.NET Core开发日志——HttpClientFactory

当需要向某特定URL地址发送HTTP请求并得到相应响应时，通常会用到HttpClient类。该类包含了众多有用的方法，可以满足绝大多数的需求。但是如果对其使用不当时，可能会出现意想不到的事情。

博客园官方团队就遇上过这样的问题，国外博主也记录过类似的情况，YOU'RE USING HTTPCLIENT WRONG AND IT IS DESTABILIZING YOUR SOFTWARE。

究其缘由是一句看似正确的代码引起的：

using(var client = new HttpClient())

对象所占用资源应该确保及时被释放掉，但是，对于网络连接而言，这是错误的。

原因有二，网络连接是需要耗费一定时间的，频繁开启与关闭连接，性能会受影响；再者，开启网络连接时会占用底层socket资源，但在HttpClient调用其本身的Dispose方法时，并不能立刻释放该资源，这意味着你的程序可能会因为耗尽连接资源而产生预期之外的异常。

所以比较好的解决方法是延长HttpClient对象的使用寿命，比如对其建一个静态的对象：

private static HttpClient Client = new HttpClient();

但从程序员的角度来看，这样的代码或许不够优雅。

所以在.NET Core 2.1中引入了新的HttpClientFactory类。

它的用法很简单，首先是对其进行IoC的注册：

public void ConfigureServices(IServiceCollection services)
{
    services.AddHttpClient();
    services.AddMvc();
}

然后通过IHttpClientFactory创建一个HttpClient对象，之后的操作如旧，但不需要担心其内部资源的释放：

public class MyController : Controller
{
    IHttpClientFactory _httpClientFactory;

    public MyController(IHttpClientFactory httpClientFactory)
    {
        _httpClientFactory = httpClientFactory;
    }

    public IActionResult Index()
    {
        var client = _httpClientFactory.CreateClient();
        var result = client.GetStringAsync("http://myurl/");
        return View();
    }
}

第一眼瞧去，可能不明白AddHttpClient方法与IHttpClientFactory有什么关系，但查到其源码后就能一目了然：

public static IServiceCollection AddHttpClient(this IServiceCollection services)
{
    if (services == null)
    {
        throw new ArgumentNullException(nameof(services));
    }

    services.AddLogging();
    services.AddOptions();

    //
    // Core abstractions
    //
    services.TryAddTransient<HttpMessageHandlerBuilder, DefaultHttpMessageHandlerBuilder>();
    services.TryAddSingleton<IHttpClientFactory, DefaultHttpClientFactory>();

    //
    // Typed Clients
    //
    services.TryAdd(ServiceDescriptor.Singleton(typeof(ITypedHttpClientFactory<>), typeof(DefaultTypedHttpClientFactory<>)));

    //
    // Misc infrastructure
    //
    services.TryAddEnumerable(ServiceDescriptor.Singleton<IHttpMessageHandlerBuilderFilter, LoggingHttpMessageHandlerBuilderFilter>());

    return services;
}

它的内部为IHttpClientFactory接口绑定了DefaultHttpClientFactory类。

再看IHttpClientFactory接口中关键的CreateClient方法：

public HttpClient CreateClient(string name)
{
    if (name == null)
    {
        throw new ArgumentNullException(nameof(name));
    }

    var entry = _activeHandlers.GetOrAdd(name, _entryFactory).Value;
    var client = new HttpClient(entry.Handler, disposeHandler: false);

    StartHandlerEntryTimer(entry);

    var options = _optionsMonitor.Get(name);
    for (var i = 0; i < options.HttpClientActions.Count; i++)
    {
        options.HttpClientActions[i](client);
    }

    return client;
}

HttpClient的创建不再是简单的new HttpClient()，而是传入了两个参数：HttpMessageHandler handler与bool disposeHandler。disposeHandler参数为false值时表示要重用内部的handler对象。handler参数则从上一句的代码可以看出是以name为键值从一字典中取出，又因为DefaultHttpClientFactory类是通过TryAddSingleton方法注册的，也就意味着其为单例，那么这个内部字典便是唯一的，每个键值对应的ActiveHandlerTrackingEntry对象也是唯一，该对象内部中包含着handler。

下一句代码StartHandlerEntryTimer(entry); 开启了ActiveHandlerTrackingEntry对象的过期计时处理。默认过期时间是2分钟。

internal void ExpiryTimer_Tick(object state)
{
    var active = (ActiveHandlerTrackingEntry)state;

    // The timer callback should be the only one removing from the active collection. If we can't find
    // our entry in the collection, then this is a bug.
    var removed = _activeHandlers.TryRemove(active.Name, out var found);
    Debug.Assert(removed, "Entry not found. We should always be able to remove the entry");
    Debug.Assert(object.ReferenceEquals(active, found.Value), "Different entry found. The entry should not have been replaced");

    // At this point the handler is no longer 'active' and will not be handed out to any new clients.
    // However we haven't dropped our strong reference to the handler, so we can't yet determine if
    // there are still any other outstanding references (we know there is at least one).
    //
    // We use a different state object to track expired handlers. This allows any other thread that acquired
    // the 'active' entry to use it without safety problems.
    var expired = new ExpiredHandlerTrackingEntry(active);
    _expiredHandlers.Enqueue(expired);

    Log.HandlerExpired(_logger, active.Name, active.Lifetime);

    StartCleanupTimer();
}

先是将ActiveHandlerTrackingEntry对象传入新的ExpiredHandlerTrackingEntry对象。

public ExpiredHandlerTrackingEntry(ActiveHandlerTrackingEntry other)
{
    Name = other.Name;

    _livenessTracker = new WeakReference(other.Handler);
    InnerHandler = other.Handler.InnerHandler;
}

在其构造方法内部，handler对象通过弱引用方式关联着，不会影响其被GC释放。

然后新建的ExpiredHandlerTrackingEntry对象被放入专用的队列。

最后开始清理工作，定时器的时间间隔设定为每10秒一次。

internal void CleanupTimer_Tick(object state)
{
    // Stop any pending timers, we'll restart the timer if there's anything left to process after cleanup.
    //
    // With the scheme we're using it's possible we could end up with some redundant cleanup operations.
    // This is expected and fine.
    // 
    // An alternative would be to take a lock during the whole cleanup process. This isn't ideal because it
    // would result in threads executing ExpiryTimer_Tick as they would need to block on cleanup to figure out
    // whether we need to start the timer.
    StopCleanupTimer();

    try
    {
        if (!Monitor.TryEnter(_cleanupActiveLock))
        {
            // We don't want to run a concurrent cleanup cycle. This can happen if the cleanup cycle takes
            // a long time for some reason. Since we're running user code inside Dispose, it's definitely
            // possible.
            //
            // If we end up in that position, just make sure the timer gets started again. It should be cheap
            // to run a 'no-op' cleanup.
            StartCleanupTimer();
            return;
        }

        var initialCount = _expiredHandlers.Count;
        Log.CleanupCycleStart(_logger, initialCount);

        var stopwatch = ValueStopwatch.StartNew();

        var disposedCount = 0;
        for (var i = 0; i < initialCount; i++)
        {
            // Since we're the only one removing from _expired, TryDequeue must always succeed.
            _expiredHandlers.TryDequeue(out var entry);
            Debug.Assert(entry != null, "Entry was null, we should always get an entry back from TryDequeue");

            if (entry.CanDispose)
            {
                try
                {
                    entry.InnerHandler.Dispose();
                    disposedCount++;
                }
                catch (Exception ex)
                {
                    Log.CleanupItemFailed(_logger, entry.Name, ex);
                }
            }
            else
            {
                // If the entry is still live, put it back in the queue so we can process it 
                // during the next cleanup cycle.
                _expiredHandlers.Enqueue(entry);
            }
        }

        Log.CleanupCycleEnd(_logger, stopwatch.GetElapsedTime(), disposedCount, _expiredHandlers.Count);
    }
    finally
    {
        Monitor.Exit(_cleanupActiveLock);
    }

    // We didn't totally empty the cleanup queue, try again later.
    if (_expiredHandlers.Count > 0)
    {
        StartCleanupTimer();
    }
}

上述方法核心是判断是否handler对象已经被GC，如果是的话，则释放其内部资源，即网络连接。

回到最初创建HttpClient的代码，会发现并没有传入任何name参数值。这是得益于HttpClientFactoryExtensions类的扩展方法。

public static HttpClient CreateClient(this IHttpClientFactory factory)
{
    if (factory == null)
    {
        throw new ArgumentNullException(nameof(factory));
    }

    return factory.CreateClient(Options.DefaultName);
}

Options.DefaultName的值为string.Empty。

DefaultHttpClientFactory缺少无参数的构造方法，唯一的构造方法需要传入多个参数，这也意味着构建它时需要依赖其它一些类，所以目前只适用于在ASP.NET程序中使用，还无法应用到诸如控制台一类的程序，希望之后官方能够对其继续增强，使得应用范围变得更广。

public DefaultHttpClientFactory(
    IServiceProvider services,
    ILoggerFactory loggerFactory,
    IOptionsMonitor<HttpClientFactoryOptions> optionsMonitor,
    IEnumerable<IHttpMessageHandlerBuilderFilter> filters)