JarCastingFactor: An experiment in Functional Actors
Factor is an experimental Scala API layer built on top of Akka and Cats Effect that offers a significantly different programming model to classic actors.
Instead of sending a message to an actor who then interpret it, a Factor ("functional actor") is sent an effectful function to be run asynchronously in its process and on its state. The factor then replies with a value sent back to the caller.
If classic actors extend the ideas of message-passing object-orientation into a concurrent, distributed context, then Factor could be described as extending the Haskell-style applying functions over exposed data types into a distributed scenario.
Factor reuses the quality engineering of Akka to power its distribution mechanism, but turns the user-programming model on its head.
SBT Coordinates
"com.github.benhutchison" %% "factor" % "0.3"
Factor is published for Scala 2.12.
API Overview
Imports Guide
//Factor library itself
import factor._
//for specifying timeout durations
import scala.concurrent.duration._
//for IO
import cats.effect._
//to enable IO operations such as `sequence`, `void`
import cats.implicits._
While Akka is used internally, its imports are not required for regular usage.
Factors
A Factor is an active, potentially remote container for some state S and an environment (or configuration) of type E. To program a factor, the programmer sends a function to the factor to be run asynchronously; the state of the factor is updated, and a reply is sent back.
Factors are not accessed directly, but via a FactorRef[E, S] which are serializable, remote handles to a Factor. The fundamental operation they provide is to asynchronously run an effectful (ie potentially side-effecting) function inside the actor, and return a response to the caller wrapped in cats.effect.IO:
trait FactorRef {
def run[A](f: (E, S) => IO[(S, A)])(implicit timeout: FiniteDuration): IO[A]
}
That is, the passed function can read the current environment and state, and can respond with new state, and a reply A. The reply is returned asynchronously inside an cats.effect.IO. Note the passed function, any values it has "closed over", and the reply value must be Serializable.
A number of variants of this primitive operation are built on top of it:
runPure[A](f: (E, S) => (S, A))(implicit timeout: FiniteDuration): IO[A]: The sent function is pure.runS[A](f: (S) => IO[(S, A)])(implicit timeout: FiniteDuration): IO[A]: The sent function reads the Factor state but not environment.runSPure[A](f: (S) => (S, A))(implicit timeout: FiniteDuration): IO[A]: A pure function reads the Factor state but not environment.runUpdatePure(f: (E, S) => S)(implicit timeout: FiniteDuration): IO[Unit]: A pure function updates the state.runSUpdatePure(f: (S) => S)(implicit timeout: FiniteDuration): IO[Unit]: A pure function reads only the Factor state and updates the state.getS(implicit timeout: FiniteDuration): IO[S]: read and return the current state.putS(newState: S)(implicit timeout: FiniteDuration): IO[Unit]: set the current stategetE(implicit timeout: FiniteDuration): IO[E]: read the environment or configuration.
In the event of a reply not received within the timeout, the IO will complete with a FactorTimeout runtime exception.
Creating or Obtaining a FactorRef
Each system that is going to host or call Factors needs to first start a FactorSystem which wraps an underlying Akka ActorSystem.
val myAddress = SystemAddress("MySystemName", "127.0.0.1", 3456)
val system = new FactorSystem(address)
Then a Factor may be created in a local or remote system via a createFactor call:
type E = String
type S = Int
val ref: FactorRef[E, S] = system.createFactor[E, S](
"immutable Factor environment or config value",
initialState = 1,
"name of the factor",
localOrRemoteAddress
)
The recommendation is to create a factor ref and then pass it to any processes that need to interact with the Factor.
However, if you need obtain a ref to a Factor you somehow know exists, but didn't create, you can use findFactor. Be warned that this lookup will not verify the types E and S you specify actually match those of the looked-up Factor (think of them as an assertion you make using some external knowledge):
val ref: FactorRef[E, S] = system.findFactor[E, S](
"name of the factor",
localOrRemoteAddress
)
Running a Computation
Factor inherits IOs deferred execution behavior. That is, the structure of a computation is declared, and then it is triggered, typically once at the "end of the world" (ie program), using eg runUnsafeSync. See the example code for an sample.
Termination
Factors persist until they are stopped. stop on FactorRef will stop a Factor, while terminate on FactorSystem will terminate the whole system on that machine including all Factors.
Coordinating Multi-Factor Computations
A key aim of the Factor model is to allow a distributed actor system to be coordinated like a chain of local asynchronous Futures, IOs or Tasks, by using monadic flatMap/binds to sequence steps in a multi-part computation.
IO also provides good support for running parallel computations. For example, you can dispatch two calls to two factors concurrently using IO.parTraverse(f1, f2) and continue when the results from each are back.
Keep in mind that calls to a Factor can be effectful, so they can read or write to databases, message queues, the cloud, etc as well as manipulating the Factor state. Sometimes the goal of a call is to cause effects, and the Factor state is just a book-keeping track the status of effects.
runAsync for Remote Orchestration
Another operation in the factor API exists to support orchestrating asynchronous calls and/or external effects remotely:
case class FactorSystem {
def runAsync[A](f: ()=>IO[A], address: SystemAddress)(implicit timeout: FiniteDuration): IO[A]
}
This will run a potentially async, slow or effectful operation on the address specified, but outside of any Factor. An ephermal one-shot Akka actor is created to run the f and then auto-disposed. It allows the control of a computation to migrate to another machine, either so chatty interactions Factors are fast local calls, or because external asynch calls should happen there.
The reason runAsync exists is that operations run inside a Factor should be brief and blocking. While processing a message, you can think of that message as having an exclusive lock to see and update the state of the actor. In Factor, the messages are pieces of runnable code. But if that code makes a non-blocking asynch call and then exists the message handler, its forfeited its right to update the actor state when the async call comes back.
I know its unfashionable to advise synchronous code in 2018, but the reality is that the Actor model is built on an requirement that message processing is synchronous.
Factor & Classic Actors Compared
Compared to classic actors, Factor takes a different position on encapsulation & data hiding. Factor exposes the state of its actors as a core feature of the programming model. This supports the capability to send mobile code from afar into a factor to be run.
In contrast, classic actors follow the object-oriented tradition of hiding the state of actors away and exposing only a set of messages, to indirectly manipulate the state.
Over the past 30 years or so, much has been written of the benefits of encapsulation; too much perhaps, so that the costs of encapsulation have seen too little recognition or discussion. My own background started in object-oriented Java, and the need for encapsulation had been drummed into me for years. I distinctly remember being quite surprised when I first studied Haskell, that idiomatic Haskell programming style chose exposed- over encapsulated- data, and what's more, they seemed to get along just fine!
So what are the trade offs? When we encapsulate, we protect ourselves against the unstable or untrusted outside world. But this comes at the cost of maintaining a hidden internal state, plus an exterior API, and a means of translating messages into state changes and back again. And this is a good approach when the outside world is unstable or untrusted.
But in many scenarios, we control the actor's environment ourselves. We have a system composed of many actors that we have fashioned to work together. Just as we normally wouldn't keep the interior doors in our house locked, erecting walls inside our system just adds overhead and complexity we don't want or need.
When you have a coherent system that needs to include distribution and concurrency, Factor can let your express your system's behavior more clearly and concisely. You can write all your code in one place, declaratively describe which parts should execute where, and you'll typically expend less effort on packaging and distribution overheads.
Serialization
Factor is built on top of Akka and uses it's serialization for transporting functions and data across the network. This delegates to Java's inbuilt serialization
A common gotcha when serializing functions is closing over, or capturing references to, values in the callers context that are not serializable. Lets look at a simple example of how this happens:
class NonSerializablePredicate {def isLucky(n: Int): Boolean = n != 13}
val p = new NonSerializablePredicate
system.createFactorS[Int](7, "example", anAddress).runSUpdatePure(n => if (p.isLucky(n)) n else 0)
In the closure passed to the Factor, value p has been captured and must be serialized. A runtime exception will be thrown here because p isn't Serializable.
Example
The sausage factory example shows a working Factor system.
It includes unit tests and a unit-integration test which starts 3 JVMs to run the system in full distributed mode.
Note the example also uses Monocle lenses for elegant updates to immutable data
Status & Roadmap
Factor is currently an experiment. Source compatiblity may break between releases. Your feedback on the APIs or desirable features is welcome.
Once I get accumulate sufficient experience using Factor myself, and gather some feedback from other users, I intend to publish a stable version.
Features I am planning or considering adding to Factor:
- Machine checked docs
- Kryo as the default serialization mechanism
- Asbtract away from IO to the Cats Effect typeclasses
Contributing
At this stage, user feedback, features requests and problem reports are particularly useful.
Contributions that are aligned with the design intent of the library are welcome. Go ahead and send PRs for minor changes, but please raise an issue first for any major feature or change.