Permalink
Switch branches/tags
Nothing to show
Find file Copy path
407 lines (345 sloc) 29.6 KB

Async Streams

  • Proposed
  • Prototype
  • Implementation
  • Specification

Summary

C# has support for iterator methods and async methods, but no support for a method that is both an iterator and an async method. We should rectify this by allowing for await to be used in a new form of async iterator, one that returns an IAsyncEnumerable<T> or IAsyncEnumerator<T> rather than an IEnumerable<T> or IEnumerator<T>, with IAsyncEnumerable<T> consumable in a new await foreach. An IAsyncDisposable interface is also used to enable asynchronous cleanup.

Related discussion

Detailed design

Interfaces

IAsyncDisposable

There has been much discussion of IAsyncDisposable (e.g. https://github.com/dotnet/roslyn/issues/114) and whether it's a good idea. However, it's a required concept to add in support of async iterators. Since finally blocks may contain awaits, and since finally blocks need to be run as part of disposing of iterators, we need async disposal. It's also just generally useful any time cleaning up of resources might take any period of time, e.g. closing files (requiring flushes), deregistering callbacks and providing a way to know when deregistration has completed, etc.

The following interface is added to the core .NET libraries (e.g. System.Private.CoreLib / System.Runtime):

namespace System
{
    public interface IAsyncDisposable
    {
        ValueTask DisposeAsync();
    }
}

As with Dispose, invoking DisposeAsync multiple times is acceptable, and subsequent invocations after the first should be treated as nops, returning a synchronously completed successful task (DisposeAsync need not be thread-safe, though, and need not support concurrent invocation). Further, types may implement both IDisposable and IAsyncDisposable, and if they do, it's similarly acceptable to invoke Dispose and then DisposeAsync or vice versa, but only the first should be meaningful and subsequent invocations of either should be a nop. As such, if a type does implement both, consumers are encouraged to call once and only once the more relevant method based on the context, Dispose in synchronous contexts and DisposeAsync in asynchronous ones.

(I'm leaving discussion of how IAsyncDisposable interacts with using to a separate discussion. And coverage of how it interacts with foreach is handled later in this proposal.)

Alternatives considered:

  • DisposeAsync accepting a CancellationToken: while in theory it makes sense that anything async can be canceled, disposal is about cleanup, closing things out, free'ing resources, etc., which is generally not something that should be canceled; cleanup is still important for work that's canceled. The same CancellationToken that caused the actual work to be canceled would typically be the same token passed to DisposeAsync, making DisposeAsync worthless because cancellation of the work would cause DisposeAsync to be a nop. If someone wants to avoid being blocked waiting for disposal, they can avoid waiting on the resulting ValueTask, or wait on it only for some period of time.
  • DisposeAsync returning a Task: Now that a non-generic ValueTask exists and can be constructed from an IValueTaskSource, returning ValueTask from DisposeAsync allows an existing object to be reused as the promise representing the eventual async completion of DisposeAsync, saving a Task allocation in the case where DisposeAsync completes asynchronously.
  • Configuring DisposeAsync with a bool continueOnCapturedContext (ConfigureAwait): While there may be issues related to how such a concept is exposed to using, foreach, and other language constructs that consume this, from an interface perspective it's not actually doing any await'ing and there's nothing to configure... consumers of the ValueTask can consume it however they wish.
  • IAsyncDisposable inheriting IDisposable: Since only one or the other should be used, it doesn't make sense to force types to implement both.
  • IDisposableAsync instead of IAsyncDisposable: We've been following the naming that things/types are an "async something" whereas operations are "done async", so types have "Async" as a prefix and methods have "Async" as a suffix.

IAsyncEnumerable / IAsyncEnumerator

Two interfaces are added to the core .NET libraries:

namespace System.Collections.Generic
{
    public interface IAsyncEnumerable<out T>
    {
        IAsyncEnumerator<T> GetAsyncEnumerator();
    }

    public interface IAsyncEnumerator<out T> : IAsyncDisposable
    {
        ValueTask<bool> MoveNextAsync();
        T Current { get; }
    }
}

Typical consumption (without additional language features) would look like:

IAsyncEnumerator<T> enumerator = enumerable.GetAsyncEnumerator();
try
{
    while (await enumerator.MoveNextAsync())
    {
        Use(enumerator.Current);
    }
}
finally { await enumerator.DisposeAsync(); }

Discarded options considered:

  • Task<bool> MoveNextAsync(); T current { get; }: Using Task<bool> would support using a cached task object to represent synchronous, successful MoveNextAsync calls, but an allocation would still be required for asynchronous completion. By returning ValueTask<bool>, we enable the enumerator object to itself implement IValueTaskSource<bool> and be used as the backing for the ValueTask<bool> returned from MoveNextAsync, which in turn allows for significantly reduced overheads.
  • ValueTask<(bool, T)> MoveNextAsync();: It's not only harder to consume, but it means that T can no longer be covariant.
  • ValueTask<T?> TryMoveNextAsync();: Not covariant.
  • Task<T?> TryMoveNextAsync();: Not covariant, allocations on every call, etc.
  • ITask<T?> TryMoveNextAsync();: Not covariant, allocations on every call, etc.
  • ITask<(bool,T)> TryMoveNextAsync();: Not covariant, allocations on every call, etc.
  • Task<bool> TryMoveNextAsync(out T result);: The out result would need to be set when the operation returns synchronously, not when it asynchronously completes the task potentially sometime long in the future, at which point there'd be no way to communicate the result.
  • IAsyncEnumerator<T> not implementing IAsyncDisposable: We could choose to separate these. However, doing so complicates certain other areas of the proposal, as code must then be able to deal with the possibility that an enumerator doesn't provide disposal, which makes it difficult to write pattern-based helpers. Further, it will be common for enumerators to have a need for disposal (e.g. any C# async iterator that has a finally block, most things enumerating data from a network connection, etc.), and if one doesn't, it is simple to implement the method purely as public ValueTask DisposeAsync() => default(ValueTask); with minimal additional overhead.

Viable alternative:

namespace System.Collections.Generic
{
    public interface IAsyncEnumerable<out T>
    {
        IAsyncEnumerator<T> GetAsyncEnumerator();
    }

    public interface IAsyncEnumerator<out T> : IAsyncDisposable
    {
        ValueTask<bool> WaitForNextAsync();
        T TryGetNext(out bool success);
    }
}

TryGetNext is used in an inner loop to consume items with a single interface call as long as they're available synchronously. When the next item can't be retrieved synchronously, it returns false, and any time it returns false, a caller must subsequently invoke WaitForNextAsync to either wait for the next item to be available or to determine that there will never be another item. Typical consumption (without additional language features) would look like:

IAsyncEnumerator<T> enumerator = enumerable.GetAsyncEnumerator();
try
{
    while (await enumerator.WaitForNextAsync())
    {
        while (true)
        {
            int item = enumerator.TryGetNext(out bool success);
            if (!success) break;
            Use(item);
        }
    }
}
finally { await enumerator.DisposeAsync(); }

The advantage of this is two-fold, one minor and one major:

  • Minor: Allows for an enumerator to support multiple consumers. There may be scenarios where it's valuable for an enumerator to support multiple concurrent consumers. That can't be achieved when MoveNextAsync and Current are separate such that an implementation can't make their usage atomic. In contrast, this approach provides a single method TryGetNext that supports pushing the enumerator forward and getting the next item, so the enumerator can enable atomicity if desired. However, it's likely that such scenarios could also be enabled by giving each consumer its own enumerator from a shared enumerable. Further, we don't want to enforce that every enumerator support concurrent usage, as that would add non-trivial overheads to the majority case that doesn't require it, which means a consumer of the interface generally couldn't rely on this any way.
  • Major: Performance. The MoveNextAsync/Current approach requires two interface calls per operation, whereas the best case for WaitForNextAsync/TryGetNext is that most iterations complete synchronously, enabling a tight inner loop with TryGetNext, such that we only have one interface call per operation. This can have a measurable impact in situations where the interface calls dominate the computation.

However, there are non-trivial downsides, including significantly increased complexity when consuming these manually, and an increased chance of introducing bugs when using them. And while the performance benefits show up in microbenchmarks, we don't believe they'll be impactful in the vast majority of real usage. If it turns out they are, we can introduce a second set of interfaces in a light-up fashion.

Discarded options considered:

  • ValueTask<bool> WaitForNextAsync(); bool TryGetNext(out T result);: out parameters can't be covariant. There's also a small impact here (an issue with the try pattern in general) that this likely incurs a runtime write barrier for reference type results.

Cancellation

There are several possible approaches to supporting cancellation:

  1. IAsyncEnumerable<T>/IAsyncEnumerator<T> are cancellation-agnostic: CancellationToken doesn't appear anywhere. Cancellation is achieved by logically baking the CancellationToken into the enumerable and/or enumerator in whatever manner is appropriate, e.g. when calling an iterator, passing the CancellationToken as an argument to the iterator method and using it in the body of the iterator, as is done with any other parameter.
  2. IAsyncEnumerator<T>.GetAsyncEnumerator(CancellationToken): You pass a CancellationToken to GetAsyncEnumerator, and subsequent MoveNextAsync operations respect it however it can.
  3. IAsyncEnumerator<T>.MoveNextAsync(CancellationToken): You pass a CancellationToken to each individual MoveNextAsync call.
  4. 1 && 2: You both embed CancellationTokens into your enumerable/enumerator and pass CancellationTokens into GetAsyncEnumerator.
  5. 1 && 3: You both embed CancellationTokens into your enumerable/enumerator and pass CancellationTokens into MoveNextAsync.

From a purely theoretical perspective, (5) is the most robust, in that (a) MoveNextAsync accepting a CancellationToken enables the most fine-grained control over what's canceled, and (b) CancellationToken is just any other type that can passed as an argument into iterators, embedded in arbitrary types, etc.

However, there are multiple problems with that approach:

  • How does a CancellationToken passed to GetAsyncEnumerator make it into the body of the iterator? We could expose a new iterator keyword that you could dot off of to get access to the CancellationToken passed to GetEnumerator, but a) that's a lot of additional machinery, b) we're making it a very first-class citizen, and c) the 99% case would seem to be the same code both calling an iterator and calling GetAsyncEnumerator on it, in which case it can just pass the CancellationToken as an argument into the method.
  • How does a CancellationToken passed to MoveNextAsync get into the body of the method? This is even worse, as if it's exposed off of an iterator local object, its value could change across awaits, which means any code that registered with the token would need to unregister from it prior to awaits and then re-register after; it's also potentially quite expensive to need to do such registering and unregistering in every MoveNextAsync call, regardless of whether implemented by the compiler in an iterator or by a developer manually.
  • How does a developer cancel a foreach loop? If it's done by giving a CancellationToken to an enumerable/enumerator, then either a) we need to support foreach'ing over enumerators, which raises them to being first-class citizens, and now you need to start thinking about an ecosystem built up around enumerators (e.g. LINQ methods) or b) we need to embed the CancellationToken in the enumerable anyway by having some WithCancellation extension method off of IAsyncEnumerable<T> that would store the provided token and then pass it into the wrapped enumerable's GetAsyncEnumerator when the GetAsyncEnumerator on the returned struct is invoked (ignoring that token). Or, you can just use the CancellationToken you have in the body of the foreach.
  • If/when query comprehensions are supported, how would the CancellationToken supplied to GetEnumerator or MoveNextAsync be passed into each clause? The easiest way would simply be for the clause to capture it, at which point whatever token is passed to GetAsyncEnumerator/MoveNextAsync is ignored.

Due to all of this, the simplest and most consistent solution is simply to do (1): IAsyncEnumerable<T>/IAsyncEnumerator<T> are cancellation-agnostic. If you want to cancel a foreach loop, you can use a CancellationToken in the body and in any methods you call:

CancellationToken ct = ...;
await foreach (var i in GetData())
{
    ct.ThrowIfCancellationRequested();
    await UseAsync(i, ct);
    ...
}

If you want to pass a CancellationToken into an iterator, you simply do so as an argument, just as with other async methods:

static async IAsyncEnumerable<T> GetData(CancellationToken cancellationToken = default)
{
    using (cancellationToken.Register(...))
    {
        ...
        await ...
        ...
    }
}
...
await foreach (T i in GetData(ct))
{
    ...
}

If you want to use a CancellationToken in a query comprehension, you just capture it:

CancellationToken ct = ...
IAsyncEnumerable<string> results = from url in source
                                   select DownloadAsync(url, ct);

Etc.

For now, we should pursue (1).

foreach

foreach will be augmented to support IAsyncEnumerable<T> in addition to its existing support for IEnumerable<T>. And it will support the equivalent of IAsyncEnumerable<T> as a pattern if the relevant members are exposed publicly, falling back to using the interface directly if not, in order to enable struct-based extensions that avoid allocating as well as using alternative awaitables as the return type of MoveNextAsync and DisposeAsync.

Syntax

Using the syntax:

foreach (var i in enumerable)

C# will continue to treat enumerable as a synchronous enumerable, such that even if it exposes the relevant APIs for async enumerables (exposing the pattern or implementing the interface), it will only consider the synchronous APIs.

To force foreach to instead only consider the asynchronous APIs, await is inserted as follows:

await foreach (var i in enumerable)

No syntax would be provided that would support using either the async or the sync APIs; the developer must choose based on the syntax used.

Discarded options considered:

  • foreach (var i in await enumerable): This is already valid syntax, and changing its meaning would be a breaking change. This means to await the enumerable, get back something synchronously iterable from it, and then synchronously iterate through that.
  • foreach (var i await in enumerable), foreach (var await i in enumerable), foreach (await var i in enumerable): These all suggest that we're awaiting the next item, but there are other awaits involved in foreach, in particular if the enumerable is an IAsyncDisposable, we will be await'ing its async disposal. That await is as the scope of the foreach rather than for each individual element, and thus the await keyword deserves to be at the foreach level. Further, having it associated with the foreach gives us a way to describe the foreach with a different term, e.g. a "await foreach". But more importantly, there's value in considering foreach syntax at the same time as using syntax, so that they remain consistent with each other, and using (await ...) is already valid syntax.
  • foreach await (var i in enumerable)

Still to consider:

  • foreach today does not support iterating through an enumerator. We expect it will be more common to have IAsyncEnumerator<T>s handed around, and thus it's tempting to support await foreach with both IAsyncEnumerable<T> and IAsyncEnumerator<T>. But once we add such support, it introduces the question of whether IAsyncEnumerator<T> is a first-class citizen, and whether we need to have overloads of combinators that operate on enumerators in addition to enumerables? Do we want to encourage methods to return enumerators rather than enumerables? We should continue to discuss this. If we decide we don't want to support it, we might want to introduce an extension method public static IAsyncEnumerable<T> AsEnumerable<T>(this IAsyncEnumerator<T> enumerator); that would allow an enumerator to still be foreach'd. If we decide we do want to support it, we'll need to also decide on whether the await foreach would be responsible for calling DisposeAsync on the enumerator, and the answer is likely "no, control over disposal should be handled by whoever called GetEnumerator."

Pattern-based Compilation

The compiler will bind to the pattern-based APIs if they exist, preferring those over using the interface (the pattern may be satisfied with instance methods or extension methods). The requirements for the pattern are:

  • The enumerable must expose a GetAsyncEnumerator method that may be called with no arguments and that returns an enumerator that meets the relevant pattern.
  • The enumerator must expose a MoveNextAsync method that may be called with no arguments and that returns something which may be awaited and whose GetResult() returns a bool.
  • The enumerator must also expose Current property whose getter returns a T representing the kind of data being enumerated.
  • The enumerator may optionally expose a DisposeAsync method that may be invoked with no arguments and that returns something that can be awaited and whose GetResult() returns void.

This code:

var enumerable = ...;
await foreach (T item in enumerable)
{
   ...
}

is translated to the equivalent of:

var enumerable = ...;
var enumerator = enumerable.GetAsyncEnumerator();
try
{
    while (await enumerator.MoveNextAsync())
    {
       T item = enumerator.Current;
       ...
    }
}
finally
{
    await enumerator.DisposeAsync(); // omitted, along with the try/finally, if the enumerator doesn't expose DisposeAsync
}

If the iterated type doesn't expose the right pattern, the interfaces will be used.

ConfigureAwait

This pattern-based compilation will allow ConfigureAwait to be used on all of the awaits, via a ConfigureAwait extension method:

await foreach (T item in enumerable.ConfigureAwait(false))
{
   ...
}

This will be based on types we'll add to .NET as well, likely to System.Threading.Tasks.Extensions.dll:

// Approximate implementation, omitting arg validation and the like
namespace System.Threading.Tasks
{
    public static class AsyncEnumerableExtensions
    {
        public static ConfiguredAsyncEnumerable<T> ConfigureAwait<T>(this IAsyncEnumerable<T> enumerable, bool continueOnCapturedContext) =>
            new ConfiguredAsyncEnumerable<T>(enumerable, continueOnCapturedContext);

        public struct ConfiguredAsyncEnumerable<T>
        {
            private readonly IAsyncEnumerable<T> _enumerable;
            private readonly bool _continueOnCapturedContext;

            internal ConfiguredAsyncEnumerable(IAsyncEnumerable<T> enumerable, bool continueOnCapturedContext)
            {
                _enumerable = enumerable;
                _continueOnCapturedContext = continueOnCapturedContext;
            }

            public ConfiguredAsyncEnumerator<T> GetAsyncEnumerator() =>
                new ConfiguredAsyncEnumerator<T>(_enumerable.GetAsyncEnumerator(), _continueOnCapturedContext);

            public struct Enumerator
            {
                private readonly IAsyncEnumerator<T> _enumerator;
                private readonly bool _continueOnCapturedContext;

                internal Enumerator(IAsyncEnumerator<T> enumerator, bool continueOnCapturedContext)
                {
                    _enumerator = enumerator;
                    _continueOnCapturedContext = continueOnCapturedContext;
                }

                public ConfiguredValueTaskAwaitable<bool> MoveNextAsync() =>
                    _enumerator.MoveNextAsync().ConfigureAwait(_continueOnCapturedContext);

                public T Current => _enumerator.Current;

                public ConfiguredValueTaskAwaitable DisposeAsync() =>
                    _enumerator.DisposeAsync().ConfigureAwait(_continueOnCapturedContext);
            }
        }
    }
}

Note that this approach will not enable ConfigureAwait to be used with pattern-based enumerables, but then again it's already the case that the ConfigureAwait is only exposed as an extension on Task/Task<T>/ValueTask/ValueTask<T> and can't be applied to arbitrary awaitable things, as it only makes sense when applied to Tasks (it controls a behavior implemented in Task's continuation support), and thus doesn't make sense when using a pattern where the awaitable things may not be tasks. Anyone returning awaitable things can provide their own custom behavior in such advanced scenarios.

(If we can come up with some way to support a scope- or assembly-level ConfigureAwait solution, then this won't be necessary.)

Async Iterators

The language / compiler will support producing IAsyncEnumerable<T>s and IAsyncEnumerator<T>s in addition to consuming them. Today the language supports writing an iterator like:

static IEnumerable<int> MyIterator()
{
    try
    {
        for (int i = 0; i < 100; i++)
        {
            Thread.Sleep(1000);
            yield return i;
        }
    }
    finally
    {
        Thread.Sleep(200);
        Console.WriteLine("finally");
    }
}

but await can't be used in the body of these iterators. We will add that support.

Syntax

The existing language support for iterators infers the iterator nature of the method based on whether it contains any yields. The same will be true for async iterators. Such async iterators will be demarcated and differentiated from synchronous iterators via adding async to the signature, and must then also have either IAsyncEnumerable<T> or IAsyncEnumerator<T> as its return type. For example, the above example could be written as an async iterator as follows:

static async IAsyncEnumerable<int> MyIterator()
{
    try
    {
        for (int i = 0; i < 100; i++)
        {
            await Task.Delay(1000);
            yield return i;
        }
    }
    finally
    {
        await Task.Delay(200);
        Console.WriteLine("finally");
    }
}

Alternatives considered:

  • Not using async in the signature: Using async is likely technically required by the compiler, as it uses it to determine whether await is valid in that context. But even if it's not required, we've established that await may only be used in methods marked as async, and it seems important to keep the consistency.
  • Enabling custom builders for IAsyncEnumerable<T>: That's something we could look at for the future, but the machinery is complicated and we don't support that for the synchronous counterparts.
  • Having an iterator keyword in the signature: Async iterators would use async iterator in the signature, and yield could only be used in async methods that included iterator; iterator would then be made optional on synchronous iterators. Depending on your perspective, this has the benefit of making it very clear by the signature of the method whether yield is allowed and whether the method is actually meant to return instances of type IAsyncEnumerable<T> rather than the compiler manufacturing one based on whether the code uses yield or not. But it is different from synchronous iterators, which don't and can't be made to require one. Plus some developers don't like the extra syntax. If we were designing it from scratch, we'd probably make this required, but at this point there's much more value in keeping async iterators close to sync iterators.

LINQ

There are over ~200 overloads of methods on the System.Linq.Enumerable class, all of which work in terms of IEnumerable<T>; some of these accept IEnumerable<T>, some of them produce IEnumerable<T>, and many do both. Adding LINQ support for IAsyncEnumerable<T> would likely entail duplicating all of these overloads for it, for another ~200. And since IAsyncEnumerator<T> is likely to be more common as a standalone entity in the asynchronous world than IEnumerator<T> is in the synchronous world, we could potentially need another ~200 overloads that work with IAsyncEnumerator<T>. Plus, a large number of the overloads deal with predicates (e.g. Where that takes a Func<T, bool>), and it may be desirable to have IAsyncEnumerable<T>-based overloads that deal with both synchronous and asynchronous predicates (e.g. Func<T, ValueTask<bool>> in addition to Func<T, bool>). While this isn't applicable to all of the now ~400 new overloads, a rough calculation is that it'd be applicable to half, which means another ~200 overloads, for a total of ~600 new methods.

That is a staggering number of APIs, with the potential for even more when extension libraries like Interactive Extensions (Ix) is considered. But Ix already has an implementation of many of these, and there doesn't seem to be a great reason to duplicate that work; we should instead help the community improve Ix and recommend it for when developers want to use LINQ with IAsyncEnumerable<T>.

There is also the issue of query comprehension syntax. The pattern-based nature of query comprehensions would allow them to "just work" with some operators, e.g. if Ix provides the following methods:

public static IAsyncEnumerable<TResult> Select<TSource, TResult>(this IAsyncEnumerable<TSource> source, Func<TSource, TResult> func);
public static IAsyncEnumerable<T> Where(this IAsyncEnumerable<T> source, Func<T, bool> func);

then this C# code will "just work":

IAsyncEnumerable<int> enumerable = ...;
IAsyncEnumerable<int> result = from item in enumerable
                               where item % 2 == 0
                               select item * 2;

However, there is no query comprehension syntax that supports using await in the clauses, so if Ix added, for example:

public static IAsyncEnumerable<TResult> Select<TSource, TResult>(this IAsyncEnumerable<TSource> source, Func<TSource, ValueTask<TResult>> func);

then this would "just work":

IAsyncEnumerable<string> result = from url in urls
                                  where item % 2 == 0
                                  select SomeAsyncMethod(item);

async ValueTask<int> SomeAsyncMethod(int item)
{
    await Task.Yield();
    return item * 2;
}

but there'd be no way to write it with the await inline in the select clause. As a separate effort, we could look into adding async { ... } expressions to the language, at which point we could allow them to be used in query comprehensions and the above could instead be written as:

IAsyncEnumerable<int> result = from item in enumerable
                               where item % 2 == 0
                               select async
                               {
                                   await Task.Yield();
                                   return item * 2;
                               };

or to enabling await to be used directly in expressions, such as by supporting async from. However, it's unlikely a design here would impact the rest of the feature set one way or the other, and this isn't a particularly high-value thing to invest in right now, so the proposal is to do nothing additional here right now.

Integration with other asynchronous frameworks

Integration with IObservable<T> and other asynchronous frameworks (e.g. reactive streams) would be done at the library level rather than at the language level. For example, all of the data from an IAsyncEnumerator<T> can be published to an IObserver<T> simply by await foreach'ing over the enumerator and OnNext'ing the data to the observer, so an AsObservable<T> extension method is possible. Consuming an IObservable<T> in a await foreach requires buffering the data (in case another item is pushed while the previous item is still being processing), but such a push-pull adapter can easily be implemented to enable an IObservable<T> to be pulled from with an IAsyncEnumerator<T>. Etc. Rx/Ix already provide prototypes of such implementations, and libraries like https://github.com/dotnet/corefx/tree/master/src/System.Threading.Channels provide various kinds of buffering data structures. The language need not be involved at this stage.