Introducing structured concurrency in Swift

In my earlier tutorial we have talked about the brand new async/await function in Swift, after that I’ve created a weblog submit about thread protected concurrency utilizing actors, now it’s time to get began with the opposite main concurrency function in Swift, known as structured concurrency. 🔀

What’s structured concurrency? Nicely, lengthy story quick, it is a new task-based mechanism that enables builders to carry out particular person process gadgets in concurrently. Usually if you await for some piece of code you create a possible suspension level. If we take our quantity calculation instance from the async/await article, we may write one thing like this:

let x = await calculateFirstNumber()
let y = await calculateSecondNumber()
let z = await calculateThirdNumber()
print(x + y + z)

I’ve already talked about that every line is being executed after the earlier line finishes its job. We create three potential suspension factors and we await till the CPU has sufficient capability to execute & end every process. This all occurs in a serial order, however generally this isn’t the habits that you really want.

If a calculation will depend on the results of the earlier one, this instance is ideal, since you should use x to calculate y, or x & y to calculate z. What if we would wish to run these duties in parallel and we do not care the person outcomes, however we’d like all of them (x,y,z) as quick as we will? 🤔

async let x = calculateFirstNumber()
async let y = calculateSecondNumber()
async let z = calculateThirdNumber()

let res = await x + y + z
print(res)

I already confirmed you ways to do that utilizing the async let bindings proposal, which is a form of a excessive stage abstraction layer on prime of the structured concurrency function. It makes ridiculously straightforward to run async duties in parallel. So the large distinction right here is that we will run all the calculations directly and we will await for the outcome “group” that incorporates each x, y and z.

Once more, within the first instance the execution order is the next:

• await for x, when it’s prepared we transfer ahead
• await for y, when it’s prepared we transfer ahead
• await for z, when it’s prepared we transfer ahead
• sum the already calculated x, y, z numbers and print the outcome

We may describe the second instance like this

• Create an async process merchandise for calculating x
• Create an async process merchandise for calculating y
• Create an async process merchandise for calculating z
• Group x, y, z process gadgets collectively, and await sum the outcomes when they’re prepared
• print the ultimate outcome

As you’ll be able to see this time we do not have to attend till a earlier process merchandise is prepared, however we will execute all of them in parallel, as a substitute of the common sequential order. We nonetheless have 3 potential suspension factors, however the execution order is what actually issues right here. By executing duties parallel the second model of our code might be means quicker, because the CPU can run all of the duties directly (if it has free employee thread / executor). 🧵

At a really fundamental stage, that is what structured concurrency is all about. After all the async let bindings are hiding a lot of the underlying implementation particulars on this case, so let’s transfer a bit right down to the rabbit gap and refactor our code utilizing duties and process teams.

await withTaskGroup(of: Int.self) { group in
    group.async {
        await calculateFirstNumber()
    }
    group.async {
        await calculateSecondNumber()
    }
    group.async {
        await calculateThirdNumber()
    }

    var sum: Int = 0
    for await res in group {
        sum += res
    }
    print(sum)
}

In accordance with the present model of the proposal, we will use duties as fundamental models to carry out some form of work. A process might be in one in every of three states: suspended, working or accomplished. Activity additionally assist cancellation and so they can have an related precedence.

Duties can type a hierarchy by defining youngster duties. Presently we will create process teams and outline youngster gadgets by means of the group.async operate for parallel execution, this youngster process creation course of might be simplified through async let bindings. Kids robotically inherit their father or mother duties’s attributes, similar to precedence, task-local storage, deadlines and they are going to be robotically cancelled if the father or mother is cancelled. Deadline assist is coming in a later Swift launch, so I will not discuss extra about them.

A process execution interval is known as a job, every job is working on an executor. An executor is a service which may settle for jobs and arranges them (by precedence) for execution on obtainable thread. Executors are at the moment supplied by the system, however afterward actors will be capable to outline customized ones.

That is sufficient idea, as you’ll be able to see it’s attainable to outline a process group utilizing the withTaskGroup or the withThrowingTaskGroup strategies. The one distinction is that the later one is a throwing variant, so you’ll be able to attempt to await async capabilities to finish. ✅

A process group wants a ChildTaskResult kind as a primary parameter, which needs to be a Sendable kind. In our case an Int kind is an ideal candidate, since we’ll acquire the outcomes utilizing the group. You may add async process gadgets to the group that returns with the correct outcome kind.

We will collect particular person outcomes from the group by awaiting for the the following ingredient (await group.subsequent()), however because the group conforms to the AsyncSequence protocol we will iterate by means of the outcomes by awaiting for them utilizing a regular for loop. 🔁

That is how structured concurrency works in a nutshell. The perfect factor about this complete mannequin is that through the use of process hierarchies no youngster process shall be ever in a position to leak and hold working within the background accidentally. This a core cause for these APIs that they need to at all times await earlier than the scope ends. (thanks for the ideas @ktosopl). ❤️

Let me present you a number of extra examples…

Ready for dependencies

You probably have an async dependency in your process gadgets, you’ll be able to both calculate the outcome upfront, earlier than you outline your process group or inside a gaggle operation you’ll be able to name a number of issues too.

import Basis

func calculateFirstNumber() async -> Int {
    await withCheckedContinuation { c in
        DispatchQueue.major.asyncAfter(deadline: .now() + 2) {
            c.resume(with: .success(42))
        }
    }
}

func calculateSecondNumber() async -> Int {
    await withCheckedContinuation { c in
        DispatchQueue.major.asyncAfter(deadline: .now() + 1) {
            c.resume(with: .success(6))
        }
    }
}

func calculateThirdNumber(_ enter: Int) async -> Int {
    await withCheckedContinuation { c in
        DispatchQueue.major.asyncAfter(deadline: .now() + 4) {
            c.resume(with: .success(9 + enter))
        }
    }
}

func calculateFourthNumber(_ enter: Int) async -> Int {
    await withCheckedContinuation { c in
        DispatchQueue.major.asyncAfter(deadline: .now() + 3) {
            c.resume(with: .success(69 + enter))
        }
    }
}

@major
struct MyProgram {
    
    static func major() async {

        let x = await calculateFirstNumber()
        await withTaskGroup(of: Int.self) { group in
            group.async {
                await calculateThirdNumber(x)
            }
            group.async {
                let y = await calculateSecondNumber()
                return await calculateFourthNumber(y)
            }
            

            var outcome: Int = 0
            for await res in group {
                outcome += res
            }
            print(outcome)
        }
    }
}

It’s value to say that if you wish to assist a correct cancellation logic you have to be cautious with suspension factors. This time I will not get into the cancellation particulars, however I will write a devoted article in regards to the subject sooner or later in time (I am nonetheless studying this too… 😅).

Duties with completely different outcome sorts

In case your process gadgets have completely different return sorts, you’ll be able to simply create a brand new enum with related values and use it as a standard kind when defining your process group. You need to use the enum and field the underlying values if you return with the async process merchandise capabilities.

import Basis

func calculateNumber() async -> Int {
    await withCheckedContinuation { c in
        DispatchQueue.major.asyncAfter(deadline: .now() + 4) {
            c.resume(with: .success(42))
        }
    }
}

func calculateString() async -> String {
    await withCheckedContinuation { c in
        DispatchQueue.major.asyncAfter(deadline: .now() + 2) {
            c.resume(with: .success("The that means of life is: "))
        }
    }
}

@major
struct MyProgram {
    
    static func major() async {
        
        enum TaskSteps {
            case first(Int)
            case second(String)
        }

        await withTaskGroup(of: TaskSteps.self) { group in
            group.async {
                .first(await calculateNumber())
            }
            group.async {
                .second(await calculateString())
            }

            var outcome: String = ""
            for await res in group {
                swap res {
                case .first(let worth):
                    outcome = outcome + String(worth)
                case .second(let worth):
                    outcome = worth + outcome
                }
            }
            print(outcome)
        }
    }
}

After the duties are accomplished you’ll be able to swap the sequence components and carry out the ultimate operation on the outcome primarily based on the wrapped enum worth. This little trick will will let you run all form of duties with completely different return sorts to run parallel utilizing the brand new Duties APIs. 👍

Unstructured and indifferent duties

As you may need observed this earlier than, it isn’t attainable to name an async API from a sync operate. That is the place unstructured duties can assist. Crucial factor to notice right here is that the lifetime of an unstructured process shouldn’t be sure to the creating process. They will outlive the father or mother, and so they inherit priorities, task-local values, deadlines from the father or mother. Unstructured duties are being represented by a process deal with that you should use to cancel the duty.

import Basis

func calculateFirstNumber() async -> Int {
    await withCheckedContinuation { c in
        DispatchQueue.major.asyncAfter(deadline: .now() + 3) {
            c.resume(with: .success(42))
        }
    }
}

@major
struct MyProgram {
    
    static func major() {
        Activity(precedence: .background) {
            let deal with = Activity { () -> Int in
                print(Activity.currentPriority == .background)
                return await calculateFirstNumber()
            }
            
            let x = await deal with.get()
            print("The that means of life is:", x)
            exit(EXIT_SUCCESS)
        }
        dispatchMain()
    }
}

You will get the present precedence of the duty utilizing the static currentPriority property and test if it matches the father or mother process precedence (in fact it ought to match it). ☺️

So what is the distinction between unstructured duties and indifferent duties? Nicely, the reply is kind of easy: unstructured process will inherit the father or mother context, however indifferent duties will not inherit something from their father or mother context (priorities, task-locals, deadlines).

@major
struct MyProgram {
    
    static func major() {
        Activity(precedence: .background) {
            Activity.indifferent {
                
                print(Activity.currentPriority == .background)
                let x = await calculateFirstNumber()
                print("The that means of life is:", x)
                exit(EXIT_SUCCESS)
            }
        }
        dispatchMain()
    }
}

You may create a indifferent process through the use of the indifferent methodology, as you’ll be able to see the precedence of the present process contained in the indifferent process is unspecified, which is certainly not equal with the father or mother precedence. By the best way additionally it is attainable to get the present process through the use of the withUnsafeCurrentTask operate. You need to use this methodology too to get the precedence or test if the duty is cancelled. 🙅‍♂️

@major
struct MyProgram {
    
    static func major() {
        Activity(precedence: .background) {
            Activity.indifferent {
                withUnsafeCurrentTask { process in
                    print(process?.isCancelled ?? false)
                    print(process?.precedence == .unspecified)
                }
                let x = await calculateFirstNumber()
                print("The that means of life is:", x)
                exit(EXIT_SUCCESS)
            }
        }
        dispatchMain()
    }
}

There’s yet another large distinction between indifferent and unstructured duties. In the event you create an unstructured process from an actor, the duty will execute instantly on that actor and NOT in parallel, however a indifferent process shall be instantly parallel. Which means that an unstructured process can alter inner actor state, however a indifferent process cannot modify the internals of an actor. ⚠️

You too can benefit from unstructured duties in process teams to create extra complicated process buildings if the structured hierarchy will not suit your wants.

Activity native values

There’s yet another factor I would like to indicate you, we have talked about process native values various occasions, so this is a fast part about them. This function is mainly an improved model of the thread-local storage designed to play good with the structured concurrency function in Swift.

Typically you would like to hold on customized contextual knowledge along with your duties and that is the place process native values are available. For instance you would add debug info to your process objects and use it to search out issues extra simply. Donny Wals has an in-depth article about process native values, if you’re extra about this function, you must positively learn his submit. 💪

So in observe, you’ll be able to annotate a static property with the @TaskLocal property wrapper, after which you’ll be able to learn this metadata inside an one other process. Any further you’ll be able to solely mutate this property through the use of the withValue operate on the wrapper itself.

import Basis

enum TaskStorage {
    @TaskLocal static var identify: String?
}

@major
struct MyProgram {
    
    static func major() async {
        await TaskStorage.$identify.withValue("my-task") {
            let t1 = Activity {
                print("unstructured:", TaskStorage.identify ?? "n/a")
            }
            
            let t2 = Activity.indifferent {
                print("indifferent:", TaskStorage.identify ?? "n/a")
            }
            
            _ = await [t1.value, t2.value]
        }
    }
}

Duties will inherit these native values (besides indifferent) and you’ll alter the worth of process native values inside a given process as effectively, however these adjustments shall be solely seen for the present process & youngster duties. To sum this up, process native values are at all times tied to a given process scope.

As you’ll be able to see structured concurrency in Swift is rather a lot to digest, however when you perceive the fundamentals all the things comes properly along with the brand new async/await options and Duties you’ll be able to simply assemble jobs for serial or parallel execution. Anyway, I hope you loved this text. 🙏