Introduction

The next step in this series on creating my own IObservable<T> implementation is concurrency. (See below for links to previous parts, Part 5 has a full code listing.)

This is the hard part. Too easy to think it is OK when it isn’t, and too hard to validate that is really thread safe.

The only option is careful analysis, and then double (triple, …) check everything.

After all even the best set of unit tests will fail to find problems if it takes a particular combination of OS version, hardware and other processes running to show the problem. With race conditions, deadlocks and all the other concurrency fun such is not just a possibility but likely.

But first, the new 1.0.2350.0 (2010-03-15) build of Rx makes no change to my code, all tests pass without any changes. And this is now running on my main system (Rx has go-live licence, so why not) with its multiple codes (previously was working on a VM with a single virtual CPU).

What Needs Protection?

At the heart of my implementation there is some shared data. This needs protecting where multiple threads could operate concurrently. E.g. while events are pushed to subscribers using a copy of the list of subscribers (this allows re-entrant subscription and un-subscription) taking a copy of the values of a Dictionary<K,V> certainly is not safe.

Some operations are inherently safe (defined by the platforms object model, see Joe Duffy’s blog for more details). In this case reading or writing an aligned, volatile int or reference is safe (such reads and writes are inherently atomic). And only a single field needs to be checked, written to, or both compared & exchanged (via Interlocked.CompareExchange):

• Checking if the endOfEvents field (reference to a Notification<T>) is null.

• The boolean field pushingEvents used only in PushEvents (the method that loops through a snapshot of the current subscribers) to avoid re-entrant or multi-threaded pushing of events.

  When it is read, and true in a simple check there is no problem (just return from PushEvents. But if it is false it then needs to be set. So two operations are needed: a read and a write. Interlocked.CompareExchange will take care of this without a lock. (Before returning from pushing events it is set back to false, but a single write is also OK.)

  As there is no overload of Interlocked.CompareExchange for bool the field will need to be an int with values 0 and 1.

• To avoid races in Next, Completed and Error between checking for endOfEvents and setting it one can use Interlocked.CompareExchange.

There are also two pieces of functionality that require multiple operations on fields (or objects that are not thread safe). In each case a lock is needed.

• The collection of subscribers and subscriber id counter. With operations to add or remove a subscriber, and to snapshot a list of all subscribers. This will require a lock, there is no way (I can see) around this.

• To operate on the queue of pending events can be handled by changing to an instance of ConcurrentQueue<T> with modifications in PushEvents to use TryDequeue rather than a combination of checking the count and then Dequeue.

What Doesn’t Need Protection

This implementation is not about being fair. So if two threads both try and use an instance together, with one calling Next and the other Completed only the Windows thread scheduler gets any say in which wins (if Completed is called first, no subscriber will see the value passed to Next).

Ready to Implement?

Even as I wrote the analysis above (which was already the second pass through what would be needed) I noted that I could avoid a lock for Next, Completed, and Error and use CompareExchange. So I’ve probably still missed something: time for another pass over the code.

And another question is how to test…? Throwing a load of cocurrent tasks from the Parallel Extensions (see here) will be a start, but cannot really cover mixing in Completed and Error because I cannot determine which thread will win and which will not, and thus no way to verify the “right one” won in a unit test assert.

More to work on.

Previously

• Part 1: The basic implementation of Subscribe, unsubscribe and Next.

• Part 2: Making Subscribe and unsubscribe re-entrant.

• Part 3: Why re-entrancy is important.

• Part 4: Making Next re-entrant.

• Part 5: Implementing OnCompleted and OnError to complete support for full IObserver<T> interface.

• Part 5a: Updates for build 1.0.2317.0 of the Reactive Extensions (2010-03-05).

The new code drop came through, 2½ months after the last drop, on 5^th March. The release notes are suitably long. With a fair amount of change, much of it under the covers to edge cases.

But a fair number of things have moved assemblies and changed name.

Impact on My Implementation

Given all the change, there was relatively little impact. The only thing that broke was my use of Notification<T>: the Current property is now called Value. And it now has an Exception property, which saves a cast when handling an OnError notification.

My (unpublished) test harness used GroupDisposable to clean up subscriptions, this was renamed CompositeDisposable and moved to the CoreEx assembly.

Both of these renames give slightly better and more consistent names, and the addition of the Exception property is definitely better. So overall the new drop is, for me, an improvement so far: once the code compiled all tests just passed.

The only function that changed was SendNotification so following DRY helped. The new implementation:

private static void SendNotification(IObserver<T> obs, Notification<T> n) {
    switch (n.Kind) {
    case NotificationKind.OnNext:
        Debug.Assert(n.HasValue);
        obs.OnNext(n.Value);
        break;
    case NotificationKind.OnCompleted:
        obs.OnCompleted();
        break;
    case NotificationKind.OnError:
        obs.OnError(n.Exception);
        break;
    }
}

So Where’s The Concurrency Support…?

Coming! :-)
I've been rather head down in learning WPF for something else, and trying to avoid too many distractions while getting my head around the basics of WPF. Concurrency support is perhaps the biggest challenge here and I therefore want to be able to focus on it and get it right. This will mean taking a block of focused time.

Introduction

So far in this series (links below) the focus has been on implementing Next and Subscribe in a way that is re-entrant. But that leaves two methods uncovered. Both Completed and Error end the sequence, but in different ways. Error passes an exception indicating the source has failed, whereas Completed just indicates it has ended.

Based on Subject<T> when a sequence of events is ended:

• Existing subscribers received a call to their OnError or OnCompleted (as applicable).

• Further calls to any of Subject<T>'s IObserver<T> methods is a no-op.

• Any new subscribers immediately get a call to their OnError or OnCompleted (as applicable) and nothing more.

Previously

• Part 1: The basic implementation of Subscribe, unsubscribe and Next.

• Part 2: Making Subscribe and unsubscribe re-entrant.

• Part 3: Why re-entrancy is important.

• Part 4: Making Next re-entrant.

Notification<T> As A Way to Store An Observable Sequence

The reactive Framework (Rx) distributable comes with three assemblies:

• System.CoreEx

• System.Reactive

• System.Interactive

System.CoreEx contains types that would have been in System or System.Core had .NET been designed with Rx in it from the start. System.Interactive extends LINQ to Objects by back-porting many of IObservable<T>'s extension methods (or ’operators”) to IEnumerable<T>. System.Reactive implements LINQ like operators for IObservable<T>.

My observable implementation almost doesn’t need any of these three assemblies (IObservable<T> and IObserver<T> being defined in the mscorlib assembly).

However one rather useful group of types defined in the System.CoreEx assembly in namespace System.Collections.Generic are Notification<T> and its subtypes. These types allow the “value” of an observable event to be stored. Using one of the three subtypes (one each for OnNext, OnCompleted and OnError) defined as members of Notification<T> (hence created as, for example, new Notification<int>.OnNext(1)). Therefore a collection of events covering all three IObserver<T> methods can be easily created without some custom helper types.

These types are provided for the Materialize and Dematerialise enumerable and observable operators (the former being one the back-ported operators in System.Interactive).

Overview of The Implementation of Completed and OnError

Both follow the same pattern as the existing Next implementation:

1. If ended, return

2. Add the new event to a queue of pending events.

3. If not already pushing an event, process the queue.

The first of these steps is new, and handles the requirement that either Completed or OnError ends the sequence.

To make handling new subscriptions after the end of the sequence the endOfSequnec field is a Notification<T> reference. If null the sequence has not ended, otherwise this is the event to immediately push. Also a helper from System.Reactive is used as the return value: System.Disposables.Disposable.Empty which is an instance of a type which implements IDisposable as a no-op and these is no need to track this after the end subscriber (it will never be called again).

The New Implementation

The single biggest change to support storing notifications (rather than values of type T>) is in PushEvents where different kinds of notifications need to call different IObserver<T> methods, so a switch is needed. This is in its own method (SendNotification) because it can then be reused to send the ending event to post-end subscribers.

The implementations of Next, Completed and Error end up being very similar and simple.

Overall, compared to the code from part 4, the implementation is a somewhat longer, but certainly better factored with each method doing one thing. The most complex (PushEvents) does have to loop over both pending events and subscribers, but that is all it does.

The Code

using System;
using System.Collections.Generic;
using System.Linq;

public class Observable4<T> : IObservable<T> {
    private Dictionary<int, IObserver<T>> subscribers = new Dictionary<int, IObserver<T>>();
    private int nextSubscriber = 0;
    private Queue<Notification<T>> pendingEvents = new Queue<Notification<T>>();
    private bool pushingEvent = false;  // If true, only queue the event.
    private Notification<T> endOfEvents;   // Not null => have had ended, and this is the Completed or Error that ended it.

    public IDisposable Subscribe(IObserver<T> observer) {
        if (endOfEvents != null) {
            // Already have end of sequence, so just tell this new subscriber
            // and return no-op.
            SendNotification(observer, endOfEvents);
            return System.Disposables.Disposable.Empty;
        } else {
            subscribers.Add(nextSubscriber, observer);
            return new Observable4Disposer(this, nextSubscriber++);
        }
    }

    private void Unsubscribe(int index) {
        subscribers.Remove(index);
    }

    public void Next(T value) {
        if (endOfEvents != null) { return; }
        pendingEvents.Enqueue(new Notification<T>.OnNext(value));
        PushEvents();
    }

    public void Completed() {
        if (endOfEvents != null) { return; }
        endOfEvents = new Notification<T>.OnCompleted();
        pendingEvents.Enqueue(endOfEvents);
        PushEvents();
    }

    public void Error(Exception exn) {
        if (endOfEvents != null) { return; }
        endOfEvents = new Notification<T>.OnError(exn);
        pendingEvents.Enqueue(endOfEvents);
        PushEvents();
    }

    private void PushEvents() {
        if (!pushingEvent) {
            try {
                pushingEvent = true;
                while (pendingEvents.Count > 0) {
                    // To allow subscribers to unsubscribe while calling them, don't
                    // directly iterate over the collection, but use a copy.
                    // Take this each time, in case some event triggers a change
                    var subs = subscribers.Values.ToArray();
                    var n = pendingEvents.Dequeue();
                    foreach (var s in subs) {
                        SendNotification(s, n);
                    }
                }
            } finally {
                pushingEvent = false;
            }
        }
    }

    private static void SendNotification(IObserver<T> obs, Notification<T> n) {
        switch (n.Kind) {
        case NotificationKind.OnNext:
            obs.OnNext(n.Current);
            break;
        case NotificationKind.OnCompleted:
            obs.OnCompleted();
            break;
        case NotificationKind.OnError:
            obs.OnError(((Notification<T>.OnError)(n)).Exception);
            break;
        }
    }


    private class Observable4Disposer : IDisposable {
        private Observable4<T> target;
        private int index;
        internal Observable4Disposer(Observable4<T> target, int index) {
            this.target = target;
            this.index = index;
        }

        public void Dispose() {
            target.Unsubscribe(index);
        }
    }
}

Further Steps

The toughest part to get right is still left: concurrency. This will be difficult as much to test as to implement, tests using workers are never simple and just being confident that they are doing things concurrently is hard (after all a simple Interlocked.Increment might be slow enough to effectively eliminate any concurrency due to forcing CPU cache synchronisation even with multiple idle cores).

Introduction

In building my own implementation of IObservable<T> across a series of posts (Part 1, Part 2, Part 3 and Part 4) I’ve been making heavy use of Observer.Create<T> to create helper observer instances to ensure that the right methods of the subscribed observers are called, the right number of times (and even—although somewhat harder to track in all but the simplest cases—the right order).

The Discovery

In working towards Part 5 which will cover implementing the two, so far missing, methods of IObserver<T>: OnCompleted and OnError I was rather surprise when this test passed:

[TestMethod]
public void Observable4_CallingCompletedAgainIsNoOp() {
    bool hasCompleted = false;
    var sub = Obs.Create<int>(
        i => { Assert.Fail("OnNext should not have been called"); },
        exn => { Assert.Fail("OnError should not have been called"); },
        () => {
            Assert.IsFalse(hasCompleted, "OnCompleted for single subscriber called multiple times");
            hasCompleted = true;
        });

    var source = new Observable4<int>();
    using (var unsub = source.Subscribe(sub)) {
        source.Completed();
        source.Completed();
    }
    Assert.IsTrue(hasCompleted);
}

When I knew that Observable4<int>.Completed had no logic to prevent IObserver<T>.OnComplete being called multiple times. Under the debugger it was quite clear that the third lambda was not being called (briefly I considered a debugger bug, but decided to check more fully first).

I looked at the implementation of Observer.Create<T> in Reflector. That factory method creates an instance of an internal type, which derives from another internal type AbstractObserver<T>. So far largely as expected.

But in looking at the implementation of the IObserver<T> methods I saw:

public void OnCompleted() {
    if (!this.IsStopped) {
        this.IsStopped = true;
        this.Completed();
    }
}

I.e. it keeps track of a call to OnCompleted and blocks further events. All three IObserver<T> methods check for IsStopped, and it is also set in OnError.

The Implication

If I want an observer that I can use to test an observable for compliance to the IObserver<T> semantic contract, I cannot use Observer.Create<T>. It will hide breaking the “OnCompleted or OnError end the observable’s event sequence”.

The Solution

In the end the fix is rather easy: make my own Observer.Create<T> which doesn’t contain any such logic. This is rather simple:

public static class Obs {
    public static AnonObserver<T> Create<T>(Action<T> onNext) {
        return new AnonObserver<T>(onNext);
    }
    public static AnonObserver<T> Create<T>(Action<T> onNext, Action<Exception> onError, Action onComplete) {
        return new AnonObserver<T>(onNext, onError, onComplete);
    }
}

public class AnonObserver<T> : IObserver<T> {
    private Action<T> onNext;
    private Action onComplete;
    private Action<Exception> onError;

    public AnonObserver(Action<T> onNext) {
        if (onNext == null) { throw new ArgumentNullException("onNext"); }
        this.onNext = onNext;
    }
    public AnonObserver(Action<T> onNext, Action<Exception> onError, Action onComplete) {
        if (onNext == null) { throw new ArgumentNullException("onNext"); }
        if (onError == null) { throw new ArgumentNullException("onError"); }
        if (onComplete == null) { throw new ArgumentNullException("onComplete"); }
        this.onNext = onNext;
        this.onError = onError;
        this.onComplete = onComplete;
    }

    public void OnCompleted() {
        if (onComplete != null) {
            onComplete();
        }
    }

    public void OnError(Exception error) {
        if (onError != null) {
            onError(error);
        }
    }

    public void OnNext(T value) {
        onNext(value);
    }
}

I now need to just replace all the use of Observer.Create<T> with Obs.Create<T> (and try and think of a better name).

Introduction

In Part 3 I looked at the effect of calling Subject<T>.OnNext from within a subscriber to that Subject<T>. I really do not think that the out of order events that creates are going to create anything other than confusion.

In my implementation, continuing on from Part 1 and Part 2, I want to avoid the implicit re-ordering that happens when Next(value) is called from within a subscriber.

The Re-Entrant Implementation

In the basic implementation a snapshot of the subscribers is taken (to avoid problems with the collection being modified by re-entrant subscribe and unsubscribe operations), then iterated over:

var subs = subscribers.Values.ToArray();
foreach (var sub in subs) {
   sub.OnNext(value);
}

In the case of a re-entrant Next call this loop will simply exist on the stack twice:

Observable2.Next()
ATestObserver.OnNext()
Observable2.Next()

With the second call iterating through the list of subscribers before the first moves on to the next.

The solution here is to recognise that an iteration is taking place already when Next is called and, rather than just looping to call all subscribers, buffer up the new value. The original call keeps looping until all buffered values are sent out. (Once, later in this series, concurrency is introduced locking is going to be needed here.)

First two new fields are needed. A Queue to act as buffer, and a flag to indicate events are currently being pushed to the subscribers:

private Queue<T> pendingEvents = new Queue<T>();
private bool pushingEvent = false;

Then reworking the Next implementation:

pendingEvents.Enqueue(value);
if (!pushingEvent) {
    try {
        pushingEvent = true;
        while (pendingEvents.Count > 0) {
            var subs = subscribers.Values.ToArray();
            var v = pendingEvents.Dequeue();
            foreach (var s in subs) {
                s.OnNext(v);
            }
        }
    } finally {
        pushingEvent = false;
    }
}

A few points worth noting:

• By always putting the new event value in the queue there is no special cases needed for the first event

• The try…finally ensures the pushingEvent gets reset, but this could leave un-pushed events. This is a consequence of the RX exception model (propogate back to the called), but with the added twist that a re-entrant call to this will not get the exception, the original caller of Next will.

• I have chosen to re-snapshot the subscribers for each event. This will avoid a subscriber that unsubscribes in response to one event getting more, even if already queued up.

Next Steps

This is still missing Exception and Complete methods (to call IObserver<T>.Error() and .OnCompleted.

…and concurrency support of course.

Introduction

Continuing from Part 1 and Part 2 where the first steps to a reusable IObservable<T> were taken.

But this is something of an aside. Because in starting to look at the reasons to explicitly support this, I found that the RX type Subject<T> (in assembly "System.Reactive", namespace System.Collections.Generic) has exactly the same behaviour as Observable2<T> from Part 2.

Personally I feel that this behaviour will lead to inconsistent and unexpected ordering of events as seen by subscribers, depending on the order in which the observers subscribed and the details of the internals of the observable (what kind of collection is used to store the list of subscribers). This opens up the potential of races to subscribe with different timing leading to seeing out of order events.

One final note here, where I started drafting my implementation of IObservable<T> I was not aware of Subject<T>. When I became aware of this type, I almost decided to give up on my own implementation. With this discovery I am not rather glad I didn't.

The Test Code

var obs = new Subject<int>();
var subCalls = 0;
var sub1Calls = 0;

var sub0 = Observer.Create<int>(i => {
    ++subCalls;
    Console.WriteLine("Sub 0 ({2}): sub1 called {0} times (total subscriber calls {1})", sub1Calls, subCalls, i);
});
var sub1 = Observer.Create<int>(i => {
    ++subCalls;
    ++sub1Calls;
    Console.WriteLine("Sub 1 ({2}): sub1 called {0} times (total subscriber calls: {1})", sub1Calls, subCalls, i);
    if (i > 0 && i < 3) {
	obs.OnNext(-1 * i);
    }
});
var sub2 = Observer.Create<int>(i => {
    ++subCalls;
    Console.WriteLine("Sub 2 ({2}): sub1 called {0} times (total subscriber calls: {1})", sub1Calls, subCalls, i);
});

using (var d0 = obs.Subscribe(sub0))
using (var d1 = obs.Subscribe(sub1))
using (var d2 = obs.Subscribe(sub2)) {
    obs.OnNext(1);
}

Console.WriteLine("END: sub1 called {0} times, all subscribers {1} times", sub1Calls, subCalls);

In this the three subscribers (imaginatively named sub0, sub1 and sub2) all increment a count of subscriber calls, and all print the same details to the console: which subscriber, the event value, the total number of subscriber calls and the number of calls to the second subscriber (sub1

Additionally the second subscriber will raise an event itself, recursively publishing an event if the current value is 1≤i≤2 with value -1×i. Thus there is no more than a single level of recursion.

Sub 0 (1): sub1 called 0 times (total subscriber calls 1)
Sub 1 (1): sub1 called 1 times (total subscriber calls: 2)
Sub 0 (-1): sub1 called 1 times (total subscriber calls 3)
Sub 1 (-1): sub1 called 2 times (total subscriber calls: 4)
Sub 2 (-1): sub1 called 2 times (total subscriber calls: 5)
Sub 2 (1): sub1 called 2 times (total subscriber calls: 6)
END: sub1 called 2 times, all subscribers 6 times