Structured Concurrency is a set of features to augment concurrency in the Swift language. It provides

  • Continuations: async/await and async-let.
  • A cooperative thread pool to run tasks with priorities, dependencies, local values, and groups.
  • Sendable: A marker protocol that signifies thread-safe types.
  • Actors: compiler-implemented reference types that ensure mutual exclusion..


An asynchronous function is one that suspends itself while waiting for a slow computation to complete. The key innovation is that suspension does not block the thread, as is typically the case with synchronous functions.

Consider the following example:

func execute(request: URLRequest) async throws -> (Data, URLResponse) {
    try await request, delegate: nil)

let request = URLRequest(url: URL(string: "")!)
let (data, response) = try await execute(request: request)

Two aspects are noteworthy:

  • It runs a network request without a completion handler.
  • The function is declared with async and called with await.
// declaration examples
func f() async { ... }
func f() async -> X { ... }
func f() async throws { ... }
func f() async throws -> X { ... }

// call examples
await f()
try await f()
let x = await f()
let x = try await f()

See also:

  • Example
  • Why is this important?
  • How does it work?
  • Explicit continuations, API
  • Don’t mix SC and GCD


import Foundation

// Return the first 'lines' of the given URL.
func head(url: URL, lines: Int) async throws -> String {
    try await url.lines.prefix(lines).reduce("", +)

// Async functions can only be called from 'async contexts'
// Here, an async context is created with Task {}
Task {
    let url = URL(string: "")!
    let firstLine = try await head(url: url, lines: 1)

// Wait 2 seconds for the async function to complete
    mode: RunLoop.Mode.default, 
    before: NSDate(timeIntervalSinceNow: 2) as Date
Things of note:
  • head is an async function that calls another async function (URL.lines).
  • Async functions must be called from async contexts. An async context is created here using Task {}. We’ll see tasks in a bit.

This code compiles since Xcode 13 on Big Sur, or swiftc on macOS Monterey. I prefer CodeRunner for small examples as playgrounds can be cumbersome.

Starting with Xcode 13.2 beta, deployment to previous OS versions is supported: macOS 10.15, iOS 13, tvOS 13, and watchOS 6 or newer.

Why is this important?

async/await avoids the problems of completion handlers
  • Less verbose –no pyramid of doom.
  • One execution path –no nesting or many paths for error/success.
  • Allows async functions to return a value –prevents missed completion calls.
  • Eliminates the need to avoid unintentional captures using weak/unowned.
  • Offers less cognitive overhead and improved performance compared to reactive programming.

GCD can be problematic

Apple advises using one serial GCD queue per subsystem in your app. However:

  • If the queue is not running, scheduled work runs without a context switch.
  • If the queue is running, the calling thread is blocked. We say that the queue is under contention.
When a thread blocks, the system may attempt to make progress launching more threads, potentially causing thread explosion: a system _overcommitted_ with more threads than cores. This can lead to:
  • Possible deadlocks from threads competing for the same lock.
  • Overhead that may negate the benefits of asynchronous work.
    • Memory overhead from the stack and kernel data structures tracking the thread.
    • Scheduling overhead.
    • Context-switching overhead.

This can be ameliorated by scheduling work with dispatch async, at the cost of creating a new thread to schedule the work on the serial queue. Although this requires a context switch and incurs extra overhead, it prevents blocking the thread.

In short: GCD is difficult to use and prone to overuse. Concurrency with GCD is often less efficient than employing a serial queue.

How does it work?

Here is an async function that downloads a URL as plain text:
func get(url: URL) async throws -> String {
    let (data, response) = try await URL)
    return String(data: data, encoding: .utf8)!
I could turn an asynchronous function to synchronous by adding a lock:
func get(url: URL) -> String {
    let lock = os_unfair_lock_s()
    var string: String! { 
        string = ... $0
    return string

This gets rid of the completion handler but it’s verbose and incurs overhead –besides hiding the internal async call. Instead, async functions use some magic sauce to avoid it.

When you call an async function this happens:

  1. The caller code is suspended and stored as a heap “async frame” –not on the stack!.
  2. The caller thread becomes free to execute any other async frame –it is not blocked!.
  3. The async function called executes on any available thread.
  4. When the async function finishes, the first thread available resumes executing the async frame of the caller.

Threads are not blocked, so there is no chance for deadlock or thread explosion.

Explicit continuations

We mentioned that calling an async function puts your code in a suspended state. A continuation is the runtime representation of that suspended state. Coding one explicitly is useful to migrate completion handlers to async functions.
import Foundation

func functionWithCompletionHandler(_ handler: (Int) -> Void) {

func asyncFunction() async throws -> Int  {
    try await withCheckedThrowingContinuation { (continuation: CheckedContinuation<Int, Error>) in
        functionWithCompletionHandler { result in
            continuation.resume(with: .success(result))
    mode: RunLoop.Mode.default, 
    before: NSDate(timeIntervalSinceNow: 2) as Date
There are a few variants you can use


Don’t mix SC and GCD

Don’t mix SC code with code that makes assumptions about threads, like GCD.

  • e.g. let’s say your GCD code spawns execution on a dispatch queue and makes an async call. The async call changes the thread but your GCD code still thinks it is on the thread of the dispatch queue. This may result in a crash.
  • e.g. let’s say your async code calls a sync function that uses GCD semaphores and dispatch queues and returns a value. This is perfectly safe because you didn’t mix both paradigms.


An async-let binds the result of an async function but doesn’t await until the result is first referenced.

async let result = f()
// ... execution continues while f is executing
await print(result) // caller suspends awaiting the result

See also: example.


import Foundation

func get(url: URL) async throws -> String {
    let (data, _) = try await url)
    return String(data: data, encoding: .utf8)!

Task {
    async let string = get(url: url)
    print("runs immediately")
    await print(try string.prefix(15).description)

// wait 2 seconds so the async function can finish
    mode: RunLoop.Mode.default, 
    before: NSDate(timeIntervalSinceNow: 2) as Date

This prints runs immediately followed by the first 15 characters of the page. It does not call get(url: url) until the result (string) is referenced.

Behind the scenes, async-let spawns a child Task that runs the async function and is implicitly awaited when the result is first referenced.



A task is a unit of asynchronous work.

Task features:

  • They express child-parent dependencies –this prevents priority inversion.
  • They can be canceled, queried, re-prioritized.
  • They store task-local values that are inherited by child tasks.

Because Task have the concept of child-parent, they are called different things in respect to each other:

Task { ... } // unstructured task

Task.detached { ... } // detached task

let parentTask = Task { // parent task
    Task {} // child task
    Task.detached {} // child detached task

parentTask.cancel() // cancels parent and child, but not the detached task

Task eats exceptions

If the code inside a Task throws an exception you won’t receive any warning. If you want to remind yourself to catch errors, you can write this instead:

typealias SafeTask<T> = Task<T, Never>
// usage
SafeTask<Void> { 
    // no throwing allowed here
    // if anything throws you must wrap it in a do/catch

Detached tasks

A detached task doesn’t inherit anything from the parent task

  • doesn’t inherit priorities
  • doesn’t inherit local task storage
  • it’s not canceled when the parent is

A detached task completes even when there are no references pointing to it.

See also:

  • Why do tasks exist?
  • Return a result from a task
  • Using Task() without priority argument
  • API

Why do tasks exist?

Before structured programming, control flow used global state and jumped all over using goto. It was hard to follow because it lacked local reasoning: the ability to figure out the code in front of your eyes without reading elsewhere.

Similarly multithreaded code is unstructured:

  1. GCD blocks can’t express dependencies between them.
  2. Code is all over the place: multiple paths, hard to read, verbose handling, must call completion only once.

Because of this, the compiler can’t analyze our code. This changes with async functions and tasks. In short:

  • async functions simplify asynchronous code
  • tasks express relationships between work items
And tieing both together: every async function runs inside a Task.

Return a result from a task

import Foundation

Task {
    // to access a value returned use `value`
    print(await Task(operation: { 0 }).value) // 0
    // to access a value or error use `result`
    struct MyError: Error {}
    switch await Task(operation: { throw MyError() }).result {
        case .success(let int): print(int)
        case .failure(let error): print("\(type(of: error))") // MyError
    // in a Task Result, failure is error or never
    let _: Result<Int, Never> = await Task(operation: { 0 }).result
    let _: Result<Int, Error> = await Task(operation: { throw MyError() }).result
    mode: RunLoop.Mode.default, 
    before: NSDate(timeIntervalSinceNow: 1) as Date

An interesting use case is that waiting on a task.value let us share computation results. For instance, let say you store the task reference for an ongoing network request, then you request the same endpoint while the first one is ongoing. You don’t need to repeat the request, just await on firstTask.value and you’ll be reading the result. Is this confusing? check this annotated example.

Using Task() without priority argument

Using Task() without specifying a priority argument can lead to unintended consequences, as it inherits the priority of the code that created it.

For example, if you spawn multiple Tasks from the main thread, they will all inherit and compete with your UI code for CPU resources. This can result in:

  • UI hiccups or stuttering.
  • Slower overall code execution.
  • Potential deadlocks when tasks depend on each other.
  • iOS suspending some tasks to avoid exceeding the number of device cores.

To prevent these issues, specify the priority using, for example, Task(priority: .medium).

If deadlocks persist, the talk Swift concurrency: Behind the scenes introduced a flag LIBDISPATCH_COOPERATIVE_POOL_STRICT=1. This flag limits the simulator to one concurrent task per priority, ensuring progress. However, it deviates further from actual device behavior, which allows multiple tasks per priority.


  • Task.sleep() suspends the task, not the thread.
  • task.isCancelled returns a boolean, doesn’t throw.
  • task.checkCancellation() throws if task is cancelled.
  • task.result or task.value reads the value returned (if any).


TaskGroup is a container designed to hold an arbitrary number of tasks.

let result = await withTaskGroup(of: Int.self) { group -> Int in    
    group.addTask { ... }
    group.addTask { ... }
    return await group.reduce(0, +)

When a TaskGroup is executed, all added tasks begin running immediately and in any order. Once the initialization closure exits, all child tasks are implicitly awaited. If a task throws an error, all tasks are canceled, the group waits for them to finish, and then the group throws the original error.

See also: example 1, example 2.


    import Foundation
    Task {
        let values = [1, 3, 5, 7]
        let sum = await withTaskGroup(of: Int.self) { group -> Int in    
            for value in values {
                group.addTask { value }
            return await group.reduce(0, +)
        mode: RunLoop.Mode.default, 
        before: NSDate(timeIntervalSinceNow: 2) as Date
In this code
  • line 5: withTaskGroup(of: Int.self) creates a group with return type Int.
  • line 6: Iterates through the values, adding an arbitrary number of tasks, each returning an Int.
  • line 9: Awaits the completion of all tasks (imagine they are slow) and sums the results.

If there were 400 values to process, the number of active tasks would likely be limited to the machine's number of cores, although this is not customizable or guaranteed.

Example 2: Map Reduce

    import UIKit
    import XCTest
    struct MapReduce
        static func mapReduce<T,U>(inputs: [T], process: @escaping (T) async throws -> U ) async throws -> [U] {
            try await withThrowingTaskGroup(of: U.self) { group in
                // map
                for input in inputs {
                    group.addTask {
                        try await process(input)
                // reduce
                var collected = [U]()
                for try await value in group {
                return collected
    final class PhotosTests: XCTestCase
        func testDownloader() async throws {
            let input = [
                URL(string: "")!,
                URL(string: "")!
            let download: (URL) async throws -> UIImage = { url in
                let (data, _) = try await url)
                return UIImage(data: data)!
            let images = try await MapReduce.mapReduce(inputs: input, process: download)
            XCTAssertEqual(images.count, 2)
This example distributes an arbitrary number of items in tasks, then collects the results. Note that if you intend to download images you are probably better served with a cache.


Sendable is a marker that indicates types that can be safely transferred across threads. There is a Sendable marker protocol and a @Sendable function attribute.

final class Counter: Sendable { ... }

func bar() { ... }

See also:

  • What is thread-safety?
  • What is a marker protocol?
  • What types are thread-safe? (sorry, long list)
  • Compiler checks on Sendable

Thread safety

Thread-safe code is code that remains correct when called from multiple threads.

Correct code is code that conforms to its specification. A good specification defines

  • Invariants constraining the state.
  • Preconditions and postconditions describing the effects of the operations.

For thread-safe code this means:

  • No sequence of operations can violate the specification. That is, see the object in an invalid state or violate the pre/post conditions.
  • Invariants and conditions will hold during multithread execution without requiring additional synchronization by the client.
For thread safety purposes code has three kinds of state:
  • Immutable.
  • Mutable private.
  • Mutable shared (aka public).
To create thread-safe code we have to do just one thing: regulate the access to mutable shared state. And there are three ways to do it:
  • Prevent access.
  • Make the state immutable.
  • Synchronize the access.
The first two are simple. The third one requires preventing a number of issues on concurrent systems:
  • liveness
    • deadlock: two threads block permanently waiting for each other to release a needed resource.
    • livelock: a thread is busy working but it's unable to make any progress.
    • starvation: a thread is perpetually denied access to resources it needs to make progress.
  • safe publication: both the reference and the state of the published object must be made visible to other threads at the same time.
  • race conditions: a defect where the output is dependent on the timing of uncontrollable events. In other words, a race condition happens when getting the right answer relies on lucky timing.
The approach of Swift Concurrency to deal with all this is: simplify async code, don’t block threads, be explicit about dependencies and which objects are thread-safe. This is not fool-proof but provides some compilation errors for unsafe situations.

Marker protocol

A marker protocol is one with these traits:

  • It indicates an intent.
  • It has no requirements.
  • It can’t inherit from a non-marker protocol.
  • It can’t be used with is or as?.

Conforming a type to a marker protocol maintains binary compatibility. Obviously, source compatibility is broken if the marker protocol wasn’t previously known.

Thread-safe types

Thread-safe Types: structs, enums, tuples of safe types, actors, immutable classes, internally-synchronized classes, and functions/closures whose capture list (if any) contains safe types.

Not Thread-safe: reference types with unsynchronized mutable state, or value types that point to them, or functions that capture them.

Thread-safe types which have automatic conformance to Sendable

  • Nearly every standard library type with value semantics with few exceptions: ManagedBuffer, pointer types, types returned by lazy algorithms.
  • Error, because by throwing, any function sends the error across concurrency domains.
  • Tuples whose elements are Sendable.
  • Metatypes, because they are immutable.
  • Key paths literals when they only capture values of Sendable types.
  • Actors because they have internal synchronization.
  • Closures and functions
  • Classes
    • Classes marked with @unchecked that provide their own synchronization.
    • Other classes that fulfill all these conditions:
      • They are final
      • All properties are immutable Sendable types
      • They are root classes or inherit directly from NSObject
      • Generic values (if any) are guaranteed to be of Sendable type (e. g. <T:Sendable>)
  • Structs that fulfill the following:
    • All members are Sendable.
    • Generic values (if any) are guaranteed to be of Sendable type (e. g. <T:Sendable>).
    • They are either frozen public structs or non-public structs that are not @usableFromInline
  • Enums that fulfill the following:
    • Their associated values are Sendable. If they are generic, the generic must be defined as <T:Sendable>.
    • They are either frozen public enums or non-public enums that are not @usableFromInline
  • Some C types
    • C enum.
    • C struct if all properties conform to Sendable.
    • C function pointers (because they cannot capture values).

Compiler checks on Sendable

For Sendable the compiler checks the following:

  • Non-thread-safe types that inherit from Error are Sendable but shouldn’t be. The compiler will issue warnings for those to ease the transition.
  • It is a compiler error to mark Sendable on an object that isn’t. But this error can be suppressed by adding @unchecked for classes that use internal synchronization to guarantee thread safety. Example:
  • final class Foo {} // OK
    class Foo: Sendable { } // error
    class Foo: @unchecked Sendable {}  // OK

For @Sendable the compiler checks the following:

  • No mutable captures.
  • Cannot be both synchronous and actor isolated –this would allow the closure to be sent to another thread and run code on the actor (because it is actor isolated).
  • Captures must be of Sendable type.
  • Accessors are currently not allowed.
  • Closures that have @Sendable function type can only use by-value captures. Example:
  • import Foundation
    func count() {
        var n = 0
        @Sendable func increment() {
            n += 1 // Error: Mutation of captured var 'n' in concurrently-executing code


A task local value is a Sendable value associated with a task context. They are inherited by the task’s children. When the root task context ends, the value is discarded.

To declare a property as task-local add static and annotate with @TaskLocal:

enum Request {
    @TaskLocal static var id = 0

To get/set a task local value

print( // 0

Request.$id.withValue( + 1) {
    print( // 1

See also:

  • Where can I use them?
  • API

Where can I use them?

You may use task local values from any task context:
  • a Task {} or Task.detached {}
  • or any of their Task children, even a child within a nested group
But surprisingly also from:
  • an async function –because it runs in a task internally,
  • a sync function –because it emulates task local values with a thread-local implementation,
  • a function declared outside but called from within the context
However, do not set local values within a group but outside a task. I mean:
    import Foundation
    enum Request {
        @TaskLocal static var id = 0
    Task {
        await withTaskGroup(of: String.self) { group in
            print( // OK
            Request.$id.withValue( + 1) {} // ABORT signal
        mode: RunLoop.Mode.default, 
        before: NSDate(timeIntervalSinceNow: 2) as Date



An actor is a reference type with mutual exclusion –meaning: only one call to the actor is active at a time. This is also known as actor isolation.

This is implemented using a serial executor which serializes calls from inside or outside the actor. Calls inside the actor run uninterrumpted to their end so they are synchronous. Calls from other types may need to wait for their turn so they must be asynchronous.

nonisolated is a keyword that disables isolation on selected methods and variables. It is allowed as long as nonisolated code doesn’t interact with isolated code.

actor Counter 
    // this is safe to mark nonisolated since it is a constant
    nonisolated let name = "my counter"
    var count = 0
    func incrementAge() -> Int {
        age += 1
        return value

See also: Priority inversion.

Priority inversion

Priority inversion is a low priority task holding the progress of a high priority task. For instance, because both need to access a resource guarded with mutual exclusion, e.g. another actor. This is not a problem with actors, let’s see why.

Actor hopping is calling an actor from another actor. This is allowed as long as they read an immutable property or invoke an asynchronous function.

An actor’s function runs uninterrupted (because of mutual exclusion) until it invokes an async function of another actor. At that point, it is suspended. When execution resumes, there is a chance that the actor executed other calls that changed its state, so the rest of the function shouldn’t rely on pre-existing actor’s state.

Actor reentrancy is scheduling work on an actor that already has suspended work items. Whenever the actor has several work items suspended, it is free to reorder their execution. Being able to reorder pending work avoids priority inversion.

Global actors

A global actor is a singleton actor available globally. It can be used to isolate a whole object, individual properties, or methods.

import Foundation

// declaration of a custom global actor
actor MyActor {
    static let shared = MyActor()

// isolation of a whole object
final class SomeClass {
    // opting out on selected members
    nonisolated let foo = "bar"

// isolation of individual properties and methods
final class SomeOtherClass {
    @MyActor var state = 0
    @MyActor func foo() {}

One such actor is the @MainActor, which is an attribute that indicates code (a function, property, or a whole type) should run on the main thread. This eliminates the guesswork about whether code should run on the main thread.

Running from an async context is the equivalent to GCD’s DispatchQueue.main.async:

Task {
    await {
        // same as DispatchQueue.main.async { .. }

In a function annotated @MainActor, any further async calls don’t block the main thread. For instance:

func fetchImage(for url: URL) async throws -> UIImage {
    // doesn’t block main
    let (data, _) = try await url) 
    // image decompression does block main
    guard let image = UIImage(data: data) else { 
        throw ImageFetchingError.imageDecodingFailed
    return image

See also:

  • API
  • Global actor inference
  • Minimize hoping on main
  • How to run fully / on main


Global actor inference

Declarations not marked with an actor can infer actor isolation according to the rules of Global actor inference. For instance, assuming that the actor is @MainActor, the rules are as follow:
  • A class that inherits from a @MainActor superclass.
  • A method overriding a @MainActor method.
  • A struct or class using a property wrapper with @MainActor for its wrapped value.
  • A method that implements a protocol that declares that method as @MainActor. There is an exception where this doesn’t happen: a method implemented in an extension that does not declare conformance to the protocol. See example below.
    protocol Configurable {
        @MainActor func configure()
    struct Cell: Configurable {}
    extension Cell { // this extension doesn’t declare conformance
        func configure() // not main actor
    struct Cell {}
    extension Cell: Configurable { // this extension declares conformance
        func configure() // main actor
  • A type that implements a protocol that declares @MainActor for the whole protocol. There is an exception where this doesn’t happen: a type that adds conformance in an extension. See example below.
    protocol Configurable {}
    // main type
    struct Cell: Configurable {} 
    // not main type
    struct Cell {}
    extension Cell: Configurable {}

These rules imply that if we create an async context and access an object annotated with @MainActor (e.g. a UILabel) the call will automatically be routed through the main thread. However, this is only true if we are in an async context.

For instance, in the following code let uilabel be a UILabel, that is now declared with @MainActor in UIKit.

func f() {
    Task {
        uilabel.text = "..." // OK, done on main
    } .background).async {
        uilabel.text = "..." // Error!

Minimize hoping on main

Structured concurrency runs on “cooperative thread pools”, which seem to be dispatch queues. The main actor is an exception because it runs on the main thread. For instance,

    Task(priority: TaskPriority.background, operation: {
        await MyMainActor.shared.f(value: 0)

creates two threads

       Thread 1 Queue : (serial)
       Thread 2 Queue : (serial)

The cooperative pool reuses threads for actors, async functions, etc. But calls to the main actor require a context switch to the main thread, which could decrease your FPS.

As Apple said in #10254 Behind the scenes

You should weigh the entertainment value of incremental updates (multiple switching) against the efficiency of batching up work (one switch).

How to run fully / on main

If you have a code path that you want to

  • Execute fully (without reentrant calls): put it inside an actor and don’t spawn additional async calls.
  • Execute on main: annotate it with @MainActor and don’t spawn additional async calls to code not protected by @MainActor.

Do you know why?

  • An actor prevents parallel execution (two active threads running the same code at the same time), not concurrent execution (two threads running the same code one at a time).
  • @MainActor guarantees that code runs in main, but async calls to unprotected code MAY suspend your @MainActor code, run the unprotected code on a thread different than main, and cause a reentrant call to your @MainActor code.


AsyncSequence is an ordered, asynchronously generated sequence of elements. Basically, a Sequence with an asynchronous Example, API.

Many library objects now return async sequences:

let url = URL(string: "")!
let firstTwoLines = try await url.lines.prefix(2).reduce("", +)


import Foundation

private enum Source {

    case local
    case remote
    case terminated

final class CollectionSequence: AsyncSequence, AsyncIteratorProtocol {

    let backend: BackendService
    let persistence: CollectionContainer

    init(backend: BackendService, persistence: CollectionContainer) {
        self.backend = backend
        self.persistence = persistence

    private func localCollections() async throws -> [Collection] {
        try await persistence.collections()

    private func remoteCollections() async throws -> [Collection] {
        let dtos = try await backend.perform(operation: .getAllCollections())
        try await dtos)
        return try await localCollections()

    private var source = Source.local

    // MARK: - AsyncSequence

    typealias Element = [Collection]

    func makeAsyncIterator() -> CollectionSequence {

    // MARK: - AsyncIteratorProtocol

    func next() async throws -> Element? {
        switch source {
        case .local:
            source = .remote
            return try await localCollections()
        case .remote:
            source = .terminated
            return try await remoteCollections()
        case .terminated:
            return nil

for try await value in CollectionSequence() {
    print("collections: \(value)")


AsyncStream is an asynchronous sequence generated from a closure that calls a continuation to produce new elements. Example, API. Things of note regarding AsyncStream:

  • Iterating over an AsyncStream multiple times, or creating multiple iterators is considered a programmer error.
  • The closure used to create the AsyncStream can’t call async code. If you need to await, use AsyncSequence instead.
  • If your code can throw, use AsyncThrowingStream instead.


    import Foundation
    let stream = AsyncStream(Int?.self) { continuation in
        continuation.onTermination = { @Sendable reason in
            print("Termination reason: \(reason)") // finished or cancelled
        for n in [0, nil, 2] {
            // ... imagine a slow async process returning 'n'

    Task {
        for await value in stream { 
            print(value as Any) // 0, nil, 2
        mode: RunLoop.Mode.default, 
        before: NSDate(timeIntervalSinceNow: 2) as Date

Iteration ends when either throws or returns nil.

When iterating a collection of optionals the equivalent of nil is Optional.none, not Optional.some(nil). For instance, in this example next() returned:

  • .some(0)
  • .some(nil)
  • .some(2)
  • .nonemarks the end so the loop body is not executed


Preparing for Swift 6

The article Concurrency in Swift 5 and 6 talks about restrictions that will be applied in Swift 6. To enable some of them pass these flags to the compiler:

OTHER_SWIFT_FLAGS: -Xfrontend -warn-concurrency -Xfrontend -enable-actor-data-race-checks

Or from SPM add this to your target

swiftSettings: [
        "-Xfrontend", "-warn-concurrency",
        "-Xfrontend", "-enable-actor-data-race-checks",

Sometimes the solution is straightforward, e.g. use weak self to refer to controllers, e.g. add mutated objects to the capture list, etc. Some errors are a pain in the ass, e.g. Data is not Sendable so you can’t pass it around and it may be too expensive to copy.

These flags may or may not be useful. They

  • reveal some errors that may pass undetected in the default compiler
  • force you to consider edge cases in your code
  • are at times too restrictive to be taken seriously

According to Doug Gregor:

To be very clear, I don’t actually suggest that you use -warn-concurrency in Swift 5.5. It’s both too noisy in some cases and misses other cases. Swift 5.6 brings a model that’s designed to deal with gradual adoption of concurrency checking.

See also: Staging in Sendable checking.


Not much is new. From XCTestCase:

If your app uses Swift Concurrency, annotate test methods with async or async throws instead to test asynchronous operations, and use standard Swift concurrency patterns in your tests.

For instance

class MyTests: XCTestCase
    override func setUp() async throws { ... }
    func testMyCode() async throws { ... }

However, class methods seem to be not supported. Using NSLock or os_unfair_lock_lock primitives inside are ignored. I guess you could use a runloop and check for a side effect but it’s too ugly.

// not supported
override class func setUp() async throws { ... }

See also