具有指定结果的任务并行库 WaitAny

Question

我正在尝试编写一些代码来并行调用多个不同服务器的 Web 服务，因此 TPL 似乎是显而易见的选择。

我的 Web 服务调用中只有一个会返回我想要的结果，而其他所有调用都不会。我正在尝试找出一种有效地拥有Task.WaitAny 的方法，但只有在第一个匹配条件的Task 返回时才解除阻塞。

我尝试使用WaitAny，但无法确定过滤器的放置位置。我已经走到这一步了：

public void SearchServers()
{
    var servers = new[] {"server1", "server2", "server3", "server4"};
    var tasks = servers
                 .Select(s => Task<bool>.Factory.StartNew(server => CallServer((string)server), s))
                 .ToArray();

    Task.WaitAny(tasks); //how do I say "WaitAny where the result is true"?

    //Omitted: cancel any outstanding tasks since the correct server has been found
}

private bool CallServer(string server)
{
    //... make the call to the server and return the result ...
}

编辑：快速澄清，以防上面有任何混淆。我正在尝试执行以下操作：

1. 对于每个服务器，启动一个Task 来检查它
2. 或者，等到服务器返回 true（最多只有 1 个服务器会返回 true）
3. 或者，等到所有服务器都返回 false，即没有匹配项。

Answer 1

我能想到的最好的方法是为每个Task 指定一个ContinueWith，检查结果，如果true 取消其他任务。对于取消任务，您可能需要使用CancellationToken。

var tasks = servers
    .Select(s => Task.Run(...)
        .ContinueWith(t =>
            if (t.Result) {
                // cancel other threads
            }
        )
    ).ToArray();

更新：另一种解决方案是 WaitAny 直到正确的任务完成（但它有一些缺点，例如从列表中删除已完成的任务并从剩余的任务中创建一个新数组是相当繁重的操作） ：

List<Task<bool>> tasks = servers.Select(s => Task<bool>.Factory.StartNew(server => CallServer((string)server), s)).ToList();

bool result;
do {
    int idx = Task.WaitAny(tasks.ToArray());
    result = tasks[idx].Result;
    tasks.RemoveAt(idx);
} while (!result && tasks.Count > 0);

// cancel other tasks

更新 2：现在我会用 Rx 来做：

[Fact]
public async Task AwaitFirst()
{
    var servers = new[] { "server1", "server2", "server3", "server4" };
    var server = await servers
        .Select(s => Observable
            .FromAsync(ct => CallServer(s, ct))
            .Where(p => p)
            .Select(_ => s)
        )
        .Merge()
        .FirstAsync();
    output.WriteLine($"Got result from {server}");
}

private async Task<bool> CallServer(string server, CancellationToken ct)
{
    try
    {
        if (server == "server1")
        {
            await Task.Delay(TimeSpan.FromSeconds(1), ct);
            output.WriteLine($"{server} finished");
            return false;
        }
        if (server == "server2")
        {
            await Task.Delay(TimeSpan.FromSeconds(2), ct);
            output.WriteLine($"{server} finished");
            return false;
        }
        if (server == "server3")
        {
            await Task.Delay(TimeSpan.FromSeconds(3), ct);
            output.WriteLine($"{server} finished");
            return true;
        }
        if (server == "server4")
        {
            await Task.Delay(TimeSpan.FromSeconds(4), ct);
            output.WriteLine($"{server} finished");
            return true;
        }
    }
    catch(OperationCanceledException)
    {
        output.WriteLine($"{server} Cancelled");
        throw;
    }

    throw new ArgumentOutOfRangeException(nameof(server));
}

在我的机器上测试需要 3.32 秒（这意味着它没有等待第 4 个服务器），我得到以下输出：

server1 finished
server2 finished
server3 finished
server4 Cancelled
Got result from server3

Answer 2

您可以使用OrderByCompletion() from the AsyncEx library，它会在任务完成时返回它们。您的代码可能类似于：

var tasks = servers
    .Select(s => Task.Factory.StartNew(server => CallServer((string)server), s))
    .OrderByCompletion();

foreach (var task in tasks)
{
    if (task.Result)
    {
        Console.WriteLine("found");
        break;
    }
    Console.WriteLine("not found yet");
}

// cancel any outstanding tasks since the correct server has been found

Answer 3

使用 Interlocked.CompareExchange 就可以做到这一点，只有一个 Task 能够在 serverReturedData 上写入

    public void SearchServers()
        {
            ResultClass serverReturnedData = null;
            var servers = new[] {"server1", "server2", "server3", "server4"};
            var tasks = servers.Select(s => Task<bool>.Factory.StartNew(server => 
            {
               var result = CallServer((string)server), s);
               Interlocked.CompareExchange(ref serverReturnedData, result, null);

            }).ToArray();

            Task.WaitAny(tasks); //how do I say "WaitAny where the result is true"?
        //
        // use serverReturnedData as you want.
        // 
        }

编辑：正如 Jasd 所说，上面的代码可以在变量 serverReturnedData 具有有效值之前返回（如果服务器返回空值，这可能会发生），以确保您可以将结果包装在自定义对象中。

Answer 4

这是基于 svick 回答的通用解决方案：

public static async Task<T> GetFirstResult<T>(
this IEnumerable<Func<CancellationToken, Task<T>>> taskFactories, 
Action<Exception> exceptionHandler,
Predicate<T> predicate)
{
    T ret = default(T);
    var cts = new CancellationTokenSource();
    var proxified = taskFactories.Select(tf => tf(cts.Token)).ProxifyByCompletion();
    int i;
    for (i = 0; i < proxified.Length; i++)
    {
        try
        {
            ret = await proxified[i].ConfigureAwait(false);
        }
        catch (Exception e)
        {
            exceptionHandler(e);
            continue;
        }
        if (predicate(ret))
        {
            break;
        }
    }

    if (i == proxified.Length)
    {
        throw new InvalidOperationException("No task returned the expected value");
    }
    cts.Cancel(); //we have our value, so we can cancel the rest of the tasks
    for (int j = i+1; j < proxified.Length; j++)
    {
        //observe remaining tasks to prevent process crash 
        proxified[j].ContinueWith(
         t => exceptionHandler(t.Exception), TaskContinuationOptions.OnlyOnFaulted)
                   .Forget();
    }
    return ret;
}

ProxifyByCompletion 的实现方式为：

public static Task<T>[] ProxifyByCompletion<T>(this IEnumerable<Task<T>> tasks)
{
    var inputTasks = tasks.ToArray();
    var buckets = new TaskCompletionSource<T>[inputTasks.Length];
    var results = new Task<T>[inputTasks.Length];
    for (int i = 0; i < buckets.Length; i++)
    {
        buckets[i] = new TaskCompletionSource<T>();
        results[i] = buckets[i].Task;
    }
    int nextTaskIndex = -1;
    foreach (var inputTask in inputTasks)
    {
        inputTask.ContinueWith(completed =>
        {
            var bucket = buckets[Interlocked.Increment(ref nextTaskIndex)];
            if (completed.IsFaulted)
            {
                Trace.Assert(completed.Exception != null);
                bucket.TrySetException(completed.Exception.InnerExceptions);
            }
            else if (completed.IsCanceled)
            {
                bucket.TrySetCanceled();
            }
            else
            {
                bucket.TrySetResult(completed.Result);
            }
        }, CancellationToken.None, 
           TaskContinuationOptions.ExecuteSynchronously, TaskScheduler.Default);
    }
    return results;
}

而Forget是抑制CS4014的空方法：

public static void Forget(this Task task) //suppress CS4014
{
}