An asynchronous programming facility for Scala
retronym Merge pull request #198 from retronym/topic/dot-dse
Performance improvements to the macro
Latest commit b662af6 Aug 6, 2018

README.md

scala-async

Supported Scala versions

This branch targets Scala 2.12 and 2.13.

Support for Scala 2.11 is on a branch.

Support for Scala 2.10 is on a branch.

Quick start

To include scala-async in an existing project use the library published on Maven Central. For sbt projects add the following to your build definition - build.sbt or project/Build.scala:

libraryDependencies += "org.scala-lang.modules" %% "scala-async" % "0.10.0"

For Maven projects add the following to your (make sure to use the correct Scala version suffix to match your project’s Scala binary version):

<dependency>
	<groupId>org.scala-lang.modules</groupId>
	<artifactId>scala-async_2.12</artifactId>
	<version>0.10.0</version>
</dependency>

After adding a scala-async to your classpath, write your first async block:

import scala.concurrent.ExecutionContext.Implicits.global
import scala.async.Async.{async, await}

val future = async {
  val f1 = async { ...; true }
  val f2 = async { ...; 42 }
  if (await(f1)) await(f2) else 0
}

What is async?

async marks a block of asynchronous code. Such a block usually contains one or more await calls, which marks a point at which the computation will be suspended until the awaited Future is complete.

By default, async blocks operate on scala.concurrent.{Future, Promise}. The system can be adapted to alternative implementations of the Future pattern.

Consider the following example:

def slowCalcFuture: Future[Int] = ...             // 01
def combined: Future[Int] = async {               // 02
  await(slowCalcFuture) + await(slowCalcFuture)   // 03
}
val x: Int = Await.result(combined, 10.seconds)   // 05

Line 1 defines an asynchronous method: it returns a Future.

Line 2 begins an async block. During compilation, the contents of this block will be analyzed to identify the await calls, and transformed into non-blocking code.

Control flow will immediately pass to line 5, as the computation in the async block is not executed on the caller's thread.

Line 3 begins by triggering slowCalcFuture, and then suspending until it has been calculated. Only after it has finished, we trigger it again, and suspend again. Finally, we add the results and complete combined, which in turn will release line 5 (unless it had already timed out).

It is important to note that while lines 1-4 are non-blocking, they are not parallel. If we wanted to parallelize the two computations, we could rearrange the code as follows:

def combined: Future[Int] = async {
  val future1 = slowCalcFuture
  val future2 = slowCalcFuture
  await(future1) + await(future2)
}

Comparison with direct use of Future API

This computation could also be expressed by directly using the higher-order functions of Futures:

def slowCalcFuture: Future[Int] = ...
val future1 = slowCalcFuture
val future2 = slowCalcFuture
def combined: Future[Int] = for {
  r1 <- future1
  r2 <- future2
} yield r1 + r2

The async approach has two advantages over the use of map and flatMap:

  1. The code more directly reflects the programmer's intent, and does not require us to name the results r1 and r2. This advantage is even more pronounced when we mix control structures in async blocks.
  2. async blocks are compiled to a single anonymous class, as opposed to a separate anonymous class for each closure required at each generator (<-) in the for-comprehension. This reduces the size of generated code, and can avoid boxing of intermediate results.

Comparison with CPS plugin

The existing continuations (CPS) plugin for Scala can also be used to provide a syntactic layer like async. This approach has been used in Akka's Dataflow Concurrency (now deprecated in favour of this library).

CPS-based rewriting of asynchronous code also produces a closure for each suspension. It can also lead to type errors that are difficult to understand.

How it works

  • The async macro analyses the block of code, looking for control structures and locations of await calls. It then breaks the code into 'chunks'. Each chunk contains a linear sequence of statements that concludes with a branching decision, or with the registration of a subsequent state handler as the continuation.
  • Before this analysis and transformation, the program is normalized into a form amenable to this manipulation. This is called the "A Normal Form" (ANF), and roughly means that:
    • if and match constructs are only used as statements; they cannot be used as an expression.
    • calls to await are not allowed in compound expressions.
  • Identify vals, vars and defs that are accessed from multiple states. These will be lifted out to fields in the state machine object.
  • Synthesize a class that holds:
    • an integer representing the current state ID.
    • the lifted definitions.
    • an apply(value: Try[Any]): Unit method that will be called on completion of each future. The behavior of this method is determined by the current state. It records the downcast result of the future in a field, and calls the resume() method.
    • the resume(): Unit method that switches on the current state and runs the users code for one 'chunk', and either: a) registers the state machine as the handler for the next future b) completes the result Promise of the async block, if at the terminal state.
    • an apply(): Unit method that starts the computation.

Limitations

  • See the neg test cases for constructs that are not allowed in an async block.
  • See the issue list for which of these restrictions are planned to be dropped in the future.
  • See #32 for why await is not possible in closures, and for suggestions on ways to structure the code to work around this limitation.