Request/Response 模式与 TPL 数据流

Question

我们在使用 TPL 数据流库时遇到问题，需要 request/response 模式。我们的问题是我们有一个调用依赖服务的 .NET 核心 API。依赖服务限制并发请求。我们的API不限制并发请求；因此，我们一次可以收到数千个请求。在这种情况下，依赖服务将在达到其限制后拒绝请求。因此，我们实现了一个BufferBlock<T>和一个TransformBlock<TIn, TOut>。表演很扎实，效果很好。我们测试了我们的 API 前端，有 1000 个用户发布了 100 个 requests/sec，0 个问题。缓冲块缓冲请求，转换块并行执行我们所需数量的请求。依赖服务接收我们的请求并响应。我们 return 在变换块操作中的响应一切都很好。我们的问题是缓冲块和转换块断开连接，这意味着 requests/responses 不同步。我们遇到了一个请求将收到另一个请求者的响应的问题（请参阅下面的代码）。

具体到下面的代码，我们的问题出在了GetContent这个方法上。该方法是从我们 API 中的服务层调用的，最终是从我们的控制器调用的。下面的代码和服务层都是单例。缓冲区的 SendAsync 与转换块 ReceiveAsync 断开连接，因此可以 return 编辑任意响应，不一定是发出的请求。

所以，我们的问题是：有没有办法使用数据流块来关联 request/responses？最终目标是请求进入我们的 API，被发送到依赖服务，然后被 return 发送到客户端。我们的数据流实现代码如下。

public class HttpClientWrapper : IHttpClientManager
{
    private readonly IConfiguration _configuration;
    private readonly ITokenService _tokenService;
    private HttpClient _client;

    private BufferBlock<string> _bufferBlock;
    private TransformBlock<string, JObject> _actionBlock;

    public HttpClientWrapper(IConfiguration configuration, ITokenService tokenService)
    {
        _configuration = configuration;
        _tokenService = tokenService;

        _bufferBlock = new BufferBlock<string>();

        var executionDataFlowBlockOptions = new ExecutionDataflowBlockOptions
        {
            MaxDegreeOfParallelism = 10
        };

        var dataFlowLinkOptions = new DataflowLinkOptions
        {
            PropagateCompletion = true
        };

        _actionBlock = new TransformBlock<string, JObject>(t => ProcessRequest(t),
            executionDataFlowBlockOptions);

        _bufferBlock.LinkTo(_actionBlock, dataFlowLinkOptions);
    }

    public void Connect()
    {
        _client = new HttpClient();

        _client.DefaultRequestHeaders.Add("x-ms-client-application-name",
            "ourappname");
    }

    public async Task<JObject> GetContent(string request)
    {
        await _bufferBlock.SendAsync(request);

        var result = await _actionBlock.ReceiveAsync();

        return result;
    }

    private async Task<JObject> ProcessRequest(string request)
    {
        if (_client == null)
        {
            Connect();
        }

        try
        {
            var accessToken = await _tokenService.GetTokenAsync(_configuration);

            var httpRequestMessage = new HttpRequestMessage(HttpMethod.Post,
                new Uri($"https://{_configuration.Uri}"));

            // add the headers
            httpRequestMessage.Headers.Add("Authorization", $"Bearer {accessToken}");
            // add the request body
            httpRequestMessage.Content = new StringContent(request, Encoding.UTF8,
                "application/json");

            var postRequest = await _client.SendAsync(httpRequestMessage);

            var response = await postRequest.Content.ReadAsStringAsync();

            return JsonConvert.DeserializeObject<JObject>(response);
        }
        catch (Exception ex)
        {
            // log error

            return new JObject();
        }
    }
}

Answer 1

您需要做的是用一个 id 标记每个传入的项目，以便您可以将数据输入与结果输出相关联。这是一个如何做到这一点的例子：

namespace ConcurrentFlows.DataflowJobs {
    using System;
    using System.Collections.Concurrent;
    using System.Collections.Generic;
    using System.Threading.Tasks;
    using System.Threading.Tasks.Dataflow;

    /// <summary>
    /// A generic interface defining that:
    /// for a specified input type => an awaitable result is produced.
    /// </summary>
    /// <typeparam name="TInput">The type of data to process.</typeparam>
    /// <typeparam name="TOutput">The type of data the consumer expects back.</typeparam>
    public interface IJobManager<TInput, TOutput> {
        Task<TOutput> SubmitRequest(TInput data);
    }

    /// <summary>
    /// A TPL-Dataflow based job manager.
    /// </summary>
    /// <typeparam name="TInput">The type of data to process.</typeparam>
    /// <typeparam name="TOutput">The type of data the consumer expects back.</typeparam>
    public class DataflowJobManager<TInput, TOutput> : IJobManager<TInput, TOutput> {

        /// <summary>
        /// It is anticipated that jobHandler is an injected
        /// singleton instance of a Dataflow based 'calculator', though this implementation
        /// does not depend on it being a singleton.
        /// </summary>
        /// <param name="jobHandler">A singleton Dataflow block through which all jobs are processed.</param>
        public DataflowJobManager(IPropagatorBlock<KeyValuePair<Guid, TInput>, KeyValuePair<Guid, TOutput>> jobHandler) {
            if (jobHandler == null) { throw new ArgumentException("Argument cannot be null.", "jobHandler"); }

            this.JobHandler = JobHandler;
            if (!alreadyLinked) {
                JobHandler.LinkTo(ResultHandler, new DataflowLinkOptions() { PropagateCompletion = true });
                alreadyLinked = true;
            }
        }

        private static bool alreadyLinked = false;            

        /// <summary>
        /// Submits the request to the JobHandler and asynchronously awaits the result.
        /// </summary>
        /// <param name="data">The input data to be processd.</param>
        /// <returns></returns>
        public async Task<TOutput> SubmitRequest(TInput data) {
            var taggedData = TagInputData(data);
            var job = CreateJob(taggedData);
            Jobs.TryAdd(job.Key, job.Value);
            await JobHandler.SendAsync(taggedData);
            return await job.Value.Task;
        }

        private static ConcurrentDictionary<Guid, TaskCompletionSource<TOutput>> Jobs {
            get;
        } = new ConcurrentDictionary<Guid, TaskCompletionSource<TOutput>>();

        private static ExecutionDataflowBlockOptions Options {
            get;
        } = GetResultHandlerOptions();

        private static ITargetBlock<KeyValuePair<Guid, TOutput>> ResultHandler {
            get;
        } = CreateReplyHandler(Options);

        private IPropagatorBlock<KeyValuePair<Guid, TInput>, KeyValuePair<Guid, TOutput>> JobHandler {
            get;
        }

        private KeyValuePair<Guid, TInput> TagInputData(TInput data) {
            var id = Guid.NewGuid();
            return new KeyValuePair<Guid, TInput>(id, data);
        }

        private KeyValuePair<Guid, TaskCompletionSource<TOutput>> CreateJob(KeyValuePair<Guid, TInput> taggedData) {
            var id = taggedData.Key;
            var jobCompletionSource = new TaskCompletionSource<TOutput>();
            return new KeyValuePair<Guid, TaskCompletionSource<TOutput>>(id, jobCompletionSource);
        }

        private static ExecutionDataflowBlockOptions GetResultHandlerOptions() {
            return new ExecutionDataflowBlockOptions() {
                MaxDegreeOfParallelism = Environment.ProcessorCount,
                BoundedCapacity = 1000
            };
        }

        private static ITargetBlock<KeyValuePair<Guid, TOutput>> CreateReplyHandler(ExecutionDataflowBlockOptions options) {
            return new ActionBlock<KeyValuePair<Guid, TOutput>>((result) => {
                RecieveOutput(result);
            }, options);
        }

        private static void RecieveOutput(KeyValuePair<Guid, TOutput> result) {
            var jobId = result.Key;
            TaskCompletionSource<TOutput> jobCompletionSource;
            if (!Jobs.TryRemove(jobId, out jobCompletionSource)) {
                throw new InvalidOperationException($"The jobId: {jobId} was not found.");
            }
            var resultValue = result.Value;
            jobCompletionSource.SetResult(resultValue);            
        }
    }
}

另见以供参考。

Answer 2

感谢 JSteward 提供的答案。他的方法是完全可以接受的；但是，我最终通过使用 SemaphoreSlim 来完成此操作。 SemaphoreSlim 提供了两个功能，使其成为一个强大的解决方案。首先，它提供了一个构造函数重载，您可以在其中发送一个计数。此计数指的是能够通过信号量等待机制的并发项目数。等待机制由称为 WaitAsync 的方法提供。使用以下方法，其中 Worker class 作为单例，并发请求进入，每次执行 http 请求时限制为 10 个，并且响应全部返回到正确的请求。因此，实现可能如下所示：

public class Worker: IWorker
{
    private readonly IHttpClientManager _httpClient;
    private readonly ITokenService _tokenService;

    private readonly SemaphoreSlim _semaphore;

    public Worker(IHttpClientManager httpClient, ITokenService tokenService)
    {
        _httpClient = httpClient;
        _tokenService = tokenService;

        // we want to limit the number of items here
        _semaphore = new SemaphoreSlim(10);
    }

    public async Task<JObject> ProcessRequestAsync(string request, string route)
    {
        try
        {
            var accessToken = await _tokenService.GetTokenAsync(
                _timeSeriesConfiguration.TenantId,
                _timeSeriesConfiguration.ClientId,
                _timeSeriesConfiguration.ClientSecret);

            var cancellationToken = new CancellationTokenSource();

            cancellationToken.CancelAfter(30000);

            await _semaphore.WaitAsync(cancellationToken.Token);
            var httpResponseMessage = await _httpClient.SendAsync(new HttpClientRequest
            {
                Method = HttpMethod.Post,
                Uri = $"https://someuri/someroute",
                Token = accessToken,
                Content = request
            });

            var response = await httpResponseMessage.Content.ReadAsStringAsync();

            return response;
        }
        catch (Exception ex)
        {
            // do some logging

            throw;
        }
        finally
        {
            _semaphore.Release();
        }
    }
}

Answer 3

进行简单的节流并不是 TPL 数据流库特别有吸引力的用例，而使用 SemaphoreSlim 似乎更简单也更有吸引力。但是，如果您想要更多功能，例如为每个请求强制执行最短持续时间，或者有一种方法可以等待所有待处理的请求完成，那么 TPL 数据流可以提供 SemaphoreSlim 无法提供的功能.基本思想是避免将裸输入值传递给块，然后尝试将它们与生成的结果相关联。根据请求立即创建任务、将任务发送到 ActionBlock<Task> 并让块激活并 await 使用其指定的 MaxDegreeOfParallelism 异步执行这些任务要安全得多。这样输入值及其结果将永远明确地绑定在一起。

public class ThrottledExecution<T>
{
    private readonly ActionBlock<Task<Task<T>>> _actionBlock;
    private readonly CancellationToken _cancellationToken;

    public ThrottledExecution(int concurrencyLevel, int minDurationMilliseconds = 0,
        CancellationToken cancellationToken = default)
    {
        if (minDurationMilliseconds < 0) throw new ArgumentOutOfRangeException();
        _actionBlock = new ActionBlock<Task<Task<T>>>(async task =>
        {
            try
            {
                var delay = Task.Delay(minDurationMilliseconds, cancellationToken);
                task.RunSynchronously();
                await task.Unwrap().ConfigureAwait(false);
                await delay.ConfigureAwait(false);
            }
            catch { } // Ignore exceptions (errors are propagated through the task)
        }, new ExecutionDataflowBlockOptions()
        {
            MaxDegreeOfParallelism = concurrencyLevel,
            CancellationToken = cancellationToken,
        });
        _cancellationToken = cancellationToken;
    }

    public Task<T> Run(Func<Task<T>> function)
    {
        // Create a cold task (the function will be invoked later)
        var task = new Task<Task<T>>(function, _cancellationToken);
        var accepted = _actionBlock.Post(task);
        if (!accepted)
        {
            _cancellationToken.ThrowIfCancellationRequested();
            throw new InvalidOperationException(
                "The component has been marked as complete.");
        }
        return task.Unwrap();
    }

    public void Complete() => _actionBlock.Complete();
    public Task Completion => _actionBlock.Completion;
}

用法示例：

private ThrottledExecution<JObject> throttledExecution
    = new ThrottledExecution<JObject>(concurrencyLevel: 10);

public Task<JObject> GetContent(string request)
{
    return throttledExecution.Run(() => ProcessRequest(request));
}

Request/Response 模式与 TPL 数据流

Request/Response pattern with TPL Dataflow

c#

.net

task-parallel-library

tpl-dataflow