Develop Reference

How do I signal completion of my dataflow?

I've got a class the implements a dataflow composed of 3 steps using TPL Dataflow.
In the constructor I create the steps as TransformBlocks and link them up using LinkTo with DataflowLinkOptions.PropagateCompletion set to true. The class exposes a single method which kicks of the workflow by calling SendAsync on the 1st step. The method returns the "Completion" property of the final step of the workflow.
At the moment the steps in the workflow appear to execute as expected but final step never completes unless I explicitly call Complete on it. But doing that short-circuits the workflow and none of the steps are executed? What am I doing wrong?
public class MessagePipeline {
private TransformBlock<object, object> step1;
private TransformBlock<object, object> step2;
private TransformBlock<object, object> step3;
public MessagePipeline() {
var linkOptions = new DataflowLinkOptions { PropagateCompletion = true };
step1 = new TransformBlock<object, object>(
x => {
Console.WriteLine("Step1...");
return x;
});
step2 = new TransformBlock<object, object>(
x => {
Console.WriteLine("Step2...");
return x;
});
step3 = new TransformBlock<object, object>(
x => {
Console.WriteLine("Step3...");
return x;
});
step1.LinkTo(step2, linkOptions);
step2.LinkTo(step3, linkOptions);
}
public Task Push(object message) {
step1.SendAsync(message);
step1.Complete();
return step3.Completion;
}
}
...
public class Program {
public static void Main(string[] args) {
var pipeline = new MessagePipeline();
var result = pipeline.Push("Hello, world!");
result.ContinueWith(_ => Console.WriteLine("Completed"));
Console.ReadLine();
}
}

When you link the steps, you need to pass a DataflowLinkOptions with the the PropagateCompletion property set to true to propagate both completion and errors. Once you do that, calling Complete() on the first block will propagete completion to downstream blocks.
Once a block receives the completion event, it finishes processing then notifies its linked downstream targets.
This way you can post all your data to the first step and call Complete(). The final block will only complete when all upstream blocks have completed.
For example,
var linkOptions=new DataflowLinkOptions { PropagateCompletion = true};
myFirstBlock.LinkTo(mySecondBlock,linkOptions);
mySecondBlock.LinkTo(myFinalBlock,linkOptions);
foreach(var message in messages)
{
myFirstBlock.Post(message);
}
myFirstBlock.Complete();
......
await myFinalBlock.Completion;
PropagateCompletion isn't true by default because in more complex scenarios (eg non-linear flows, or dynamically changing flows) you don't want completion and errors to propagate automatically. You may also want to avoid automatic completion if you want to handle errors without terminating the entire flow.
Way back when TPL Dataflow was in beta the default was true but this was changed on RTM
UPDATE
The code never completes because the final step is a TransformBlock with no linked target to receive its output. This means that even though the block received the completion signal, it hasn't finished all its work and can't change its own Completion status.
Changing it to an ActionBlock<object> removes the issue.

You need to explicitly call Complete.

Unit tests failing with Observable.FromAsync and Observable.Switch

I'm having troubles testing a class that makes use of Observable.FromAsync<T>() and Observable.Switch<T>(). What it does is to wait for a trigger observable to produce a value, then it starts an async operation, and finally recollects all operations' results in a single output sequence. The gist of it is something like:
var outputStream = triggerStream
.Select(_ => Observable
.FromAsync(token => taskProducer.DoSomethingAsync(token)))
.Switch();
I put up some sanity check tests with the bare minimum parts to understand what's going on, here's the test with results in comments:
class test_with_rx : nspec
{
void Given_async_task_and_switch()
{
Subject<Unit> triggerStream = null;
TaskCompletionSource<long> taskDriver = null;
ITestableObserver<long> testObserver = null;
IDisposable subscription = null;
before = () =>
{
TestScheduler scheduler = new TestScheduler();
testObserver = scheduler.CreateObserver<long>();
triggerStream = new Subject<Unit>();
taskDriver = new TaskCompletionSource<long>();
// build stream under test
IObservable<long> streamUnderTest = triggerStream
.Select(_ => Observable
.FromAsync(token => taskDriver.Task))
.Switch();
/* Also tried with this Switch() overload
IObservable<long> streamUnderTest = triggerStream
.Select(_ => taskDriver.Task)
.Switch(); */
subscription = streamUnderTest.Subscribe(testObserver);
};
context["Before trigger"] = () =>
{
it["Should not notify"] = () => testObserver.Messages.Count.Should().Be(0);
// PASSED
};
context["After trigger"] = () =>
{
before = () => triggerStream.OnNext(Unit.Default);
context["When task completes"] = () =>
{
long result = -1;
before = () =>
{
taskDriver.SetResult(result);
//taskDriver.Task.Wait(); // tried with this too
};
it["Should notify once"] = () => testObserver.Messages.Count.Should().Be(1);
// FAILED: expected 1, actual 0
it["Should notify task result"] = () => testObserver.Messages[0].Value.Value.Should().Be(result);
// FAILED: of course, index out of bound
};
};
after = () =>
{
taskDriver.TrySetCanceled();
taskDriver.Task.Dispose();
subscription.Dispose();
};
}
}
In other tests I've done with mocks too, I can see that the Func passed to FromAsync is actually invoked (e.g. taskProducer.DoSomethingAsync(token)), but then it looks like nothing more follows, and the output stream doesn't produce the value.
I also tried inserting some Task.Delay(x).Wait(), or some taskDriver.Task.Wait() before hitting expectations, but with no luck.
I read this SO thread and I'm aware of schedulers, but at a first look I thought I didn't need them, no ObserveOn() is being used. Was I wrong? What am I missing? TA
Just for completeness, testing framework is NSpec, assertion library is FluentAssertions.

What you're hitting is a case of testing Rx and TPL together.
An exhaustive explanation can be found here but I'll try to give advice for your particular code.
Basically your code is working fine, but your test is not.
Observable.FromAsync will transform into a ContinueWith on the provided task, which will be executed on the taskpool, hence asynchronously.
Many ways to fix your test: (from ugly to complex)
Sleep after result set (note wait doesn't work because Wait doesn't wait for continuations)
taskDriver.SetResult(result);
Thread.Sleep(50);
Set the result before executing FromAsync (because FromAsync will return an immediate IObservable if the task is finished, aka will skip ContinueWith)
taskDriver.SetResult(result);
triggerStream.OnNext(Unit.Default);
Replace FromAsync by a testable alternative, e.g
public static IObservable<T> ToObservable<T>(Task<T> task, TaskScheduler scheduler)
{
if (task.IsCompleted)
{
return task.ToObservable();
}
else
{
AsyncSubject<T> asyncSubject = new AsyncSubject<T>();
task.ContinueWith(t => task.ToObservable().Subscribe(asyncSubject), scheduler);
return asyncSubject.AsObservable<T>();
}
}
(using either a synchronous TaskScheduler, or a testable one)

Rx and tasks - cancel running task when new task is spawned?

I have an user interaction scenario I'd like to handle with Rx.
The scenario is similar to the canonical "when user stops typing, do some work" (usually, search for what the user has typed so far) (1) - but I also need to :
(2) only get the latest of the results of "do some work" units (see below)
(3) when a new unit of work starts, cancel any work in progress (in my case it's CPU intensive)
For (1) I use an IObservable for the user events, throttled with .Throttle() to only trigger on pauses between events ("user stops typing").
From that, i .Select(_ => CreateMyTask(...).ToObservable()).
This gives me an IObservable<IObservable<T>> where each of the inner observables wraps a single task.
To get (2) I finally apply .Switch() to only get the results from the newest unit of work.
What about (3) - cancel pending tasks ?
If I understand correctly, whenever there's a new inner IObservable<T>, the .Switch() method subscribes to it and unsubscribes from the previous one(s), causing them to Dispose().
Maybe that can be somehow wired to trigger the task to cancel?

You can just use Observable.FromAsync which will generate tokens that are cancelled when the observer unsubcribes:
input.Throttle(...)
.Select(_ => Observable.FromAsync(token => CreateMyTask(..., token)))
.Switch()
.Subscribe(...);
This will generate a new token for each unit of work and cancel it every time Switch switches to the new one.

Do you have to work with Tasks?
If you're happy to work purely with Observables then you can do this nicely yourself.
Try doing something like this:
var query =
Observable.Create<int>(o =>
{
var cancelling = false;
var cancel = Disposable.Create(() =>
{
cancelling = true;
});
var subscription = Observable.Start(() =>
{
for (var i = 0; i < 100; i++)
{
Thread.Sleep(10); //1000 ms in total
if (cancelling)
{
Console.WriteLine("Cancelled on {0}", i);
return -1;
}
}
Console.WriteLine("Done");
return 42;
}).Subscribe(o);
return new CompositeDisposable(cancel, subscription);
});
This observable is doing some hard work in the for loop with the Thread.Sleep(10);, but when the observable is disposed the loop is exited and the intensive CPU work ceases. Then you can use the standard Rx Dispose with the Switch to cancel the in progress work.
If you'd like that bundled up in a method, then try this:
public static IObservable<T> Start<T>(Func<Func<bool>, T> work)
{
return Observable.Create<T>(o =>
{
var cancelling = false;
var cancel = Disposable
.Create(() => cancelling = true);
var subscription = Observable
.Start(() => work(() => cancelling))
.Subscribe(o);
return new CompositeDisposable(cancel, subscription);
});
}
And then call it with a function like this:
Func<Func<bool>, int> work = cancelling =>
{
for (var i = 0; i < 100; i++)
{
Thread.Sleep(10); //1000 ms in total
if (cancelling())
{
Console.WriteLine("Cancelled on {0}", i);
return -1;
}
}
Console.WriteLine("Done");
return 42;
};
Here's my code that proved this worked:
var disposable =
ObservableEx
.Start(work)
.Subscribe(x => Console.WriteLine(x));
Thread.Sleep(500);
disposable.Dispose();
I got "Cancelled on 50" (sometime "Cancelled on 51") as my output.

TPL Dataflow, can I query whether a data block is marked complete but has not yet completed?

Given the following:
BufferBlock<int> sourceBlock = new BufferBlock<int>();
TransformBlock<int, int> targetBlock = new TransformBlock<int, int>(element =>
{
return element * 2;
});
sourceBlock.LinkTo(targetBlock, new DataflowLinkOptions { PropagateCompletion = true });
//feed some elements into the buffer block
for(int i = 1; i <= 1000000; i++)
{
sourceBlock.SendAsync(i);
}
sourceBlock.Complete();
targetBlock.Completion.ContinueWith(_ =>
{
//notify completion of the target block
});
The targetBlock never seems to complete and I think the reason is that all the items in the TransformBlock targetBlock are waiting in the output queue as I have not linked the targetBlock to any other Dataflow block. However, what I actually want to achieve is a notification when (A) the targetBlock is notified of completion AND (B) the input queue is empty. I do not want to care whether items still sit in the output queue of the TransformBlock. How can I go about that? Is the only way to get what I want to query the completion status of the sourceBlock AND to make sure the InputCount of the targetBlock is zero? I am not sure this is very stable (is the sourceBlock truly only marked completed if the last item in the sourceBlock has been passed to the targetBlock?). Is there a more elegant and more efficient way to get to the same goal?
Edit: I just noticed even the "dirty" way to check on completion of the sourceBlock AND InputCount of the targetBlock being zero is not trivial to implement. Where would that block sit? It cannot be within the targetBlock because once above two conditions are met obviously no message is processed within targetBlock anymore. Also checking on the completion status of the sourceBlock introduces a lot of inefficiency.

I believe you can't directly do this. It's possible you could get this information from some private fields using reflection, but I wouldn't recommend doing that.
But you can do this by creating custom blocks. In the case of Complete() it's simple: just create a block that forwards each method to the original block. Except Complete(), where it will also log it.
In the case of figuring out when processing of all items is complete, you could link your block to an intermediate BufferBlock. This way, the output queue will be emptied quickly and so checking Completed of the internal block would give you fairly accurate measurement of when the processing is complete. This would affect your measurements, but hopefully not significantly.
Another option would be to add some logging at the end of the block's delegate. This way, you could see when processing of the last item was finished.

It would be nice if the TransformBlock had a ProcessingCompleted event that would fire when the block has completed the processing of all messages in its queue, but there is no such event. Below is an attempt to rectify this omission. The CreateTransformBlockEx method accepts an Action<Exception> handler, that is invoked when this "event" occurs.
The intention was to always invoke the handler before the final completion of the block. Unfortunately in the case that the supplied CancellationToken is canceled, the completion (cancellation) happens first, and the handler is invoked some milliseconds later. To fix this inconsistency would require some tricky workarounds, and may had other unwanted side-effects, so I am leaving it as is.
public static IPropagatorBlock<TInput, TOutput>
CreateTransformBlockEx<TInput, TOutput>(Func<TInput, Task<TOutput>> transform,
Action<Exception> onProcessingCompleted,
ExecutionDataflowBlockOptions dataflowBlockOptions = null)
{
if (onProcessingCompleted == null)
throw new ArgumentNullException(nameof(onProcessingCompleted));
dataflowBlockOptions = dataflowBlockOptions ?? new ExecutionDataflowBlockOptions();
var transformBlock = new TransformBlock<TInput, TOutput>(transform,
dataflowBlockOptions);
var bufferBlock = new BufferBlock<TOutput>(dataflowBlockOptions);
transformBlock.LinkTo(bufferBlock);
PropagateCompletion(transformBlock, bufferBlock, onProcessingCompleted);
return DataflowBlock.Encapsulate(transformBlock, bufferBlock);
async void PropagateCompletion(IDataflowBlock block1, IDataflowBlock block2,
Action<Exception> completionHandler)
{
try
{
await block1.Completion.ConfigureAwait(false);
}
catch { }
var exception =
block1.Completion.IsFaulted ? block1.Completion.Exception : null;
try
{
// Invoke the handler before completing the second block
completionHandler(exception);
}
finally
{
if (exception != null) block2.Fault(exception); else block2.Complete();
}
}
}
// Overload with synchronous lambda
public static IPropagatorBlock<TInput, TOutput>
CreateTransformBlockEx<TInput, TOutput>(Func<TInput, TOutput> transform,
Action<Exception> onProcessingCompleted,
ExecutionDataflowBlockOptions dataflowBlockOptions = null)
{
return CreateTransformBlockEx<TInput, TOutput>(
x => Task.FromResult(transform(x)), onProcessingCompleted,
dataflowBlockOptions);
}
The code of the local function PropagateCompletion mimics the source code of the LinkTo built-in method, when invoked with the PropagateCompletion = true option.
Usage example:
var httpClient = new HttpClient();
var downloader = CreateTransformBlockEx<string, string>(async url =>
{
return await httpClient.GetStringAsync(url);
}, onProcessingCompleted: ex =>
{
Console.WriteLine($"Download completed {(ex == null ? "OK" : "Error")}");
}, new ExecutionDataflowBlockOptions()
{
MaxDegreeOfParallelism = 10
});

First thing it is not right to use a IPropagator Block as a leaf terminal. But still your requirement can be fulfilled by asynchronously checking the output buffer of the TargetBlock for output messages and then consuming then so that the buffer could be emptied.
` BufferBlock<int> sourceBlock = new BufferBlock<int>();
TransformBlock<int, int> targetBlock = new TransformBlock<int, int>
(element =>
{
return element * 2;
});
sourceBlock.LinkTo(targetBlock, new DataflowLinkOptions {
PropagateCompletion = true });
//feed some elements into the buffer block
for (int i = 1; i <= 100; i++)
{
sourceBlock.SendAsync(i);
}
sourceBlock.Complete();
bool isOutputAvailable = await targetBlock.OutputAvailableAsync();
while(isOutputAvailable)
{
int value = await targetBlock.ReceiveAsync();
isOutputAvailable = await targetBlock.OutputAvailableAsync();
}
await targetBlock.Completion.ContinueWith(_ =>
{
Console.WriteLine("Target Block Completed");//notify completion of the target block
});
`

TPL Dataflow, guarantee completion only when ALL source data blocks completed

How can I re-write the code that the code completes when BOTH transformblocks completed? I thought completion means that it is marked complete AND the " out queue" is empty?
public Test()
{
broadCastBlock = new BroadcastBlock<int>(i =>
{
return i;
});
transformBlock1 = new TransformBlock<int, string>(i =>
{
Console.WriteLine("1 input count: " + transformBlock1.InputCount);
Thread.Sleep(50);
return ("1_" + i);
});
transformBlock2 = new TransformBlock<int, string>(i =>
{
Console.WriteLine("2 input count: " + transformBlock1.InputCount);
Thread.Sleep(20);
return ("2_" + i);
});
processorBlock = new ActionBlock<string>(i =>
{
Console.WriteLine(i);
});
//Linking
broadCastBlock.LinkTo(transformBlock1, new DataflowLinkOptions { PropagateCompletion = true });
broadCastBlock.LinkTo(transformBlock2, new DataflowLinkOptions { PropagateCompletion = true });
transformBlock1.LinkTo(processorBlock, new DataflowLinkOptions { PropagateCompletion = true });
transformBlock2.LinkTo(processorBlock, new DataflowLinkOptions { PropagateCompletion = true });
}
public void Start()
{
const int numElements = 100;
for (int i = 1; i <= numElements; i++)
{
broadCastBlock.SendAsync(i);
}
//mark completion
broadCastBlock.Complete();
processorBlock.Completion.Wait();
Console.WriteLine("Finished");
Console.ReadLine();
}
}
I edited the code, adding an input buffer count for each transform block. Clearly all 100 items are streamed to each of the transform blocks. But as soon as one of the transformblocks finishes the processorblock does not accept any more items and instead the input buffer of the incomplete transformblock just flushes the input buffer.

The issue is exactly what casperOne said in his answer. Once the first transform block completes, the processor block goes into “finishing mode”: it will process remaining items in its input queue, but it won't accept any new items.
There is a simpler fix than splitting your processor block in two though: don't set PropagateCompletion, but instead set completion of the processor block manually when both transform blocks complete:
Task.WhenAll(transformBlock1.Completion, transformBlock2.Completion)
.ContinueWith(_ => processorBlock.Complete());

The issue here is that you are setting the PropagateCompletion property each time you call the LinkTo method to link the blocks and the different in wait times in your transformation blocks.
From the documentation for the Complete method on the IDataflowBlock interface (emphasis mine):
Signals to the IDataflowBlock that it should not accept nor produce any more messages nor consume any more postponed messages.
Because you stagger out your wait times in each of the TransformBlock<TInput, TOutput> instances, transformBlock2 (waiting for 20 ms) is finished before transformBlock1 (waiting for 50 ms). transformBlock2 completes first, and then sends the signal to processorBlock which then says "I'm not accepting anything else" (and transformBlock1 hasn't produced all of its messages yet).
Note that the processing of transformBlock1 before transformBlock1 is not absolutely guaranteed; it's feasible that the thread pool (assuming you're using the default scheduler) will process the tasks in a different order (but more than likely will not, as it will steal work from the queues once the 20 ms items are done).
Your pipeline looks like this:
broadcastBlock
/ \
transformBlock1 transformBlock2
\ /
processorBlock
In order to get around this, you want to have a pipeline that looks like this:
broadcastBlock
/ \
transformBlock1 transformBlock2
| |
processorBlock1 processorBlock2
Which is accomplished by just creating two separate ActionBlock<TInput> instances, like so:
// The action, can be a method, makes it easier to share.
Action<string> a = i => Console.WriteLine(i);
// Create the processor blocks.
processorBlock1 = new ActionBlock<string>(a);
processorBlock2 = new ActionBlock<string>(a);
// Linking
broadCastBlock.LinkTo(transformBlock1,
new DataflowLinkOptions { PropagateCompletion = true });
broadCastBlock.LinkTo(transformBlock2,
new DataflowLinkOptions { PropagateCompletion = true });
transformBlock1.LinkTo(processorBlock1,
new DataflowLinkOptions { PropagateCompletion = true });
transformBlock2.LinkTo(processorBlock2,
new DataflowLinkOptions { PropagateCompletion = true });
You then need to wait on both processor blocks instead of just one:
Task.WhenAll(processorBlock1.Completion, processorBlock2.Completion).Wait();
A very important note here; when creating an ActionBlock<TInput>, the default is to have the MaxDegreeOfParallelism property on the ExecutionDataflowBlockOptions instance passed to it set to one.
This means that the calls to the Action<T> delegate that you pass to the ActionBlock<TInput> are thread-safe, only one will execute at a time.
Because you now have two ActionBlock<TInput> instances pointing to the same Action<T> delegate, you aren't guaranteed thread-safety.
If your method is thread-safe, then you don't have to do anything (which would allow you to set the MaxDegreeOfParallelism property to DataflowBlockOptions.Unbounded, since there's no reason to block).
If it's not thread-safe, and you need to guarantee it, you need to resort to traditional synchronization primitives, like the lock statement.
In this case, you'd do it like so (although it's clearly not needed, as the WriteLine method on the Console class is thread-safe):
// The lock.
var l = new object();
// The action, can be a method, makes it easier to share.
Action<string> a = i => {
// Ensure one call at a time.
lock (l) Console.WriteLine(i);
};
// And so on...

An addition to svick's answer: to be consistent with the behaviour you get with the PropagateCompletion option, you also need to forward exceptions in case a preceding block faulted. An extension method like the following takes care of that as well:
public static void CompleteWhenAll(this IDataflowBlock target, params IDataflowBlock[] sources) {
if (target == null) return;
if (sources.Length == 0) { target.Complete(); return; }
Task.Factory.ContinueWhenAll(
sources.Select(b => b.Completion).ToArray(),
tasks => {
var exceptions = (from t in tasks where t.IsFaulted select t.Exception).ToList();
if (exceptions.Count != 0) {
target.Fault(new AggregateException(exceptions));
} else {
target.Complete();
}
}
);
}

Here is a method that is functionally equivalent to pkt's CompleteWhenAll method, but with slightly less code:
public static void PropagateCompletion(IDataflowBlock[] sources,
IDataflowBlock target)
{
// Arguments validation omitted
Task allSourcesCompletion = Task.WhenAll(sources.Select(s => s.Completion));
ThreadPool.QueueUserWorkItem(async _ =>
{
try { await allSourcesCompletion.ConfigureAwait(false); } catch { }
Exception exception = allSourcesCompletion.IsFaulted ?
allSourcesCompletion.Exception : null;
if (exception is null) target.Complete(); else target.Fault(exception);
});
}
Usage example:
PropagateCompletion(new[] { transformBlock1, transformBlock2 }, processorBlock);
The PropagateCompletion method is a variant of a more general method with the same name, that I have posted here.

Other answers are quite clear about why PropagateCompletion=true mess things up when a block has more than two sources.
To provide a simple solution to the problem, you may want to look at an open source library DataflowEx that solves this kind of problem with smarter completion rules built-in. (It uses TPL Dataflow linking internally but supports complex completion propagation. The implementation looks similiar to WhenAll but also handles the dynamic link adding. Please check Dataflow.RegisterDependency() and TaskEx.AwaitableWhenAll() for impl detail.)
I slightly changed your code to make everything work using DataflowEx:
public CompletionDemo1()
{
broadCaster = new BroadcastBlock<int>(
i =>
{
return i;
}).ToDataflow();
transformBlock1 = new TransformBlock<int, string>(
i =>
{
Console.WriteLine("1 input count: " + transformBlock1.InputCount);
Thread.Sleep(50);
return ("1_" + i);
});
transformBlock2 = new TransformBlock<int, string>(
i =>
{
Console.WriteLine("2 input count: " + transformBlock2.InputCount);
Thread.Sleep(20);
return ("2_" + i);
});
processor = new ActionBlock<string>(
i =>
{
Console.WriteLine(i);
}).ToDataflow();
/** rather than TPL linking
broadCastBlock.LinkTo(transformBlock1, new DataflowLinkOptions { PropagateCompletion = true });
broadCastBlock.LinkTo(transformBlock2, new DataflowLinkOptions { PropagateCompletion = true });
transformBlock1.LinkTo(processorBlock, new DataflowLinkOptions { PropagateCompletion = true });
transformBlock2.LinkTo(processorBlock, new DataflowLinkOptions { PropagateCompletion = true });
**/
//Use DataflowEx linking
var transform1 = transformBlock1.ToDataflow();
var transform2 = transformBlock2.ToDataflow();
broadCaster.LinkTo(transform1);
broadCaster.LinkTo(transform2);
transform1.LinkTo(processor);
transform2.LinkTo(processor);
}
Full code is here.
Disclaimer: I am the author of DataflowEx, which is published under MIT license.