Task Parallel Library - Custom Task Schedulers

Question

I have a requirement to fire off web service requests to an online api and I thought that Parallel Extensions would be a good fit for my needs.

The web service in questi

Answer 1

I haven't used RX much, but AFAICT the Observable.Window method would work fine for this.

http://msdn.microsoft.com/en-us/library/system.reactive.linq.observable.window(VS.103).aspx

It would seem to be a better fit than Throttle which seems to throw elements away, which I'm guessing is not what you want

Answer 2

If you need to throttle by time, you should check out Quartz.net. It can facilitate consistent polling. If you care about all requests, you should consider using some sort of queueing mechanism. MSMQ is probably the right solution but there are many specific implementations if you want to go bigger and use an ESB like NServiceBus or RabbitMQ.

Update:

In that case, TPL Dataflow is your preferred solution if you can leverage the CTP. A throttled BufferBlock is the solution.

This example comes from the documentation provided by Microsoft:

// Hand-off through a bounded BufferBlock<T>
private static BufferBlock<int> m_buffer = new BufferBlock<int>(
    new DataflowBlockOptions { BoundedCapacity = 10 });

// Producer
private static async void Producer()
{
    while(true)
    {
        await m_buffer.SendAsync(Produce());
    }
}

// Consumer
private static async Task Consumer()
{
    while(true)
    {
        Process(await m_buffer.ReceiveAsync());
    }
}

// Start the Producer and Consumer
private static async Task Run()
{
    await Task.WhenAll(Producer(), Consumer());
}

Update:

Check out RX's Observable.Throttle.

Answer 3

I agree with others that TPL Dataflow sounds like a good solution for this.

To limit the processing, you could create a TransformBlock that doesn't actually transform the data in any way, it just delays it if it arrived too soon after the previous data:

static IPropagatorBlock<T, T> CreateDelayBlock<T>(TimeSpan delay)
{
    DateTime lastItem = DateTime.MinValue;
    return new TransformBlock<T, T>(
        async x =>
                {
                    var waitTime = lastItem + delay - DateTime.UtcNow;
                    if (waitTime > TimeSpan.Zero)
                        await Task.Delay(waitTime);

                    lastItem = DateTime.UtcNow;

                    return x;
                },
        new ExecutionDataflowBlockOptions { BoundedCapacity = 1 });
}

Then create a method that produces the data (for example integers starting from 0):

static async Task Producer(ITargetBlock<int> target)
{
    int i = 0;
    while (await target.SendAsync(i))
        i++;
}

It's written asynchronously, so that if the target block isn't able to process the items right now, it will wait.

Then write a consumer method:

static void Consumer(int i)
{
    Console.WriteLine(i);
}

And finally, link it all together and start it up:

var delayBlock = CreateDelayBlock<int>(TimeSpan.FromMilliseconds(500));

var consumerBlock = new ActionBlock<int>(
    (Action<int>)Consumer,
    new ExecutionDataflowBlockOptions { MaxDegreeOfParallelism = DataflowBlockOptions.Unbounded });

delayBlock.LinkTo(consumerBlock, new DataflowLinkOptions { PropagateCompletion = true });

Task.WaitAll(Producer(delayBlock), consumerBlock.Completion);

Here, delayBlock will accept at most one item every 500 ms and the Consumer() method can run multiple times in parallel. To finish processing, call delayBlock.Complete().

If you want to add some caching per your #2, you could create another TransformBlock do the work there and link it to the other blocks.

Answer 4

Honestly I would work at a higher level of abstraction and use the TPL Dataflow API for this. The only catch is you would need to write a custom block that will throttle the requests at the rate at which you need because, by default, blocks are "greedy" and will just process as fast as possible. The implementation would be something like this:

1. Start with a BufferBlock<T> which is the logical block that you would post to.
2. Link the BufferBlock<T> to a custom block which has the knowledge of requests/sec and throttling logic.
3. Link the custom block from 2 to to your ActionBlock<T>.

I don't have the time to write the custom block for #2 right this second, but I will check back later and try to fill in an implementation for you if you haven't already figured it out.