Writing complex actors

[fvasco]

I need to create a complex actor using some private methods to update the internal state, but I cannot find any useful base class to implement ActorJob.

What is the best practice in such case?

[elizarov]

I'd do something along these lines:

class MyActor {
    // your private state here
    suspend fun onReceive(msg: MyMsg) {
        // ... your code here ...
    }
}

fun myActorJob(): ActorJob<MyMsg> = actor(CommonPool) {
    with(MyActor()) {
        for (msg in channel) onReceive(msg)
    }
}

You can also pass actor's channel to MyActor constructor if it needs access to it.

[fvasco]

Hi @elizarov,
hi solved this issue in the same way, but an helper class looks like a miss of abstraction.

I want propose to introduce a dedicated interface for actor, something like:

interface Actor : ActorScope {
    fun invoke() : Unit
}

and the new function

fun launch(context: CoroutineContext, ..., actorBuild: (CoroutineContext) -> Actor) : ActorJob

I want to propose a draft implementation in the next week, what do you think about it?

[elizarov]

The key design consideration here is how a typical actor's code is going to look like. It is not clear in your example. The goal should be to minimize the amount of boilerplate that users of this library code are forced to write. I was thinking about something along the following lines:

abstract class Actor<T>(scope: ActorScope<T>): ActorScope<T> by scope {
    abstract suspend fun onReceive(msg: T)
}

An abstract class lets us share some common state among all actor implementations (in this case, the fact that ActorScope is provided to all the functions defined inside an actor impl). It also makes sense to share "for messages" loop, as every typical actor is going to have one, and leave only onReceive for override.

With this bases class, the specific actor implementation looks like this:

class MyActor(scope: ActorScope<MyMsg>): Actor<MyMsg>(scope) {
    // your private state here
    override suspend fun onReceive(msg: MyMsg) {
        // ... your code here ...
    }
}

There is still some boiler-plate (MyMsg and scope are mentioned twice), but it is not clear how to further reduce it.

Now, you can define a generic actorOf function to start any kind of actor:

fun <T> actorOf(
    context: CoroutineContext,
    init: (ActorScope<T>) -> Actor<T>
): ActorJob<T> = actor<T>(context) {
        val instance = init(this@actor)
        for (msg in channel) instance.onReceive(msg)
    }

Using this function you can start your custom actor with a natural-looking and easy-to-read invocation actorOf(CommonPool, ::MyActor) with ability to further customize construction of your actor impl by providing an explicit lambda for an init block.

[fvasco]

onReceive method is comfortable to use, but I want to expose you my concerns.

The actor function uses the iterable pattern, instead the Actor class use visitor pattern. Personally I consider the iterator pattern a better choice becouse it is possible to detect actor termination (end of iteration).

Using the explicit itaration in Actor makes it cut-and-paste compatible with the actor function.

I suggest to mantain Actor as public interface and to provide a comfortable implementation

public abstract class ActorBase<E>(actorScope: ActorScope<E>) : Actor<E>, ActorScope<E> by actorScope {

    override suspend fun invoke() {
        for (message in channel)
            onReceive(message)
    }

    public abstract suspend fun onReceive(message: E)
}

Finally actorOf misses of some parameters, we should keep this one, an extended version:

public fun <E> actorOf(
        context: CoroutineContext,
        capacity: Int = 0,
        start: CoroutineStart = CoroutineStart.DEFAULT,
        builder: (ActorScope<E>) -> Actor<E>
): ActorJob<E> = actor(context, capacity, start) {
    val actor = builder(this)
    actor.invoke()
}

or both?

[elizarov]

Makes sense. I'd keep an extended version. Maybe rename actor.invoke() to actor.run() (similar to Thread's run method). I'm not sure that an actor base class should be called ActorBase. I'd either keep it an Actor to be short or name it AbstractActor.

[elizarov]

Let me voice my concern that adding such functionality to the core is a start of a slippery road of turning it in into an actor-based programming framework. There are lots of issues to be addressed in a large-scale actor-based programming and the corresponding support libraries are only bound to grow over time. Maybe we should think-through what other things needs to be added, beyond an actor base-class and a construction function, and have a separate module to support all of that.

[elizarov]

Also, if we are starting on a path of generic actor-based programming, then we are inevitably going to be compared with other actor-based frameworks, including comparisons on the basis of performance. It does not mean that we have to worry about performance up-front, but we have to keep our designs, at least, optimizable in the future.

That is my concern for overridable invoke/run method and other potential design decisions. Let me elaborate. A generic actor can, for example, start multiple concurrent actors to process a single mailbox concurrently if the actor's logic is stateless by itself. However, a stateful actor is must typically process its mailbox in serial fashion, e.g. it is a single consumer of its mailbox.

Current implementation of Channel is multi-consumer/multi-producer queue which makes it widely-applicable (it can do both use-cases), but it inevitably costs some performance. One might implement an optimized version of channel for single-producer case to support stateful-actor use-case with better performance, but then we have to ensure that it is not possible for an actor to accidentally launch additional coroutines to read from its mailbox (to make the design less error-prone and more fool-proof). It means that we should actually hide actor's ReceiveChannel from the actor itself.

[fvasco]

At this time of development, a stateless concurrent actor can be implemented using a BroadcastChannel to create a different ActorScope for each worker actor.

However your concern sounds right to me, but I suggest to mantain an uniformity of actor function and Actor class

public interface Actor<in E> : CoroutineScope {
    public suspend fun onReceive(message: E)
}

public fun <E> actor(
        context: CoroutineContext,
        capacity: Int = 0,
        start: CoroutineStart = CoroutineStart.DEFAULT,
        onReceive: suspend CoroutineScope.(E) -> Unit
): ActorJob<E>

In such case we deprecate ActorScope in favour of Actor.

[elizarov]

The idea of concurrent stateless actor is to have multiple instance working on the same mailbox in round-robin fashion as opposed to broadcasting messages to all actors. I actually considered adding an additional concurrency: Int = 1 optional parameter to actor builder to cover this use-case, so that you can launch multiple actors under the umbrella of a single ActorJob. However, it is not that really straight-forward as it does require us to start dabbling into supervision strategies, because multiple running coroutines do make it urgent to figure out what you are going to do when one of them crashes.

[fvasco]

Yes obviously, I was wrong.

As you proposed above I consider preferable a different module for actor supervisor and so on.
In my opinion this module should offer a common pool of interfaces for asynchronous programming, an "actor" module should implement more complex use cases, but I hope that the "core" module can interoperate well with existent solutions like Quasar or Akka.

So in my limited scope of view this "core" module shoud propose an "Actor" as a "Producer" dual, all other implementation should extends core's interfaces and provides its own builders.

[elizarov]

Let me record here a draft design that I current have with respect to complex actors. The idea is to radically reduce boiler-plate that is required for writing actors that accept multiple message types by providing a handy facility that totally avoids the need to write sealed class for messages by following this pattern:

class MyActor : Actor() {
    // actor state is here, must be private

    init {
        // optional state initialization code
    }

    // public actor operation without result
    suspend fun operationA(someParams...) = act {
        // logic here
    }

    // public actor operation with result
    suspend fun operationB(otherParams...): Result = actAndReply {
        // logic here
    }

    private fun helper() { /* regular code, can access state */ }
}

The idea is that all public functions defined on actor must be suspend and must invoke a special function (tentatively called act here) that captures "message parameters" (function parameters) and stores the message in the actor's inbox for processing. When response to the caller is needed, then a separate function (tentatively called actAndReply here) is used.

We can have IDE inspections to help avoid "sharing" pitfalls to verify that all of the actor's state is private and all the public functions are properly written using act pattern. IDE inspection can also help to ensure that only immutable types are passed in/out of actor or the proper defensive copies are made on mutable data (the latter is hard to ensure, though, without complicated reference capability annotations akin to the ones used in Pony language).

The reference to MyActor instance serves as a type-safe reference to an actor and should implement a Job interface. The actor is started lazily on the first invocation of any of its public (acting) functions or by explicitly invoking start(). Lazy start feature prevents races when systems of communicating actors are written like this:

class MySystem { 
    private val actorA = object : Actor() {
         // somewhere sends message to actorB
    }

    private val actorB = object : Actor() {
        // somewhere sends message to actorA
    }

    suspend fun facadeOperation() { actorA.doSomethig() }
}

We shall also consider changing default to CoroutineStart.LAZY for a simple actor { ... } builder, too.

[fvasco]

Hi @elizarov
I wish expose some consideration about your -interesting- proposal.

Actor, as intended until now, is the dual of Producer, so your design does not help to write complex actor.

Your proposal looks like a fully synchronized, non blocking class, which is equally interesting.
I wrote a dummy implementation using your requirement plus one: avoid abstract superclass; so I wrote an task synchronization helper (TaskQueue here) and a case of use.

class TaskQueue(
        val context: CoroutineContext = DefaultDispatcher,
        val mutex: Mutex = Mutex(),
        lazyInit: (suspend CoroutineScope.() -> Unit)? = null
) {

    private var lazyInit: Deferred<Unit>?

    init {
        this.lazyInit = lazyInit?.let {
            async(context, start = CoroutineStart.LAZY, block = it)
        }
    }

    /**
     * Force lazy initialization
     */
    suspend fun init() {
        lazyInit?.run {
            await()
            lazyInit = null
        }
    }

    suspend operator fun <T> invoke(block: () -> T): T {
        init()
        return mutex.withLock(this) {
            withContext(context) {
                block()
            }
        }
    }
}

class HttpSession {

    val start = Instant.now()

    private lateinit var state: MutableMap<String, String>

    private val taskQueue = TaskQueue {
        state = mutableMapOf()
    }

    suspend fun get(key: String) = taskQueue {
        state[key]
    }

    suspend fun set(key: String, value: String) {
        taskQueue {
            state[key] = value
        }
    }
}

Plus: using this implementation and issue #94 makes easy to implement a read task queue and a write task queue.