Skip to content

Latest commit

 

History

History
1486 lines (1143 loc) · 52.5 KB

AsyncGuidance.md

File metadata and controls

1486 lines (1143 loc) · 52.5 KB
Raw

Table of contents

Asynchronous Programming

Asynchronous programming has been around for several years on the .NET platform but has historically been very difficult to do well. Since the introduction of async/await in C# 5 asynchronous programming has become mainstream. Modern frameworks (like ASP.NET Core) are fully asynchronous and it's very hard to avoid the async keyword when writing web services. As a result, there's been lots of confusion on the best practices for async and how to use it properly. This section will try to lay out some guidance with examples of bad and good patterns of how to write asynchronous code.

Asynchrony is viral

Once you go async, all of your callers SHOULD be async, since efforts to be async amount to nothing unless the entire call stack is async. In many cases, being partially asynchronous can be worse than being entirely synchronous. Therefore it is best to go all in, and make everything async at once.

❌ BAD This example uses the Task.Result and as a result blocks the current thread to wait for the result. This is an example of sync over async.

public int DoSomethingAsync()
{
    var result = CallDependencyAsync().Result;
    return result + 1;
}

✅ GOOD This example uses the await keyword to get the result from CallDependencyAsync.

public async Task<int> DoSomethingAsync()
{
    var result = await CallDependencyAsync();
    return result + 1;
}

Async void

The use of async void in ASP.NET Core applications is ALWAYS bad. Avoid it, never do it. Typically, it's used when developers are trying to implement fire-and-forget patterns triggered by a controller action. Async void methods will crash the process if an exception is thrown. We'll look at more of the patterns that cause developers to do this in ASP.NET Core applications but here's a simple example:

❌ BAD Async void methods can't be tracked and therefore unhandled exceptions can result in application crashes.

public class MyController : Controller
{
    [HttpPost("/start")]
    public IActionResult Post()
    {
        BackgroundOperationAsync();
        return Accepted();
    }
    
    public async void BackgroundOperationAsync()
    {
        var result = await CallDependencyAsync();
        DoSomething(result);
    }
}

✅ GOOD Task-returning methods are better since unhandled exceptions trigger the TaskScheduler.UnobservedTaskException.

public class MyController : Controller
{
    [HttpPost("/start")]
    public IActionResult Post()
    {
        Task.Run(BackgroundOperationAsync);
        return Accepted();
    }
    
    public async Task BackgroundOperationAsync()
    {
        var result = await CallDependencyAsync();
        DoSomething(result);
    }
}

Prefer Task.FromResult over Task.Run for pre-computed or trivially computed data

For pre-computed results, there's no need to call Task.Run, which will end up queuing a work item to the thread pool that will immediately complete with the pre-computed value. Instead, use Task.FromResult, to create a task wrapping already computed data.

❌ BAD This example wastes a thread-pool thread to return a trivially computed value.

public class MyLibrary
{
   public Task<int> AddAsync(int a, int b)
   {
       return Task.Run(() => a + b);
   }
}

✅ GOOD This example uses Task.FromResult to return the trivially computed value. It does not use any extra threads as a result.

public class MyLibrary
{
   public Task<int> AddAsync(int a, int b)
   {
       return Task.FromResult(a + b);
   }
}

💡NOTE: Using Task.FromResult will result in a Task allocation. Using ValueTask<T> can completely remove that allocation.

✅ GOOD This example uses a ValueTask<int> to return the trivially computed value. It does not use any extra threads as a result. It also does not allocate an object on the managed heap.

public class MyLibrary
{
   public ValueTask<int> AddAsync(int a, int b)
   {
       return new ValueTask<int>(a + b);
   }
}

Avoid using Task.Run for long-running work that blocks the thread

Long-running work in this context refers to a thread that's running for the lifetime of the application doing background work (like processing queue items, or sleeping and waking up to process some data). Task.Run will queue a work item to the thread pool. The assumption is that that work will finish quickly (or quickly enough to allow reusing that thread within some reasonable timeframe). Stealing a thread-pool thread for long-running work is bad since it takes that thread away from other work that could be done (timer callbacks, task continuations, etc). Instead, spawn a new thread manually to do long-running blocking work.

💡 NOTE: The thread pool grows if you block threads but it's bad practice to do so.

💡 NOTE:Task.Factory.StartNew has an option TaskCreationOptions.LongRunning that under the covers creates a new thread and returns a Task that represents the execution. Using this properly requires several non-obvious parameters to be passed in to get the right behavior on all platforms.

💡 NOTE: Don't use TaskCreationOptions.LongRunning with async code as this will create a new thread which will be destroyed after first await.

❌ BAD This example steals a thread-pool thread forever, to execute queued work on a BlockingCollection<T>.

public class QueueProcessor
{
    private readonly BlockingCollection<Message> _messageQueue = new BlockingCollection<Message>();
    
    public void StartProcessing()
    {
        Task.Run(ProcessQueue);
    }
    
    public void Enqueue(Message message)
    {
        _messageQueue.Add(message);
    }
    
    private void ProcessQueue()
    {
        foreach (var item in _messageQueue.GetConsumingEnumerable())
        {
             ProcessItem(item);
        }
    }
    
    private void ProcessItem(Message message) { }
}

✅ GOOD This example uses a dedicated thread to process the message queue instead of a thread-pool thread.

public class QueueProcessor
{
    private readonly BlockingCollection<Message> _messageQueue = new BlockingCollection<Message>();
    
    public void StartProcessing()
    {
        var thread = new Thread(ProcessQueue) 
        {
            // This is important as it allows the process to exit while this thread is running
            IsBackground = true
        };
        thread.Start();
    }
    
    public void Enqueue(Message message)
    {
        _messageQueue.Add(message);
    }
    
    private void ProcessQueue()
    {
        foreach (var item in _messageQueue.GetConsumingEnumerable())
        {
             ProcessItem(item);
        }
    }
    
    private void ProcessItem(Message message) { }
}

✅ GOOD This example utilizes a TaskFactory with TaskCreationOptions.LongRunning to process the message queue instead of creating a thread manually.

public class QueueProcessor
{
    private readonly BlockingCollection<Message> _messageQueue = new BlockingCollection<Message>();

    public Task StartProcessing() => Task.Factory.StartNew(ProcessQueue, TaskCreationOptions.LongRunning);

    public void Enqueue(Message message)
    {
        _messageQueue.Add(message);
    }

    private void ProcessQueue()
    {
        foreach (var item in _messageQueue.GetConsumingEnumerable())
        {
            ProcessItem(item);
        }
    }

    private void ProcessItem(Message message) { }
}

Utilizing TaskCreationOptions.LongRunning introduces several advantages in comparison with manual thread creation:

  • It can be easily combined with await and TPL APIs, such as Task.WhenAll, amongst others.
  • It provides a superior exception-handling mechanism. For instance, in the event of an unhandled exception in a manually created thread, the application will crash (unless handled via AppDomain.CurrentDomain.UnhandledException), but with .LongRunning, it will be wrapped into a Task as an AggregateException.

💡 NOTE: The TaskCreationOptions.LongRunning option is essentially a recommendation to the TaskScheduler, which may interpret it differently in custom TaskScheduler applications or runtimes, or future updates to the .NET runtime libraries. If your primary goal is to spawn a new dedicated thread, then you might consider using the manual thread creation approach discussed previously.

Avoid using Task.Result and Task.Wait

There are very few ways to use Task.Result and Task.Wait correctly so the general advice is to completely avoid using them in your code.

⚠️ Sync over async

Using Task.Result or Task.Wait to block waiting on an asynchronous operation to complete is MUCH worse than calling a truly synchronous API to block. This phenomenon is dubbed "Sync over async". Here is what happens at a very high level:

  • An asynchronous operation is kicked off.
  • The calling thread is blocked waiting for that operation to complete.
  • When the asynchronous operation completes, it unblocks the code waiting on that operation. This takes place on another thread.

The result is that we need to use 2 threads instead of 1 to complete synchronous operations. This usually leads to thread-pool starvation and results in service outages.

⚠️ Deadlocks

The SynchronizationContext is an abstraction that gives application models a chance to control where asynchronous continuations run. ASP.NET (non-core), WPF, and Windows Forms each have an implementation that will result in a deadlock if Task.Wait or Task.Result is used on the main thread. This behavior has led to a bunch of "clever" code snippets that show the "right" way to block waiting for a Task. The truth is, there's no good way to block waiting for a Task to complete.

💡NOTE: ASP.NET Core does not have a SynchronizationContext and is not prone to the deadlock problem.

❌ BAD The below are all examples that are, in one way or another, trying to avoid the deadlock situation but still succumb to "sync over async" problems.

public string DoOperationBlocking()
{
    // Bad - Blocking the thread that enters.
    // DoAsyncOperation will be scheduled on the default task scheduler, and remove the risk of deadlocking.
    // In the case of an exception, this method will throw an AggregateException wrapping the original exception.
    return Task.Run(() => DoAsyncOperation()).Result;
}

public string DoOperationBlocking2()
{
    // Bad - Blocking the thread that enters.
    // DoAsyncOperation will be scheduled on the default task scheduler, and remove the risk of deadlocking.
    // In the case of an exception, this method will throw the exception without wrapping it in an AggregateException.
    return Task.Run(() => DoAsyncOperation()).GetAwaiter().GetResult();
}

public string DoOperationBlocking3()
{
    // Bad - Blocking the thread that enters, and blocking the threadpool thread inside.
    // In the case of an exception, this method will throw an AggregateException containing another AggregateException, containing the original exception.
    return Task.Run(() => DoAsyncOperation().Result).Result;
}

public string DoOperationBlocking4()
{
    // Bad - Blocking the thread that enters, and blocking the threadpool thread inside.
    return Task.Run(() => DoAsyncOperation().GetAwaiter().GetResult()).GetAwaiter().GetResult();
}

public string DoOperationBlocking5()
{
    // Bad - Blocking the thread that enters.
    // Bad - No effort has been made to prevent a present SynchonizationContext from becoming deadlocked.
    // In the case of an exception, this method will throw an AggregateException wrapping the original exception.
    return DoAsyncOperation().Result;
}

public string DoOperationBlocking6()
{
    // Bad - Blocking the thread that enters.
    // Bad - No effort has been made to prevent a present SynchonizationContext from becoming deadlocked.
    return DoAsyncOperation().GetAwaiter().GetResult();
}

public string DoOperationBlocking7()
{
    // Bad - Blocking the thread that enters.
    // Bad - No effort has been made to prevent a present SynchonizationContext from becoming deadlocked.
    var task = DoAsyncOperation();
    task.Wait();
    return task.GetAwaiter().GetResult();
}

Prefer await over ContinueWith

Task existed before the async/await keywords were introduced and as such provided ways to execute continuations without relying on the language. Although these methods are still valid to use, we generally recommend that you prefer async/await to using ContinueWith. ContinueWith also does not capture the SynchronizationContext and as a result is actually semantically different to async/await.

❌ BAD The example uses ContinueWith instead of async

public Task<int> DoSomethingAsync()
{
    return CallDependencyAsync().ContinueWith(task =>
    {
        return task.Result + 1;
    });
}

✅ GOOD This example uses the await keyword to get the result from CallDependencyAsync.

public async Task<int> DoSomethingAsync()
{
    var result = await CallDependencyAsync();
    return result + 1;
}

Always create TaskCompletionSource<T> with TaskCreationOptions.RunContinuationsAsynchronously

TaskCompletionSource<T> is an important building block for libraries trying to adapt things that are not inherently awaitable to be awaitable via a Task. It is also commonly used to build higher-level operations (such as batching and other combinators) on top of existing asynchronous APIs. By default, Task continuations will run inline on the same thread that calls Try/Set(Result/Exception/Canceled). As a library author, this means having to understand that calling code can resume directly on your thread. This is extremely dangerous and can result in deadlocks, thread-pool starvation, corruption of state (if code runs unexpectedly) and more.

Always use TaskCreationOptions.RunContinuationsAsynchronously when creating the TaskCompletionSource<T>. This will dispatch the continuation onto the thread pool instead of executing it inline.

❌ BAD This example does not use TaskCreationOptions.RunContinuationsAsynchronously when creating the TaskCompletionSource<T>.

public Task<int> DoSomethingAsync()
{
    var tcs = new TaskCompletionSource<int>();
    
    var operation = new LegacyAsyncOperation();
    operation.Completed += result =>
    {
        // Code awaiting on this task will resume on this thread!
        tcs.SetResult(result);
    };
    
    return tcs.Task;
}

✅ GOOD This example uses TaskCreationOptions.RunContinuationsAsynchronously when creating the TaskCompletionSource<T>.

public Task<int> DoSomethingAsync()
{
    var tcs = new TaskCompletionSource<int>(TaskCreationOptions.RunContinuationsAsynchronously);
    
    var operation = new LegacyAsyncOperation();
    operation.Completed += result =>
    {
        // Code awaiting on this task will resume on a different thread-pool thread
        tcs.SetResult(result);
    };
    
    return tcs.Task;
}

💡NOTE: There are 2 enums that look alike. TaskCreationOptions.RunContinuationsAsynchronously and TaskContinuationOptions.RunContinuationsAsynchronously. Be careful not to confuse their usage.

Always dispose CancellationTokenSource(s) used for timeouts

CancellationTokenSource objects that are used for timeouts (are created with timers or use the CancelAfter method), can put pressure on the timer queue if not disposed.

❌ BAD This example does not dispose of the CancellationTokenSource and as a result, the timer stays in the queue for 10 seconds after each request is made.

public async Task<Stream> HttpClientAsyncWithCancellationBad()
{
    var cts = new CancellationTokenSource(TimeSpan.FromSeconds(10));

    using (var client = _httpClientFactory.CreateClient())
    {
        var response = await client.GetAsync("http://backend/api/1", cts.Token);
        return await response.Content.ReadAsStreamAsync();
    }
}

✅ GOOD This example disposes of the CancellationTokenSource and properly removes the timer from the queue.

public async Task<Stream> HttpClientAsyncWithCancellationGood()
{
    using (var cts = new CancellationTokenSource(TimeSpan.FromSeconds(10)))
    {
        using (var client = _httpClientFactory.CreateClient())
        {
            var response = await client.GetAsync("http://backend/api/1", cts.Token);
            return await response.Content.ReadAsStreamAsync();
        }
    }
}

Always flow CancellationToken(s) to APIs that take a CancellationToken

Cancellation is cooperative in .NET. Everything in the call chain has to be explicitly passed the CancellationToken in order for it to work well. This means you need to explicitly pass the token into other APIs that take a token if you want cancellation to be most effective.

❌ BAD This example neglects to pass the CancellationToken to Stream.ReadAsync making the operation effectively not cancellable.

public async Task<string> DoAsyncThing(CancellationToken cancellationToken = default)
{
   byte[] buffer = new byte[1024];
   // We forgot to pass flow cancellationToken to ReadAsync
   int read = await _stream.ReadAsync(buffer, 0, buffer.Length);
   return Encoding.UTF8.GetString(buffer, 0, read);
}

✅ GOOD This example passes the CancellationToken into Stream.ReadAsync.

public async Task<string> DoAsyncThing(CancellationToken cancellationToken = default)
{
   byte[] buffer = new byte[1024];
   // This properly flows cancellationToken to ReadAsync
   int read = await _stream.ReadAsync(buffer, 0, buffer.Length, cancellationToken);
   return Encoding.UTF8.GetString(buffer, 0, read);
}

Cancelling uncancellable operations

One of the coding patterns that appears when doing asynchronous programming is canceling an uncancellable operation. This usually means creating another task that completes when a timeout or CancellationToken fires, and then using Task.WhenAny to detect a complete or cancelled operation.

Using CancellationTokens

❌ BAD This example uses Task.Delay(-1, token) to create a Task that completes when the CancellationToken fires, but if it doesn't fire, there's no way to dispose of the CancellationTokenRegistration created inside of Task.Delay. This can lead to a memory leak.

public static async Task<T> WithCancellation<T>(this Task<T> task, CancellationToken cancellationToken)
{
    // There's no way to dispose of the registration
    var delayTask = Task.Delay(-1, cancellationToken);

    var resultTask = await Task.WhenAny(task, delayTask);
    if (resultTask == delayTask)
    {
        // Operation cancelled
        throw new OperationCanceledException();
    }

    return await task;
}

✅ GOOD This example disposes of the CancellationTokenRegistration when one of the Task(s) is complete.

public static async Task<T> WithCancellation<T>(this Task<T> task, CancellationToken cancellationToken)
{
    var tcs = new TaskCompletionSource<object>(TaskCreationOptions.RunContinuationsAsynchronously);

    // This disposes the registration as soon as one of the tasks trigger
    using (cancellationToken.Register(state =>
    {
        ((TaskCompletionSource<object>)state).TrySetResult(null);
    },
    tcs))
    {
        var resultTask = await Task.WhenAny(task, tcs.Task);
        if (resultTask == tcs.Task)
        {
            // Operation cancelled
            throw new OperationCanceledException(cancellationToken);
        }

        return await task;
    }
}

✅ GOOD Prefer Task.WaitAsync on .NET >= 6;

Using a timeout

❌ BAD This example does not cancel the timer even if the operation successfully completes. This means you could end up with lots of timers, which can flood the timer queue.

public static async Task<T> TimeoutAfter<T>(this Task<T> task, TimeSpan timeout)
{
    var delayTask = Task.Delay(timeout);

    var resultTask = await Task.WhenAny(task, delayTask);
    if (resultTask == delayTask)
    {
        // Operation cancelled
        throw new OperationCanceledException();
    }

    return await task;
}

✅ GOOD This example cancels the timer if the operation successfully completes.

public static async Task<T> TimeoutAfter<T>(this Task<T> task, TimeSpan timeout)
{
    using (var cts = new CancellationTokenSource())
    {
        var delayTask = Task.Delay(timeout, cts.Token);

        var resultTask = await Task.WhenAny(task, delayTask);
        if (resultTask == delayTask)
        {
            // Operation cancelled
            throw new OperationCanceledException();
        }
        else
        {
            // Cancel the timer task so that it does not fire
            cts.Cancel();
        }

        return await task;
    }
}

✅ GOOD Prefer Task.WaitAsync on .NET >= 6;

Always call FlushAsync on StreamWriter(s) or Stream(s) before calling Dispose

When writing to a Stream or StreamWriter, even if the asynchronous overloads are used for writing, the underlying data might be buffered. When data is buffered, disposing the Stream or StreamWriter via the Dispose method will synchronously write/flush, which results in blocking the thread and could lead to thread-pool starvation. Either use the asynchronous DisposeAsync method (for example via await using) or call FlushAsync before calling Dispose.

💡NOTE: This is only problematic if the underlying subsystem does IO.

❌ BAD This example ends up blocking the request by writing synchronously to the HTTP-response body.

app.Run(async context =>
{
    // The implicit Dispose call will synchronously write to the response body
    using (var streamWriter = new StreamWriter(context.Response.Body))
    {
        await streamWriter.WriteAsync("Hello World");
    }
});

✅ GOOD This example asynchronously flushes any buffered data while disposing the StreamWriter.

app.Run(async context =>
{
    // The implicit AsyncDispose call will flush asynchronously
    await using (var streamWriter = new StreamWriter(context.Response.Body))
    {
        await streamWriter.WriteAsync("Hello World");
    }
});

✅ GOOD This example asynchronously flushes any buffered data before disposing the StreamWriter.

app.Run(async context =>
{
    using (var streamWriter = new StreamWriter(context.Response.Body))
    {
        await streamWriter.WriteAsync("Hello World");

        // Force an asynchronous flush
        await streamWriter.FlushAsync();
    }
});

Prefer async/await over directly returning Task

There are benefits to using the async/await keyword instead of directly returning the Task:

  • Asynchronous and synchronous exceptions are normalized to always be asynchronous.
  • The code is easier to modify (consider adding a using, for example).
  • Diagnostics of asynchronous methods are easier (debugging hangs etc).
  • Exceptions thrown will be automatically wrapped in the returned Task instead of surprising the caller with an actual exception.
  • Async locals will not leak out of async methods. If you set an async local in a non-async method, it will "leak" out of that call.

❌ BAD This example directly returns the Task to the caller.

public Task<int> DoSomethingAsync()
{
    return CallDependencyAsync();
}

✅ GOOD This example uses async/await instead of directly returning the Task.

public async Task<int> DoSomethingAsync()
{
    return await CallDependencyAsync();
}

💡NOTE: There are performance considerations when using an async state machine over directly returning the Task. It's always faster to directly return the Task since it does less work but you end up changing the behavior and potentially losing some of the benefits of the async state machine.

AsyncLocal<T>

Async locals are a way to store/retrieve ambient state throughout an application. This can be a very useful alternative to flowing explicit state everywhere, especially through call sites that you do not have much control over. While it is powerful, it is also dangerous if used incorrectly. Async locals are attached to the execution context which flows everywhere implicitly. Disabling execution context flow requires the use of advanced APIs (typically prefixed with the Unsafe name). As such, there's very little control over what code will attempt to access these values.

Creating an AsyncLocal<T>

If you can avoid async locals, do so by explicitly passing state around or using techniques like inversion of control.

If you cannot avoid it, it's best to make sure that anything put into an async local is:

  1. Not disposable
  2. Immutable/read-only/thread-safe

Let's look at 2 examples:

  1. ❌ BAD A disposable object stored in an async local
using (var thing = new DisposableThing())
{
    // Make the disposable object available ambiently
    DisposableThing.Current = thing;

    Dispatch();

    // We're about to dispose the object so make sure nobody else captures this instance
    DisposableThing.Current = null;
}

void Dispatch()
{
    // Task.Run will capture the current execution context (which means async locals are captured in the callback)
    _ = Task.Run(async () =>
    {
        // Delay for a second then log
        await Task.Delay(1000);

        Log();
    });
}

void Log()
{
    try
    {
        // Get the current value and make sure it's not null before reading the value
        var thing = DisposableThing.Current;
        if (thing is not null)
        {
            Console.WriteLine($"Logging ambient value {thing.Value}");
        }
    }
    catch (Exception ex)
    {
        Console.WriteLine(ex);
    }
}

Console.ReadLine();

class DisposableThing : IDisposable
{
    private static readonly AsyncLocal<DisposableThing?> _current = new();

    private bool _disposed;

    public static DisposableThing? Current
    {
        get => _current.Value;
        set
        {
            _current.Value = value;
        }
    }

    public int Value
    {
        get
        {
            if (_disposed) throw new ObjectDisposedException(GetType().FullName);
            return 1;
        }
    }

    public void Dispose()
    {
        _disposed = true;
    }
}

This above example will always result in an ObjectDisposedException being thrown. Even though the Log method defensively checks for null before logging the value, it has a reference to the disposed of DisposableThing. Setting the AsyncLocal<DisposableThing> to null does not affect the code inside of Log, this is because the execution context is copy on write. This means that all future reads DisposableThing.Current will be null, but it won't affect any of the previous reads.

When we set DisposableThing.Current = null; we are making a new execution context, not mutating the one that was captured by Task.Run. To get a better understanding of this run the following code:

DisposableThing.Current = new DisposableThing();

Console.WriteLine("After setting thing " + ExecutionContext.Capture().GetHashCode());

DisposableThing.Current = null;

Console.WriteLine("After setting Current to null " + ExecutionContext.Capture().GetHashCode());

The hash code of the execution context is different each time we set a new value.

⚠️ It might be tempting to update the logic in DisposableThing.Current to mutate the original execution context instead of setting the async local directly (StrongBox<T> is a reference type that stores the underlying T in a mutable field):

class DisposableThing : IDisposable
{
    private static readonly AsyncLocal<StrongBox<DisposableThing?>> _current = new();

    private bool _disposed;

    public static DisposableThing? Current
    {
        get => _current.Value?.Value;
        set
        {
            var box = _current.Value;
            if (box is not null)
            {
                // Mutate the value in any execution context that was copied
                box.Value = null;
            }

            if (value is not null)
            {
                _current.Value = new StrongBox<DisposableThing?>(value);
            }
        }
    }

    public int Value
    {
        get
        {
            if (_disposed) throw new ObjectDisposedException(GetType().FullName);
            return 1;
        }
    }

    public void Dispose()
    {
        _disposed = true;
    }
}

This will have the desired effect and will set the value to null in any execution context that references this async local value.

DisposableThing.Current = new DisposableThing();

Console.WriteLine("After setting thing " + ExecutionContext.Capture().GetHashCode());

DisposableThing.Current = null;

Console.WriteLine("After setting Current to null " + ExecutionContext.Capture().GetHashCode());

⚠️ While this looks attractive, the reference to DisposableThing.Current might have still been captured before the value was set to null:

void Dispatch()
{
    // Task.Run will capture the current execution context (which means async locals are captured in the callback)
    _ = Task.Run(async () =>
    {
        // Get the current reference
        var current = DisposableThing.Current;

        // Delay for a second then log
        await Task.Delay(1000);

        Log(current);
    });
}

void Log(DisposableThing thing)
{
    try
    {
        Console.WriteLine($"Logging ambient value {thing.Value}");
    }
    catch (Exception ex)
    {
        Console.WriteLine(ex);
    }
}

There's a race condition between the capture of the DisposableThing, the disposal of DisposableThing and setting DisposableThing.Current it to null. In the end, the code is unreliable and may fail at random. Don't store disposable objects in async locals.

  1. ❌ BAD A non-thread-safe object stored in an async local
AmbientValues.Current = new Dictionary<int, string>();

Parallel.For(0, 10, i =>
{
    AmbientValues.Current[i] = "processing";
    LogCurrentValues();
    AmbientValues.Current[i] = "done";
});


void LogCurrentValues()
{
    foreach (var pair in AmbientValues.Current)
    {
        Console.WriteLine(pair);
    }
}

class AmbientValues
{
    private static readonly AsyncLocal<Dictionary<int, string>> _current = new();

    public static Dictionary<int, string> Current
    {
        get => _current.Value!;
        set => _current.Value = value;
    }
}

The above example stores a normal Dictionary<int, string> in an async local and does some parallel processing on it. While this may be obvious from the above example, async locals allow arbitrary code on arbitrary threads to access the execution context and thus any async locals associated with said context. As a result, it is important to assume that data can be accessed concurrently and should be made thread-safe as a result.

class AmbientValues
{
    private static readonly AsyncLocal<ConcurrentDictionary<int, string>> _current = new();

    public static ConcurrentDictionary<int, string> Current
    {
        get => _current.Value!;
        set => _current.Value = value;
    }
}

✅ GOOD The above uses a ConcurrentDictionary<int, string> which is thread safe.

Don't leak your AsyncLocal<T>

Async locals flow across awaits automatically and can be captured by any API that explicitly calls ExecutionContext.Capture. The latter can lead to memory leaks in certain situations.

Common APIs that capture the ExecutionContext

APIs that run user callbacks usually capture the current execution context in order to preserve async locals between callback registration and execution. Here are examples of some APIs that do this:

  • Timer
  • CancellationToken.Register
  • new FileSystemWatcher
  • SocketAsyncEventArgs
  • Task.Run
  • ThreadPool.QueueUserWorkItem

❌ BAD Here's an example of an execution context leak that causes memory pressure because of a lifetime mismatch between the API capturing the execution context, and the lifetime of the data stored in the async local.

using System.Collections.Concurrent;

// Singleton cache
var cache = new NumberCache(TimeSpan.FromHours(1));

var executionContext = ExecutionContext.Capture();

// Simulate 10000 concurrent requests
Parallel.For(0, 10000, i =>
{
    // Restore the initial ExecutionContext per "request"
    ExecutionContext.Restore(executionContext!);

    ChunkyObject.Current = new ChunkyObject();

    cache.Add(i);
});

Console.WriteLine("Before GC: " + BytesAsString(GC.GetGCMemoryInfo().HeapSizeBytes));
Console.ReadLine();

GC.Collect();
GC.WaitForPendingFinalizers();

Console.WriteLine("After GC: " + BytesAsString(GC.GetGCMemoryInfo().HeapSizeBytes));
Console.ReadLine();

static string BytesAsString(long bytes)
{
    string[] suffix = { "B", "KB", "MB", "GB", "TB" };
    int i;
    double doubleBytes = 0;

    for (i = 0; bytes / 1024 > 0; i++, bytes /= 1024)
    {
        doubleBytes = bytes / 1024.0;
    }

    return string.Format("{0:0.00} {1}", doubleBytes, suffix[i]);
}

public class NumberCache
{
    private readonly ConcurrentDictionary<int, CancellationTokenSource> _cache = new ConcurrentDictionary<int, CancellationTokenSource>();
    private TimeSpan _timeSpan;

    public NumberCache(TimeSpan timeSpan)
    {
        _timeSpan = timeSpan;
    }

    public void Add(int key)
    {
        var cts = _cache.GetOrAdd(key, _ => new CancellationTokenSource());
        // Delete entry on expiration
        cts.Token.Register((_, _) => _cache.TryRemove(key, out _), null);

        // Start count down
        cts.CancelAfter(_timeSpan);
    }
}

class ChunkyObject
{
    private static readonly AsyncLocal<ChunkyObject?> _current = new();

    // Stores lots of data (but it should be gen0)
    private readonly string _data = new string('A', 1024 * 32);

    public static ChunkyObject? Current
    {
        get => _current.Value;
        set => _current.Value = value;
    }

    public string Data => _data;
}

The above example has a singleton NumberCache that stores numbers for an hour. We have a ChunkyObject which stores a 32K string in a field, and has an async local so that any code running may access the current ChunkyObject. This object should be collected when the GC runs, but instead, we're implicitly capturing the ChunkyObject in the NumberCache via CancellationToken.Register.

Instead of just caching the number and a CancellationTokenSource, we're implicitly capturing and storing all async locals attached to the current execution context for an hour!

Try running the sample locally. Running this on my machine reports numbers like this:

Before GC: 654.65 MB
After GC: 659.68 MB

Here's a look at the heap with those objects. You can see we have stored 10,000 ChunkyObjects, strings rooted by those chunky objects. The object graph looks like CancellationTokenSource -> ExecutionContext -> AsyncLocalValueMap -> ChunkObject -> string.

image

With one small tweak to this code, we can avoid the implicit execution context capture.

✅ GOOD Use CancellationToken.UnsafeRegister to avoid capturing the execution context and any async locals as part of the NumberCache:

public class NumberCache
{
    private readonly ConcurrentDictionary<int, CancellationTokenSource> _cache = new ConcurrentDictionary<int, CancellationTokenSource>();
    private TimeSpan _timeSpan;

    public NumberCache(TimeSpan timeSpan)
    {
        _timeSpan = timeSpan;
    }

    public void Add(int key)
    {
        var cts = _cache.GetOrAdd(key, _ => new CancellationTokenSource());
        // Delete entry on expiration
        cts.Token.UnsafeRegister((_, _) => _cache.TryRemove(key, out _), null);

        // Start count down
        cts.CancelAfter(_timeSpan);
    }
}

The GC numbers after this change:

Before GC: 10.32 MB
After GC: 5.10 MB

The heap looks like we'd expect. There's no execution context capture, so the ChunkyObject isn't stored.

image

💡 NOTE: You have NO control over how APIs decide to store the execution context, but with this understanding, you should be able to minimize memory leaks by clearing the memory using the technique described in Creating an AsyncLocal<T> section.

using System.Collections.Concurrent;

// Singleton cache
var cache = new NumberCache(TimeSpan.FromHours(1));

var executionContext = ExecutionContext.Capture();

// Simulate 10000 concurrent requests
Parallel.For(0, 10000, i =>
{
    // Restore the initial ExecutionContext per "request"
    ExecutionContext.Restore(executionContext!);

    ChunkyObject.Current = new ChunkyObject();

    cache.Add(i);

    // Null out the chunky object so the GC can release the memory
    ChunkyObject.Current = default;
});

class ChunkyObject
{
    private static readonly AsyncLocal<StrongBox<ChunkyObject?>> _current = new();

    // Stores lots of data (but it should be gen0)
    private readonly string _data = new string('A', 1024 * 32);

    public static ChunkyObject? Current
    {
        get => _current.Value?.Value;
        set
        {
            var box = _current.Value;
            if (box is not null)
            {
                // Mutate the value in any execution context that was copied
                box.Value = null;
            }

            if (value is not null)
            {
                _current.Value = new StrongBox<ChunkyObject?>(value);
            }
        }
    }

    public string Data => _data;
}

This technique reduces the heap memory significantly:

Before GC: 7.91 MB
After GC: 5.66 MB

The execution context is storing StrongBox<ChunkyObject> with a null reference to the ChunkyObject. This is technically still a "leak" but we've reduced the impact significantly. Here's a look at the memory profile showing objects with 10,000 allocations (the number of requests we created). You can see the GC has collected ChunkObject instances but there are still 10,000 references to StrongBox<ChunkyObject>.

image

Avoid setting AsyncLocal<T> values outside of async methods

Async methods have a special behavior for async locals that makes sure values do not propagate outside of the async method.

❌ BAD Avoid setting async local values outside of async methods:

var local = new AsyncLocal<int>();
MethodA();
Console.WriteLine(local.Value);

void MethodA()
{
    local.Value = 1;
    MethodB();
    Console.WriteLine(local.Value);
}

void MethodB()
{
    local.Value = 2;
    Console.WriteLine(local.Value);
}

The above prints 2, 2, 2. The execution context mutations are being propagated outside of the method. This can lead to extremely confusing behavior and hard-to-track down bugs.

✅ GOOD Set async locals in async methods:

var local = new AsyncLocal<int>();
await MethodA();
Console.WriteLine(local.Value);

async Task MethodA()
{
    local.Value = 1;
    await MethodB();
    Console.WriteLine(local.Value);
}

async Task MethodB()
{
    local.Value = 2;
    Console.WriteLine(local.Value);
}

The above will print 2, 1, 0. This is because the async method restores the original execution context on exit.

ConfigureAwait

TBD

Scenarios

The above tries to distill general guidance but doesn't do justice to the kinds of real-world situations that cause code like this to be written in the first place (bad code). This section tries to take concrete examples from real applications and turn them into something simple to help you relate these problems to existing codebases.

Timer callbacks

❌ BAD The Timer callback is void-returning and we have asynchronous work to execute. This example uses async void to accomplish it and as a result, can crash the process if an exception occurs.

public class Pinger
{
    private readonly Timer _timer;
    private readonly HttpClient _client;
    
    public Pinger(HttpClient client)
    {
        _client = client;
        _timer = new Timer(Heartbeat, null, 1000, 1000);
    }

    public async void Heartbeat(object state)
    {
        await _client.GetAsync("http://mybackend/api/ping");
    }
}

❌ BAD This attempts to block the Timer callback. This may result in thread-pool starvation and is an example of sync over async

public class Pinger
{
    private readonly Timer _timer;
    private readonly HttpClient _client;
    
    public Pinger(HttpClient client)
    {
        _client = client;
        _timer = new Timer(Heartbeat, null, 1000, 1000);
    }

    public void Heartbeat(object state)
    {
        _client.GetAsync("http://mybackend/api/ping").GetAwaiter().GetResult();
    }
}

✅ GOOD This example uses an async Task-based method and discards the Task in the Timer callback. If this method fails, it will not crash the process. Instead, it will fire the TaskScheduler.UnobservedTaskException event.

public class Pinger
{
    private readonly Timer _timer;
    private readonly HttpClient _client;
    
    public Pinger(HttpClient client)
    {
        _client = client;
        _timer = new Timer(Heartbeat, null, 1000, 1000);
    }

    public void Heartbeat(object state)
    {
        // Discard the result
        _ = DoAsyncPing();
    }

    private async Task DoAsyncPing()
    {
        await _client.GetAsync("http://mybackend/api/ping");
    }
}

✅ GOOD This example uses the new PeriodicTimer introduced in .NET 6:

public class Pinger : IDisposable
{
    private readonly PeriodicTimer _timer;
    private readonly HttpClient _client;

    public Pinger(HttpClient client)
    {
        _client = client;
        _timer = new PeriodicTimer(TimeSpan.FromSeconds(1));
        _ = Task.Run(DoAsyncPings);
    }

    public void Dispose()
    {
        _timer.Dispose();
    }

    private async Task DoAsyncPings()
    {
        while (await _timer.WaitForNextTickAsync())
        {
            // TODO: Handle exceptions
            await _client.GetAsync("http://mybackend/api/ping");
        }
    }
}

Implicit async void delegates

Imagine a BackgroundQueue with a FireAndForget that takes a callback. This method will execute the callback at some time in the future.

❌ BAD This will force callers to either block in the callback or use an async void delegate.

public class BackgroundQueue
{
    public static void FireAndForget(Action action) { }
}

❌ BAD This calling code is creating an async void method implicitly. The compiler fully supports this today.

public class Program
{
    public void Main(string[] args)
    {
        var httpClient = new HttpClient();
        BackgroundQueue.FireAndForget(async () =>
        {
            await httpClient.GetAsync("http://pinger/api/1");
        });
        
        Console.ReadLine();
    }
}

✅ GOOD This BackgroundQueue implementation offers both sync and async callback overloads.

public class BackgroundQueue
{
    public static void FireAndForget(Action action) { }
    public static void FireAndForget(Func<Task> action) { }
}

ConcurrentDictionary.GetOrAdd

It's pretty common to cache the result of an asynchronous operation and ConcurrentDictionary is a good data structure for doing that. GetOrAdd is a convenience API for trying to get an item if it's already there or adding it if it isn't. The callback is synchronous so it's tempting to write code that uses Task.Result to produce the value of an asynchronous process but that can lead to thread-pool starvation.

❌ BAD This may result in thread-pool starvation since we're blocking the request thread if the person data is not cached.

public class PersonController : Controller
{
   private AppDbContext _db;
   
   // This cache needs expiration
   private static ConcurrentDictionary<int, Person> _cache = new ConcurrentDictionary<int, Person>();
   
   public PersonController(AppDbContext db)
   {
      _db = db;
   }
   
   public IActionResult Get(int id)
   {
       var person = _cache.GetOrAdd(id, (key) => _db.People.FindAsync(key).Result);
       return Ok(person);
   }
}

✅ GOOD This implementation won't result in thread-pool starvation since we're storing a task instead of the result itself.

⚠️ ConcurrentDictionary.GetOrAdd, when accessed concurrently, may run the value-constructing delegate multiple times. This can result in needlessly kicking off the same potentially expensive computation multiple times.

public class PersonController : Controller
{
   private AppDbContext _db;
   
   // This cache needs expiration
   private static ConcurrentDictionary<int, Task<Person>> _cache = new ConcurrentDictionary<int, Task<Person>>();
   
   public PersonController(AppDbContext db)
   {
      _db = db;
   }
   
   public async Task<IActionResult> Get(int id)
   {
       var person = await _cache.GetOrAdd(id, (key) => _db.People.FindAsync(key));
       return Ok(person);
   }
}

✅ GOOD This implementation prevents the delegate from being executed multiple times, by using the async lazy pattern: even if construction of the AsyncLazy instance happens multiple times ("cheap" operation), the delegate will be called only once.

public class PersonController : Controller
{
   private AppDbContext _db;
   
   // This cache needs expiration
   private static ConcurrentDictionary<int, AsyncLazy<Person>> _cache = new ConcurrentDictionary<int, AsyncLazy<Person>>();
   
   public PersonController(AppDbContext db)
   {
      _db = db;
   }
   
   public async Task<IActionResult> Get(int id)
   {
       var person = await _cache.GetOrAdd(id, (key) => new AsyncLazy<Person>(() => _db.People.FindAsync(key))).Value;
       return Ok(person);
   }
   
   private class AsyncLazy<T> : Lazy<Task<T>>
   {
      public AsyncLazy(Func<Task<T>> valueFactory) : base(valueFactory)
      {
      }
   }
}

Constructors

Constructors are synchronous. If you need to initialize some logic that may be asynchronous, there are a couple of patterns for dealing with this.

Here's an example of using a client API that needs to connect asynchronously before use.

public interface IRemoteConnectionFactory
{
   Task<IRemoteConnection> ConnectAsync();
}

public interface IRemoteConnection
{
    Task PublishAsync(string channel, string message);
    Task DisposeAsync();
}

❌ BAD This example uses Task.Result to get the connection in the constructor. This could lead to thread-pool starvation and deadlocks.

public class Service : IService
{
    private readonly IRemoteConnection _connection;
    
    public Service(IRemoteConnectionFactory connectionFactory)
    {
        _connection = connectionFactory.ConnectAsync().Result;
    }
}

✅ GOOD This implementation uses a static factory pattern in order to allow asynchronous construction:

public class Service : IService
{
    private readonly IRemoteConnection _connection;

    private Service(IRemoteConnection connection)
    {
        _connection = connection;
    }

    public static async Task<Service> CreateAsync(IRemoteConnectionFactory connectionFactory)
    {
        return new Service(await connectionFactory.ConnectAsync());
    }
}

WindowsIdentity.RunImpersonated

This API runs the specified action as the impersonated Windows identity. An asynchronous version of the callback was introduced in .NET 5.0.

❌ BAD This example tries to execute the query asynchronously, and then wait for it outside of the call to RunImpersonated. This will throw because the query might be executing outside of the impersonation context.

public async Task<IEnumerable<Product>> GetDataImpersonatedAsync(SafeAccessTokenHandle safeAccessTokenHandle)
{
    Task<IEnumerable<Product>> products = null;
    WindowsIdentity.RunImpersonated(
        safeAccessTokenHandle,
        context =>
        {
            products = _db.QueryAsync("SELECT Name from Products");
        });
    return await products;
}

❌ BAD This example uses Task.Result to execute the query synchronously (sync over async). This could lead to thread-pool starvation and deadlocks.

public IEnumerable<Product> GetDataImpersonated(SafeAccessTokenHandle safeAccessTokenHandle)
{
    return WindowsIdentity.RunImpersonated(
        safeAccessTokenHandle,
        context => _db.QueryAsync("SELECT Name from Products").Result);
}

✅ GOOD This example awaits the result of RunImpersonated (the delegate is Func<Task<IEnumerable<Product>>> in this case). It is the recommended practice in frameworks earlier than .NET 5.0.

public async Task<IEnumerable<Product>> GetDataImpersonatedAsync(SafeAccessTokenHandle safeAccessTokenHandle)
{
    return await WindowsIdentity.RunImpersonated(
        safeAccessTokenHandle, 
        context => _db.QueryAsync("SELECT Name from Products"));
}

✅ GOOD This example uses the asynchronous RunImpersonatedAsync function and awaits its result. It is available in .NET 5.0 or newer.

public async Task<IEnumerable<Product>> GetDataImpersonatedAsync(SafeAccessTokenHandle safeAccessTokenHandle)
{
    return await WindowsIdentity.RunImpersonatedAsync(
        safeAccessTokenHandle, 
        context => _db.QueryAsync("SELECT Name from Products"));
}