从头开始实现IObservable <T>

Question

Reactive Extensions带有许多辅助方法,用于将现有事件和异步操作转换为可观察对象,但是如何从头开始实现IObservable <T>？

IEnumerable有一个可爱的yield关键字,使其实现起来非常简单.

实现IObservable <T>的正确方法是什么？

我需要担心线程安全吗？

我知道有人支持在特定的同步上下文中回调,但这是否是我作为IObservable <T>作者需要担心或者以某种方式内置的内容？

更新:

这是Brian的F#解决方案的C#版本

using System;
using System.Linq;
using Microsoft.FSharp.Collections;

namespace Jesperll
{
    class Observable<T> : IObservable<T>, IDisposable where T : EventArgs
    {
        private FSharpMap<int, IObserver<T>> subscribers = 
                 FSharpMap<int, IObserver<T>>.Empty;
        private readonly object thisLock = new object();
        private int key;
        private bool isDisposed;

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

        protected virtual void Dispose(bool disposing)
        {
            if (disposing && !isDisposed)
            {
                OnCompleted();
                isDisposed = true;
            }
        }

        protected void OnNext(T value)
        {
            if (isDisposed)
            {
                throw new ObjectDisposedException("Observable<T>");
            }

            foreach (IObserver<T> observer in subscribers.Select(kv => kv.Value))
            {
                observer.OnNext(value);
            }
        }

        protected void OnError(Exception exception)
        {
            if (isDisposed)
            {
                throw new ObjectDisposedException("Observable<T>");
            }

            if (exception == null)
            {
                throw new ArgumentNullException("exception");
            }

            foreach (IObserver<T> observer in subscribers.Select(kv => kv.Value))
            {
                observer.OnError(exception);
            }
        }

        protected void OnCompleted()
        {
            if (isDisposed)
            {
                throw new ObjectDisposedException("Observable<T>");
            }

            foreach (IObserver<T> observer in subscribers.Select(kv => kv.Value))
            {
                observer.OnCompleted();
            }
        }

        public IDisposable Subscribe(IObserver<T> observer)
        {
            if (observer == null)
            {
                throw new ArgumentNullException("observer");
            }

            lock (thisLock)
            {
                int k = key++;
                subscribers = subscribers.Add(k, observer);
                return new AnonymousDisposable(() =>
                {
                    lock (thisLock)
                    {
                        subscribers = subscribers.Remove(k);
                    }
                });
            }
        }
    }

    class AnonymousDisposable : IDisposable
    {
        Action dispose;
        public AnonymousDisposable(Action dispose)
        {
            this.dispose = dispose;
        }

        public void Dispose()
        {
            dispose();
        }
    }
}

编辑:如果调用Dispose两次,不要抛出ObjectDisposedException

Answer 1

在官方文档不赞成用户实施的IObservable自己.相反,用户应该使用工厂方法Observable.Create

如果可能,通过组合现有运算符来实现新运算符.否则使用Observable.Create实现自定义运算符

碰巧Observable.Create是Reactive内部类的一个简单的包装器AnonymousObservable:

public static IObservable<TSource> Create<TSource>(Func<IObserver<TSource>, IDisposable> subscribe)
{
    if (subscribe == null)
    {
        throw new ArgumentNullException("subscribe");
    }
    return new AnonymousObservable<TSource>(subscribe);
}

我不知道他们为什么不公开他们的实施,但是,嘿,无论如何.

Answer 2

老实说,我不确定这一切是多么'正确',但如果根据我迄今为止的经验感觉相当不错.它是F#代码,但希望你能感受到它的味道.它允许您"新建"一个源对象,然后可以调用Next/Completed/Error on,它管理订阅并在源或客户端执行错误操作时尝试Assert.

type ObservableSource<'T>() =     // '
    let protect f =
        let mutable ok = false
        try 
            f()
            ok <- true
        finally
            Debug.Assert(ok, "IObserver methods must not throw!")
            // TODO crash?
    let mutable key = 0
    // Why a Map and not a Dictionary?  Someone's OnNext() may unsubscribe, so we need threadsafe 'snapshots' of subscribers to Seq.iter over
    let mutable subscriptions = Map.empty : Map<int,IObserver<'T>>  // '
    let next(x) = subscriptions |> Seq.iter (fun (KeyValue(_,v)) -> protect (fun () -> v.OnNext(x)))
    let completed() = subscriptions |> Seq.iter (fun (KeyValue(_,v)) -> protect (fun () -> v.OnCompleted()))
    let error(e) = subscriptions |> Seq.iter (fun (KeyValue(_,v)) -> protect (fun () -> v.OnError(e)))
    let thisLock = new obj()
    let obs = 
        { new IObservable<'T> with       // '
            member this.Subscribe(o) =
                let k =
                    lock thisLock (fun () ->
                        let k = key
                        key <- key + 1
                        subscriptions <- subscriptions.Add(k, o)
                        k)
                { new IDisposable with 
                    member this.Dispose() = 
                        lock thisLock (fun () -> 
                            subscriptions <- subscriptions.Remove(k)) } }
    let mutable finished = false
    // The methods below are not thread-safe; the source ought not call these methods concurrently
    member this.Next(x) =
        Debug.Assert(not finished, "IObserver is already finished")
        next x
    member this.Completed() =
        Debug.Assert(not finished, "IObserver is already finished")
        finished <- true
        completed()
    member this.Error(e) =
        Debug.Assert(not finished, "IObserver is already finished")
        finished <- true
        error e
    // The object returned here is threadsafe; you can subscribe and unsubscribe (Dispose) concurrently from multiple threads
    member this.Value = obs

我会对任何有关这里好坏的想法感兴趣; 我还没有机会看看devlabs的所有新Rx东西......

我自己的经历表明:

• 订阅observables的人不应该从订阅中抛出.当用户抛出时,观察者无法做任何合理的事情.(这与事件类似.)很可能异常只会冒泡到顶级catch-all处理程序或崩溃应用程序.
• 来源可能应该是"逻辑单线程".我认为编写可以对并发OnNext调用作出反应的客户端可能更难; 即使每个单独的调用来自不同的线程,避免并发调用也是有帮助的.
• 拥有一个强制执行某些"契约"的基础/助手类非常有用.

如果人们能够沿着这些方向展示更具体的建议,我会很好奇.

Answer 3

是的,yield关键字很可爱; 也许IObservable(OfT)会有类似的东西？[编辑:在Eric Meijer的PDC '09谈话中,他说"是的,看这个空间",以产生可观察的声明产量.

对于接近的东西(而不是自己动手),请查看" (尚未)101 Rx Samples "维基的底部,团队建议使用Subject(T)类作为"后端"来实现IObservable( OFT).这是他们的例子:

public class Order
{            
    private DateTime? _paidDate;

    private readonly Subject<Order> _paidSubj = new Subject<Order>();
    public IObservable<Order> Paid { get { return _paidSubj.AsObservable(); } }

    public void MarkPaid(DateTime paidDate)
    {
        _paidDate = paidDate;                
        _paidSubj.OnNext(this); // Raise PAID event
    }
}

private static void Main()
{
    var order = new Order();
    order.Paid.Subscribe(_ => Console.WriteLine("Paid")); // Subscribe

    order.MarkPaid(DateTime.Now);
}