close();

// or an ActorGroup
messenger = ActorGroup(Two(), size: 2);
print(await messenger.send(4)); // 8
print(await messenger.send(5)); // 10
messenger.close();

This makes it possible to write code that works the same whether the message is handled locally or in another Isolate.

More examples

• basic_example.dart (the basics of actors)
• actors_example.dart (using actors, groups, streams, local)
• example-projects/word_count (utility to count words in files)