Skip to main content

Resource

Allocation and release of resources is not easy, especially when we have multiple resources that depend on each other. The Resource DSL adds the ability to install resources and ensure proper finalization even in the face of exceptions and cancellations. Arrow's Resource co-operated with Structured Concurrency and KotlinX Coroutines.

Media resources
Where to find it

Resource management is part of the arrow-fx-coroutines library. The separate arrow-autoclose library provides a similar API but without integration with Kotlin's coroutine mechanism.

Understanding the problem​

The following program is not safe because it is prone to leak dataSource and userProcessor when an exception or cancellation signal occurs while using the service.

class UserProcessor {
  fun start(): Unit = println("Creating UserProcessor")
  fun shutdown(): Unit = println("Shutting down UserProcessor")
}

class DataSource {
  fun connect(): Unit = println("Connecting dataSource")
  fun close(): Unit = println("Closed dataSource")
}

class Service(val db: DataSource, val userProcessor: UserProcessor) {
  suspend fun processData(): List<String> = 
    throw RuntimeException("I'm going to leak resources by not closing them")
}

For example, the following application would leak resources.

suspend fun example() {
  val userProcessor = UserProcessor().also { it.start() }
  val dataSource = DataSource().also { it.connect() }
  val service = Service(dataSource, userProcessor)

  service.processData()

  dataSource.close()
  userProcessor.shutdown()
}

If we were using Kotlin JVM, we might rely on Closeable or AutoCloseable and rewrite our code.

suspend fun example() {
  UserProcessor().use { userProcessor ->
    userProcessor.start()
    DataSource().use { dataSource ->
      dataSource.connect()
      Service(dataSource, userProcessor).processData()
    }
  }
}

However, while we fixed the closing of UserProcessor and DataSource, there are still issues with this code:

  1. It requires implementing Closeable or AutoCloseable, which is only possible for Kotlin JVM but not available for Multiplatform.
  2. Requires implementing an interface or wrapping external types with something like class CloseableOf<A>(val type: A): Closeable.
  3. Requires nesting of different resources in callback tree, not composable.
  4. Enforces close method name, renamed UserProcessor#shutdown to close
  5. Cannot run suspend functions within fun close(): Unit.
  6. No exit signal; we don't know if we exited successfully, with an error or cancellation.

Resource solves these issues. The main idea is that each resource has three stages: 1️⃣ acquiring the resource, 2️⃣ using the resource, and 3️⃣ releasing the resource. With Resource, we bundle steps (1) and (3), and the implementation ensures that everything works correctly, even in the event of exceptions or cancellations.

Graceful Shutdowns

Correct release of resources when the application is terminating is important in several scenarios. SuspendApp improves on Resource to gracefully deal with shutdown and termination.

Dealing with resources properly​

You can use Arrow's Resource in two ways:

  1. Using resourceScope and functions with ResourceScope as its receiver.
  2. Wrapping the entire resource allocation and release as a Resource<A> value, which we later use in a larger block.

Using resourceScope​

The ResourceScope DSL allows you to install resources and safely interact with them. In fact, that's the only operation you need to learn about: install takes both the acquisition and release steps as arguments. The result of this function is whatever was acquired, plus the promise of running the finalizer at the end of the block.

tip

The Resource DSL gives you enough flexibility to perform different actions depending on how the execution finished: successful completion, exceptions, or cancellation. The second argument to the finalizer is of type ExitCase and represents the reason why the finalizer is run.

The code below shows our example rewritten to use resourceScope. Note that we acquire our UserProcessor and DataSource in parallel, using the parZip operation in Arrow. This means that their start and connect methods can run in parallel.

suspend fun ResourceScope.userProcessor(): UserProcessor =
  install({ UserProcessor().also { it.start() } }) { p, _ -> p.shutdown() }

suspend fun ResourceScope.dataSource(): DataSource =
  install({ DataSource().also { it.connect() } }) { ds, _ -> ds.close() }

suspend fun example(): Unit = resourceScope {
  val service = parZip({ userProcessor() }, { dataSource() }) { userProcessor, ds ->
    Service(ds, userProcessor)
  }
  val data = service.processData()
  println(data)
}

The code above also showcases a very common pattern of resource acquisition: running the constructor, followed by calling some start method using Kotlin's also scope function.

note

To achieve its behavior, install invokes the acquire and release step as NonCancellable. If a cancellation signal or an exception is received during acquire, the resource is assumed to not have been acquired and thus will not trigger the release function; any composed resources that are already acquired are guaranteed to release as expected.

Interfacing with Java​

If you're running on the JVM, Arrow provides built-in integration with AutoCloseable in the form of the closeable function.

Using Resource​

The usage of resource is very similar to install. The main difference is that the result is a value of type Resource<T>, where T is the type of the resource to acquire. But such a value doesn't run the acquisition step, it's simply a recipe describing how that's done; to actually acquire the resource, you need to call .bind() inside a resourceScope.

val userProcessor: Resource<UserProcessor> = resource({
  UserProcessor().also { it.start() }
}) { p, _ -> p.shutdown() }

val dataSource: Resource<DataSource> = resource({
  DataSource().also { it.connect() }
}) { ds, exitCase ->
  println("Releasing $ds with exit: $exitCase")
  withContext(Dispatchers.IO) { ds.close() }
}

val service: Resource<Service> = resource {
  Service(dataSource.bind(), userProcessor.bind())
}

suspend fun example(): Unit = resourceScope {
  val data = service.bind().processData()
  println(data)
}
info

Why provide two ways to accomplish the same goal? Although resourceScope provides nicer syntax in general, some usage patterns like acquiring several resources become easier when the steps are saved in an actual class.

The actual magic is that Resource is nothing more than a type alias for parameter-less function using ResourceScope,

typealias Resource<A> = suspend ResourceScope.() -> A

Although the primary usage pattern is to give resource the acquisition and release steps directly, there's another way to define a Resource<T>. Arrow provides a resource for more complex scenarios that takes a block with ResourceScope as a receiver. That allows calling install as required.

val userProcessor: Resource<UserProcessor> = resource {
  val x: UserProcessor = install(
    {  UserProcessor().also { it.start() } },
    { processor, _ -> processor.shutdown() }
  )
  x
}

Integration with typed errors​

Resource management cooperates with typed error builders. It's important to be aware that the order in which we open the scopes affects the behavior. To be more concrete, let's consider the two possible nestings of resourceScope and either.

  1. When either is in the outermost position and resourceScope is inside of it, a bind that crosses the resourceScope results in the release finalizer being called with Cancelled.

    either<String, Int> {
      resourceScope {
        val a = install({ }) { _, ex -> println("Closing A: $ex") }
        raise("Boom!")
      } // Closing A: ExitCase.Cancelled
    } // Either.Left(Boom!)

  2. With reverse nesting order of either and resourceScope, then resources are released with a normal state since nothing "failed."

    resourceScope {
      either<String, Int> {
        val a = install({ }) { _,ex -> println("Closing A: $ex") }
        raise("Boom!")
      } // Either.Left(Boom!)
    } // Closing A: ExitCase.Completed

We remark that, in both cases, resources are correctly released. If you're finalizer works in the same way for every possible ExitCase, then there's no visible difference between both.

info

If you want to know more, this conversation in the Kotlin Slack goes into more detail.