# Async Streams * [x] Proposed * [x] Prototype * [ ] Implementation * [ ] Specification ## Summary [summary]: #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` or `IAsyncEnumerator` rather than an `IEnumerable` or `IEnumerator`, with `IAsyncEnumerable` consumable in a new `await foreach`. An `IAsyncDisposable` interface is also used to enable asynchronous cleanup. ## Related discussion - https://github.com/dotnet/roslyn/issues/261 - https://github.com/dotnet/roslyn/issues/114 ## Detailed design [design]: #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 `await`s, 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): ```C# 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: ```C# namespace System.Collections.Generic { public interface IAsyncEnumerable { IAsyncEnumerator GetAsyncEnumerator(); } public interface IAsyncEnumerator : IAsyncDisposable { ValueTask MoveNextAsync(); T Current { get; } } } ``` Typical consumption (without additional language features) would look like: ```C# IAsyncEnumerator enumerator = enumerable.GetAsyncEnumerator(); try { while (await enumerator.MoveNextAsync()) { Use(enumerator.Current); } } finally { await enumerator.DisposeAsync(); } ``` Discarded options considered: - _`Task MoveNextAsync(); T current { get; }`_: Using `Task` 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`, we enable the enumerator object to itself implement `IValueTaskSource` and be used as the backing for the `ValueTask` 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 TryMoveNextAsync();`_: Not covariant. - _`Task TryMoveNextAsync();`_: Not covariant, allocations on every call, etc. - _`ITask TryMoveNextAsync();`_: Not covariant, allocations on every call, etc. - _`ITask<(bool,T)> TryMoveNextAsync();`_: Not covariant, allocations on every call, etc. - _`Task 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` 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: ```C# namespace System.Collections.Generic { public interface IAsyncEnumerable { IAsyncEnumerator GetAsyncEnumerator(); } public interface IAsyncEnumerator : IAsyncDisposable { ValueTask 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: ```C# IAsyncEnumerator 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 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`/`IAsyncEnumerator` 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.GetAsyncEnumerator(CancellationToken)`: You pass a `CancellationToken` to `GetAsyncEnumerator`, and subsequent `MoveNextAsync` operations respect it however it can. 3. `IAsyncEnumerator.MoveNextAsync(CancellationToken)`: You pass a `CancellationToken` to each individual `MoveNextAsync` call. 4. 1 && 2: You both embed `CancellationToken`s into your enumerable/enumerator and pass `CancellationToken`s into `GetAsyncEnumerator`. 5. 1 && 3: You both embed `CancellationToken`s into your enumerable/enumerator and pass `CancellationToken`s 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` 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`/`IAsyncEnumerator` 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: ```C# 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: ```C# static async IAsyncEnumerable 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: ```C# CancellationToken ct = ... IAsyncEnumerable results = from url in source select DownloadAsync(url, ct); ``` Etc. For now, we should pursue (1). ## foreach `foreach` will be augmented to support `IAsyncEnumerable` in addition to its existing support for `IEnumerable`. And it will support the equivalent of `IAsyncEnumerable` 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: ```C# 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: ```C# 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`s handed around, and thus it's tempting to support `await foreach` with both `IAsyncEnumerable` and `IAsyncEnumerator`. But once we add such support, it introduces the question of whether `IAsyncEnumerator` 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 AsEnumerable(this IAsyncEnumerator 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 `await`ed 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 `await`ed and whose `GetResult()` returns `void`. This code: ```C# var enumerable = ...; await foreach (T item in enumerable) { ... } ``` is translated to the equivalent of: ```C# 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: ```C# 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: ```C# // Approximate implementation, omitting arg validation and the like namespace System.Threading.Tasks { public static class AsyncEnumerableExtensions { public static ConfiguredAsyncEnumerable ConfigureAwait(this IAsyncEnumerable enumerable, bool continueOnCapturedContext) => new ConfiguredAsyncEnumerable(enumerable, continueOnCapturedContext); public struct ConfiguredAsyncEnumerable { private readonly IAsyncEnumerable _enumerable; private readonly bool _continueOnCapturedContext; internal ConfiguredAsyncEnumerable(IAsyncEnumerable enumerable, bool continueOnCapturedContext) { _enumerable = enumerable; _continueOnCapturedContext = continueOnCapturedContext; } public ConfiguredAsyncEnumerator GetAsyncEnumerator() => new ConfiguredAsyncEnumerator(_enumerable.GetAsyncEnumerator(), _continueOnCapturedContext); public struct Enumerator { private readonly IAsyncEnumerator _enumerator; private readonly bool _continueOnCapturedContext; internal Enumerator(IAsyncEnumerator enumerator, bool continueOnCapturedContext) { _enumerator = enumerator; _continueOnCapturedContext = continueOnCapturedContext; } public ConfiguredValueTaskAwaitable 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`/`ValueTask`/`ValueTask` 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`s and `IAsyncEnumerator`s in addition to consuming them. Today the language supports writing an iterator like: ```C# static IEnumerable 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 `yield`s. 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` or `IAsyncEnumerator` as its return type. For example, the above example could be written as an async iterator as follows: ```C# static async IAsyncEnumerable 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`_: 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` 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`; some of these accept `IEnumerable`, some of them produce `IEnumerable`, and many do both. Adding LINQ support for `IAsyncEnumerable` would likely entail duplicating all of these overloads for it, for another ~200. And since `IAsyncEnumerator` is likely to be more common as a standalone entity in the asynchronous world than `IEnumerator` is in the synchronous world, we could potentially need another ~200 overloads that work with `IAsyncEnumerator`. Plus, a large number of the overloads deal with predicates (e.g. `Where` that takes a `Func`), and it may be desirable to have `IAsyncEnumerable`-based overloads that deal with both synchronous and asynchronous predicates (e.g. `Func>` in addition to `Func`). 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) are 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`. 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: ```C# public static IAsyncEnumerable Select(this IAsyncEnumerable source, Func func); public static IAsyncEnumerable Where(this IAsyncEnumerable source, Func func); ``` then this C# code will "just work": ```C# IAsyncEnumerable enumerable = ...; IAsyncEnumerable 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: ```C# public static IAsyncEnumerable Select(this IAsyncEnumerable source, Func> func); ``` then this would "just work": ```C# IAsyncEnumerable result = from url in urls where item % 2 == 0 select SomeAsyncMethod(item); async ValueTask 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: ```C# IAsyncEnumerable 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` 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` can be published to an `IObserver` simply by `await foreach`'ing over the enumerator and `OnNext`'ing the data to the observer, so an `AsObservable` extension method is possible. Consuming an `IObservable` 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` to be pulled from with an `IAsyncEnumerator`. 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.