Implementation of a highly-scalable and ergonomic actor model for Rust

Axiom

Axiom brings a highly-scalable actor model to the Rust language based on the many lessons learned over years of Actor model implementations in Akka and Erlang. Axiom is, however, not a direct re-implementation of either of the two aforementioned actor models but rather a new implementation deriving inspiration from the good parts of those projects.

• 2019-12-19 0.2.1
  • Fixed a critical issue where pending Actor Handles were dropped early.
  • Fixed a critical issue where panics weren't caught on poll of Actor Handles.
• 2019-12-06 0.2.0
  • Massive internal refactor in order to support async Actors. There are only a few breaking changes, so porting to this version will be relatively simple.
  • BREAKING CHANGE: The signature for Processors has changed from references for Context and Message to values. For closures-as-actors, wrap the body in an async block. move |...| {...} becomes |...| async move { ... }. For regular function syntax, simply add async in front of fn.
  • NOTE: the positioning of move may need to be different, depending on semantics. Values cannot be moved out of the closure and into the async block.
  • BREAKING CHANGE: Due to the nature of futures, the actor's processor cannot be given a mutable reference to the state of the actor. The state needs to live at least as long as the future and our research could find no way to do this easily. So now when the actor returns a status it will return the new state as well. See the examples for more info. The signature for the processor is now:

  impl<F, S, R> Processor<S, R> for F where
      S: Send + Sync,
      R: Future<Output = AxiomResult<S>> + Send + 'static,
      F: (FnMut(S, Context, Message) -> R) + Send + Sync + 'static  {}
  

  • BREAKING: Actors are now panic-tolerant! This means asserts and panics will be caught and converted, treated the same as errors. Errors should already be considered fatal, as Actors should handle any errors in their own scope.
  • BREAKING: Error types have been broken up to be more context-specific.
  • BREAKING: A start_on_launch flag has been added to the ActorSystemConfig struct. This allows for an ActorSystem to be created without immediately starting it. See ActorSystem::start for how to start an unstarted ActorSystem.
  • Helper methods have been added to Status to help with the return points in Actors. Each variant has a corresponding function that takes the Actor's state. Ok(Status::Done) is instead Ok(Status::done(state)).
  • The user should take be aware that, at runtime, Actors will follow the semantics of Rust Futures. This means that an Actor awaiting a future will not process any messages nor will continue executing until that future is ready to be polled again. While async/await will provide ergonomic usage of async APIs, this can be a concern and can affect timing.
  • A prelude has been introduced. Attempts will be made at keeping the prelude relatively the same even across major versions, and we recommend using it whenever possible.
  • More log points have been added across the codebase.

Release Notes for All Versions

Getting Started

An actor model is an architectural asynchronous programming paradigm characterized by the use of actors for all processing activities.

Actors have the following characteristics:

1. An actor can be interacted with only by means of messages.
2. An actor processes only one message at a time.
3. An actor will process a message only once.
4. An actor can send a message to any other actor without knowledge of that actor's internals.
5. Actors send only immutable data as messages, though they may have mutable internal state.
6. Actors are location agnostic; they can be sent a message from anywhere in the cluster.

Note that within the language of Rust, rule five cannot be enforced by Rust but is a best practice which is important for developers creating actors based on Axiom. In Erlang and Elixir rule five cannot be violated because of the structure of the language but this also leads to performance limitations. It's better to allow internal mutable state and encourage the good practice of not sending mutable messages.

What is important to understand is that these rules combined together makes each actor operate like a micro-service in the memory space of the program using them. Since actor messages are immutable, actors can trade information safely and easily without copying large data structures.

Although programming in the actor model is quite an involved process you can get started with Axiom in only a few lines of code.

use axiom::prelude::*;
use std::sync::Arc;
use std::time::Duration;

let system = ActorSystem::create(ActorSystemConfig::default().thread_pool_size(2));

let aid = system
    .spawn()
    .with(
        0 as usize,
        |state: usize, _context: Context, _message: Message| async move {
            Ok(Status::done(state))
        }
    )
    .unwrap();

aid.send(Message::new(11)).unwrap();

// It is worth noting that you probably wouldn't just unwrap in real code but deal with
// the result as a panic in Axiom will take down a dispatcher thread and potentially
// hang the system.

// This will wrap the value `17` in a Message for you!
aid.send_new(17).unwrap();

// We can also create and send separately using just `send`, not `send_new`.
let message = Message::new(19);
aid.send(message).unwrap();

// Another neat capability is to send a message after some time has elapsed.
aid.send_after(Message::new(7), Duration::from_millis(10)).unwrap();
aid.send_new_after(7, Duration::from_millis(10)).unwrap();

This code creates an actor system, fetches a builder for an actor via the spawn() method, spawns an actor and finally sends the actor a message. Once the actor is done processing a message it returns the new state of the actor and the status after handling this message. In this case we didnt change the state so we just return it. Creating an Axiom actor is literally that easy but there is a lot more functionality available as well.

Keep in mind that if you are capturing variables from the environment you will have to wrap the async move {} block in another block and then move your variables into the first block. Please see the test cases for more examples of this.

If you want to create an actor with a struct that is simple as well. Let's create one that handles a couple of different message types:

use axiom::prelude::*;
use std::sync::Arc;

let system = ActorSystem::create(ActorSystemConfig::default().thread_pool_size(2));

struct Data {
    value: i32,
}

impl Data {
    fn handle_bool(mut self, message: bool) -> ActorResult<Self> {
        if message {
            self.value += 1;
        } else {
            self.value -= 1;
        }
        Ok(Status::done(self))
    }

    fn handle_i32(mut self, message: i32) -> ActorResult<Self> {
        self.value += message;
        Ok(Status::done(self))
    }

    async fn handle(mut self, _context: Context, message: Message) -> ActorResult<Self> {
        if let Some(msg) = message.content_as::<bool>() {
            self.handle_bool(*msg)
        } else if let Some(msg) = message.content_as::<i32>() {
            self.handle_i32(*msg)
        } else {
            panic!("Failed to dispatch properly");
        }
    }
}

let data = Data { value: 0 };
let aid = system.spawn().name("Fred").with(data, Data::handle).unwrap();

aid.send_new(11).unwrap();
aid.send_new(true).unwrap();
aid.send_new(false).unwrap();

This code creates a named actor out of an arbitrary struct. Since the only requirement to make an actor is to have a function that is compliant with the axiom::actors::Processor trait, anything can be an actor. If this struct had been declared somewhere outside of your control you could use it in an actor as state by declaring your own handler function and making the calls to the 3rd party structure.

It's important to keep in mind that the starting state is moved into the actor and you will not have external access to it afterwards. This is by design and although you could conceivably use a [Arc] or Mutex enclosing a structure as state, that would definitely be a bad idea as it would break the rules we laid out for actors.

There is a lot more to learn and explore and your best resource is the test code for Axiom. The developers have a belief that test code should be well architected and well commented to act as a set of examples for users of Axiom.

Detailed Examples

• Hello World: The obligatory introduction to any computer system.
• Dining Philosophers: An example of using Axiom to solve a classic Finite State Machine problem in computer science.
• Monte Carlo: An example of how to use Axiom for parallel computation.

Design Principals of Axiom

Based on previous experience with other actor models I wanted to design Axiom around some core principles:

1. At its core an actor is just an function that processes messages. The simplest actor is a function that takes a message and simply ignores it. The benefit to the functional approach over the Akka model is that it allows the user to create actors easily and simply. This is the notion of micro module programming; the notion of building a complex system from the smallest components. Software based on the actor model can get complicated; keeping it simple at the core is fundamental to solid architecture.
2. Actors can be a Finite State Machine (FSM). Actors receive and process messages nominally in the order received. However, there are certain circumstances where an actor has to change to another state and process other messages, skipping certain messages to be processed later.
3. When skipping messages, the messages must not move. Akka allows the skipping of messages by stashing the message in another data structure and then restoring this stash later. This process has many inherent flaws. Instead Axiom allows an actor to skip messages in its channel but leave them where they are, increasing performance and avoiding many problems.
4. Actors use a bounded capacity channel. In Axiom the message capacity for the actor's channel is bounded, resulting in greater simplicity and an emphasis on good actor design.
5. Axiom should be kept as small as possible. Axiom is the core of the actor model and should not be expanded to include everything possible for actors. That should be the job of libraries that extend Axiom. Axiom itself should be an example of micro module programming.
6. The tests are the best place for examples. The tests of Axiom will be extensive and well maintained and should be a resource for those wanting to use Axiom. They should not be a dumping ground for copy-paste or throwaway code. The best tests will look like architected code.
7. A huge emphasis is put on crate user ergonomics. Axiom should be easy to use.