Posts Tagged ‘task’

19 Aug 11

Asynchrony and sequential workflows

Rather than finishing this series on asynchrony before my Monospace 2011 talk on the same subject, I got busy finishing the materials for the talk instead. The slides and example code from the talk can be found here, and the video will be posted on InfoQ in the near future. Since the examples are slightly different and the slides less explanatory, I will continue this series of posts as a more in-depth version of the content in the slides.

In the last post, Exit… Screen Right, I talked about using the CPS (continuation passing style) to chain asynchronous methods together. While this is a proven way of dealing with asynchrony, it puts the how in front of the what, which makes our code harder to maintain and debug. Ideally we'd continue writing our code in the regular sequential style we've always used but steps that do not happen synchronously, suspend the flow and resume once the step is completed instead of blocking their thread. With a regular method that's not possible, since it controls the flow of execution until it exits, but there exists a variation of the trusty method (which really is just a subroutine), called coroutine:

Coroutines are computer program components that generalize subroutines to allow multiple entry points for suspending and resuming execution at certain locations

Wikipedia

With Coroutines, we can suspend the current execution flow, allow an async action to complete and then have its continuation resumes our flow of execution. Great. Except .NET doesn't have Coroutines, right? Not exactly, but it does have generators for building Iterators, in the form of IEnumerator<T> and the yield keyword. Since Generators are a specialized form of Coroutines it is possible to use them to mimic Coroutine behavior. In addtion, the AsyncCTP for C# 5.0 introduces the async/await constructs, which provide a better syntax for suspending and resuming a method.

DReAM Coroutines are built on top an inversion of the Iterator pattern, which allowed us to introduce them in .NET 2.0, while Task based Coroutines won't be possible in production until .NET 5.0 ships. Both provide us the ability to write sequential code that suspends on asynchronous work, preserving our workflow.

Review of our Async Workflow Methods

Below are the asynchronous method signatures from last time:

1. GetUserRecentArticles

This method implements the asynchronous workflow.

// Result
Result<DreamMessage> GetUserRecentArticles(int userId, int limit, Result<DreamMessage> response)
// Task
Task<DreamMessage> GetUserRecentArticles(int userId, int limit)

DreamMessage is DReAM's symmetric Http Request and Response data container, and serves as the value of the synchronization handle seen by the outside caller. The Http Server would call GetUserRecentArticles to start the asynchronous chain and use the synchronization handle to retrieve the final result.

2. GetFeedsForUserFromDb

This method uses asynchronous database queries to get the list of Uri's the user is subscribed to:

// Result
Result<IEnumerable<XUri>> GetFeedsForUserFromDb(int userId, Result<IEnumerable<XUri>> result)
// Task
Task<IEnumerable<XUri>> GetFeedsForUserFromDb(int userId)

3. FetchFeed

This method uses an asynchronous web request to fetch the feed for a given Uri. We will be multiplexing this call to fetch multiple feeds at once.

// Result
Result<XDoc> FetchFeed(XUri uri, Result<XDoc> result)
//Task
Task<XDoc> FetchFeed(XUri uri)

4. FilterViaDb

This method is called once for each feed that we've fetched to turn into a list of only the articles not yet seen by the user

// Result
Result<IEnumerable<XDoc>> FilterViaDb(XDoc doc, Result<IEnumerable<XDoc>> result)
// Task
Task<IEnumerable<XDoc>> FilterViaDb(XDoc doc)

5. MarkAsRead

Step 5 combines the articles, applies the limit and marks the articles to be returned as read. Only the last part is an asynchronous operation, the rest can be performed inline. The method MarkAsRead makes asynchronous calls to the database to insert the canonical Uri's of the articles to be returned to mark them as read.

//Result
Result MarkAsRead(int userId, IEnumerable<XDoc> aggregatedArticles, Result result)
// Task
Task MarkAsRead(int userId, IEnumerable<XDoc> aggregatedArticles)

DReAM coroutine

A DReAM Coroutine must produce an Enumerator of type IYield, which is the coordination construct for Coroutines and is implemented by Result, i.e. any method returning a Result can be yielded to in a Coroutine. Since this signature requirement would change the expected signature of GetUserRecentArticles from the DReAM Asynchronous Method Pattern and the usage of a Coroutine is really an implementation detail, we hide the actual Coroutine:

private Result<DreamMessage> GetUserRecentArticles(int userId, int limit, Result<DreamMessage> response) {
  return Coroutine.Invoke(GetUserRecentArticles_Co, userId, limit, response);
}

We use Coroutine.Invoke to execute the enumerator that GetUserRecentArticles_Co returns. The actual implementation now has the familiar sequential flow we are used to:

public Yield GetUserRecentArticles_Co(int userId, int limit, Result response) {

    // yield execution to fetch the user's feed uris from the DB
    IEnumerable uris = null;
    yield return GetFeedsForUserFromDb(userId, new Result>()).Set(x => uris = x);

    // start to fetch all feeds
    var feeds = uris.Select(uri => FetchFeed(uri, new Result())).ToList();

    // yield execution until all fetch calls have signaled completion
    yield return feeds.Join(new Result());

    // start to filter all feeds to only unread
    var allUnreadArticles = feeds
        .Select(feedResult => FilterViaDb(feedResult.Value, new Result>()))
        .ToList();

    // yield execution until all filter calls have signaled completion
    yield return allUnreadArticles.Join(new Result());

    // combine all filtered articles
    var combinedArticles = (from unreadResult in allUnreadArticles
                            let unreadArticles = unreadResult.Value
                            from article in unreadArticles
                            select article).Take(limit).ToList();

    // yield execution to async method marking all remaining articles as read
    yield return MarkAsRead(userId, combinedArticles, new Result());
    var aggregatedDoc = new XDoc("articles").AddAll(combinedArticles);

    // signal response with aggregated article document
    response.Return(DreamMessage.Ok(aggregatedDoc));
}

Instead of chaining .WhenDone on each async method, we now call them with yield return instead, which resumes the body of the method after the async call completes. We still use response.Return to signal completion, since the Iterator doesn't regular return value. The one caveat causing some syntactic cruft is that we cannot create and assign the value in the yield return. Instead we have to declare it before and use the .Set() extension method, which unwraps the result, allowing us to capture it with a lambda.

Coroutine flow with async/await

async/await allows the same coroutine flow we create with an Iterator and Result, but since it is custom built into the compiler around Task, it achieves additional syntactic magic for very readable and natural code:

public async Task GetUserRecentArticles(int userId, int limit) {

    // Await the fetching of the user's feed uris from the DB
    var uris = await GetFeedsForUserFromDb(userId);

    // start fetch all feeds
    var feedTasks = uris.Select(FetchFeed).ToList();

    // Await all feeds being fetched
    var feeds = await  TaskEx.WhenAll(feedTasks);

    // start to filter all feeds to only unread
    var unreadArticleTasks = feeds.Select(FilterViaDb).ToList();

    // Await feeds completing filtering
    var allUnreadArticles = await TaskEx.WhenAll(unreadArticleTasks);

    // combine all filtered articles
    var combinedArticles = (from unreadArticles in allUnreadArticles
                            from article in unreadArticles
                            select article).Take(limit).ToList();

    // Await the database call to mark all remaining articles as read
    await MarkAsRead(userId, combinedArticles);
    var aggregatedDoc = new XDoc("articles").AddAll(combinedArticles);

    // return the final result, which in turn sets it on the Task
    // that had already been returned for us
    return DreamMessage.Ok(aggregatedDoc);
}

With async/await, the only difference between a workflow where all of our calls are synchronous is that we have an await in front of every async method now. That's a really low syntax tax to pay for completely non-blocking workflows! The other thing you may notice is that the signature of the method is Task<DreamMessage> while we return just a DreamMessage. This is possible, because the method is marked with async, which tells the compiler to rewrite the method body as a series of continuations interrupted by await barriers.

But how does it work?

Both yield and async/await are constructs that tell the compiler that the body of the subroutine should be rewritten as a state machine. They make it easy for us to think of the flow of execution as a regular sequential flow, while handling the complexities of tracking the local scope when execution is suspended and resumed. In simplistic terms, a new anonymous class is generated with members to capture the locally available variables of the method. The body of the method is broken up into linear blocks of execution at the yield/await barriers. This way each block can be run until the next barrier, store to which point it was last run to and exit and upon the next invocation continue running the next block.

In my next post, I will break down how you can use this behavior to roll your own Coroutines instead of just using yield to return values in a sequence or await to wait for an asynchronous method to complete.

21 Jun 11

Exit… Screen Right

This article is a continuation in a series about asynchrony related to DReAM Coroutine's and C#5.0 async/await constructs in preparation for my talk on the same subject at monospace in July in Boston. In the last post "What is asynchrony good for", I introduced Result.WhenDone and Task.ContinueWith, two constructs that facilitate continuation passing style programming in C#. This time I want to tackle a workflow that is a little more complicated. It includes some parallelization of work and a number of I/O operations. Each of the below steps is an async barrier that would traditionally result in a blocking call of indeterminate duration:

  1. A request to fetch unread aggregated articles comes in
  2. The list of feeds the user is subscribed to are fetched
  3. All feeds are fetched in parallel
  4. All feeds are filtered via the DB to remove the already seen articles
  5. Combine all unseen articles, limit the max article to return and mark them as seen in the DB
  6. The limited, unseen entries are returned as the response

Each of the steps in the above workflow is separated from the previous by a potentially asynchronous operation. If we think of each step as a callback for the completion of the previous the continuation arguments fall out naturally.

The Players in our Async Workflow Production

For clarity and brevity, I'm going to treat each step as an existing async method with its body left as an exercise. Below I've defined the asynchronous method signatures we will need:

1. GetUserRecentArticles

This method implements the asynchronous workflow.

// Result
Result GetUserRecentArticles(int userId, int limit, Result response)
// Task
Task GetUserRecentArticles(int userId, int limit)

DreamMessage is DReAM's symmetric Http Request and Response data container, and serves as the value of the synchronization handle seen by the outside caller. The Http Server would call GetUserRecentArticles to start the asynchronous chain and use the synchronization handle to retrieve the final result.

2. GetFeedsForUserFromDb

This method uses asynchronous database queries to get the list of Uri's the user is subscribed to:

// Result
Result> GetFeedsForUserFromDb(int userId, Result> result)
// Task
Task> GetFeedsForUserFromDb(int userId)

3. FetchFeed

This method uses an asynchronous web request to fetch the feed for a given Uri. We will be multiplexing this call to fetch multiple feeds at once.

// Result
Result FetchFeed(XUri uri, Result result)
//Task
Task FetchFeed(XUri uri)

4. FilterViaDb

This method is called once for each feed that we've fetched to turn into a list of only the articles not yet seen by the user

// Result
Result> FilterViaDb(XDoc doc, Result> result)
// Task
Task> FilterViaDb(XDoc doc)

5. MarkAsRead

Step 5 combines the articles, applies the limit and marks the articles to be returned as read. Only the last part is an asynchronous operation, the rest can be performed inline. The method MarkAsRead makes asynchronous calls to the database to insert the canonical Uri's of the articles to be returned to mark them as read.

//Result
Result MarkAsRead(int userId, IEnumerable aggregatedArticles, Result result)
// Task
Task MarkAsRead(int userId, IEnumerable aggregatedArticles)

Act I: Result oriented

Below is the implementation of GetUserRecentArticles, which creates the asynchronous workflow calling the remaining methods in turn as continuations. This method will immediately return providing the Result instance to block on or attach a continuation:

public Result GetUserRecentArticles(int userId, int limit, Result response) {

    // Start to fetch the users feed uri's from the DB and set up continuation
    GetFeedsForUserFromDb(userId, new Result>()).WhenDone(r1 => {
        var uris = r1.Value;

        // start fetch all feeds
        var feeds = uris.Select(uri => FetchFeed(uri, new Result())).ToList();

        // set up continuation for when all feeds are fetched
        feeds.Join(new Result()).WhenDone(r2 => {

            // start to filter all feeds to only unread
            var allUnreadArticles = feeds
                .Select(feedResult => FilterViaDb(feedResult.Value, new Result>()))
                .ToList();

            // set up continuation for when feeds are filtered
            allUnreadArticles.Join(new Result()).WhenDone(r3 => {

                // combine all filtered articles
                var combinedArticles = (from unreadResult in allUnreadArticles
                                        let unreadArticles = unreadResult.Value
                                        from article in unreadArticles
                                        select article).Take(limit).ToList();

                // start to mark all remaining articles as read and set up continuation
                MarkAsRead(userId, combinedArticles, new Result()).WhenDone(r4 => {
                    var aggregatedDoc = new XDoc("articles").AddAll(combinedArticles);

                    // signal response with aggregated article document
                    response.Return(DreamMessage.Ok(aggregatedDoc));
                });
            });
        });
    });

    // immediately return the response Result object to the Caller
    return response;
}

As promised in the last article, the parallel fan-out for FetchFeed and FilterViaDb show the benefit of Result.Join(). Since it returns another Result, we can easily attach the next continuation for when all parallel operations complete to it, rather than requiring blocking behavior to join our workers back together.

Also note, that for simplicity, I set up the parallel loops for fetching and filtering the feeds as two separate blocks. This could easily be refined to only fan out once by attaching the FilterViaDb continuation to each FetchFeed call and joining back after each chain is finished, instead of after each fan-out.

Act II: A Task completed

The Task version of GetUserRecentArticles is very similar to the Result version and once again returns immediately with the Task to block on or attach a continuation to.

public Task GetUserRecentArticles(int userId, int limit) {
    var response = new TaskCompletionSource();

    // Start to fetch the users feed uri's from the DB and set up continuation
    GetFeedsForUserFromDb(userId).ContinueWith(t1 => {
        var uris = t1.Result;

        // start fetch all feeds
        var feedTasks = uris.Select(FetchFeed).ToList();

        // set up continuation for when all feeds are fetched
        TaskEx.WhenAll(feedTasks).ContinueWith(t2 => {
            var feeds = t2.Result;

            // start to filter all feeds to only unread
            var unreadArticleTasks = feeds.Select(FilterViaDb).ToList();

            // set up continuation for when feeds are filtered
            TaskEx.WhenAll(unreadArticleTasks).ContinueWith(t3 => {
                var allUnreadArticles = t3.Result;

                // combine all filtered articles
                var combinedArticles = (from unreadArticles in allUnreadArticles
                                        from article in unreadArticles
                                        select article).Take(limit).ToList();
                MarkAsRead(userId, combinedArticles).ContinueWith(t4 => {
                    var aggregatedDoc = new XDoc("articles").AddAll(combinedArticles);

                    // signal response with aggregated article document
                    response.SetResult(DreamMessage.Ok(aggregatedDoc));
                });
            });
        });
    });

    // immediately return the response Task object to the caller
    return response.Task;
}

The Task version also uses the simpler double-fan-out for the parallel operations the Result version did, but gains some additional compactness because of Task continuations being chainable. In Result the chaining pattern is meant as a fluent interface, i.e. the same result is returned by each operation that attaches behavior to a Result. Task on the other hand does not provide a fluent interface, requiring each method to be called on the original instance in turn; This allows its members to return other values. Task.ContinueWith in particular takes and Action or Func as its continuation and returns a new Task instance that captures the continuation's execution and can in turn take a continuation.

Another difference is that Result is meant to be used as a signal, while Task by default wraps an operation. This means  Result.Value can be set manually, while Task.Result cannot. In order for GetUserRecentArticles to return a Task than can be signaled on completion (and not schedule a new Thread to do it), we use TaskCompletionSource. This class has a Task property that we return allowing us signal completion in final continuation to return the list of filtered, aggregated articles.

Epilogue: It's Turtles All The Way Down

One drawback to asynchrony is that once you go async, you gotta go async all the way up and down the call chain. If we were to use GetUserRecentArticles inside an action method on an MVC controller, we'd have to block to get the result, which defeats the purpose of the whole exercise.

Fortunately, this can usually be accommodated, even if it means stepping outside the traditional frameworks. In most scenarios, we are performing the work in order to return a result to our caller, and commonly, this caller is either waiting at a GUI or calling across a network boundary. The former already has concepts of marshaling events occurring off the UI thread back to the UI, which allows our async workflow to return report back without blocking, while the latter involves I/O and the incoming API is likely to have a Begin*/End* variant for starting and ending the I/O request.

Another drawback to this, and all continuation passing styles, is the tendency for the code to wander off the right hand side of the screen, as each continuation introduces another level of nesting. While this may just appear to be a visual artifact that can be alleviated by formatting, it really is a warning of much more complicated issues that affect debugging and maintaince of this.

In order to do continuation passing right, there should be no further work in the context that attches the continuation, but other than diligence there's nothing to stop us from putting code after one of those }); terminators. Any code executing after the continuation has potentially started, is now in danger of the local variables being accessed concurrently, which introduces the same probems that make multi-threaded code so hard to debug.

And the problem goes beyond that: While the workflow still looks inherently linear, traditional block scope constructs are not available since we only enter, never exit chained scopes. That means using statements as well as try/catch/finally do not work. While each continuation can inspect the result of its antecedent, we now have to manually check for exceptions in the async methods and catch exceptions in our own continuation scope to trap, handle or forward errors correctly.

Next time, I'll introduce constructs that let us write the workflow in linear fashion like a regular, blocking workflow, but still take advantage of the asynchronous behavior like the above.

Copyright © 2011 MindTouch, Inc. Powered by