以管道的方式来完成复杂的流程处理
之前参与一个机票价格计算的项目,为他们设计了基本的处理流程,但是由于整个计算流程相当复杂,而且变化非常频繁,导致日常的修改、维护和升级也变得越来越麻烦,当我后来再接手的时候已经看不懂计算逻辑了。为了解决这个问题,我借鉴了“工作流”的思路,试图将整个计算过程设计成一个工作流。但是我又不想引入一个独立的工作流引擎,于是写了一个名为Pipelines的框架。顾名思义,Pipelines通过构建Pipeline的方式完成所需的处理流程,整个处理逻辑被分解并实现在若干Pipe中,这些Pipe按照指定的顺序将完成的Pipeline构建出来。Pipeline本质上就是一个简单的顺序工作流,它仅仅按序执行注册的Pipe。这个简单的Pipelines框架被放在这里,这里我不会介绍它的设计实现,只是简单地介绍它的用法,有兴趣的可以查看源代码。
一、构建并执行管道
二、Pipeline的“内部中断”
三、Pipeline的“外部中断”
四、处理层次化数据结构
五、利用扩展方法使Pipeline构建更简洁
一、构建并执行管道
Pipelines旨在提供一个用于处理数据的顺序工作流或者管道(以下简称Pipeline),该Pipeline在一个强类型的上下文中被执行,管道可以利用此上下文得到需要处理的数据,并将处理的结果(含中间结果)存储在上下文中。接下来我们来演示如何利用Pipelines框架处理人口统计数据的实例。如下所示的两个类型分别表示人口统计数据和处理上下文,后者继承基类ContextBase。
public class PopulationData { public object Statistics { get; set; } = default!; } public sealed class PopulationContext : ContextBase { public PopulationContext(PopulationData data)=> Data = data; public PopulationData Data { get; } }
Pipeline由一系列Pipe对象按照注册的顺序组合而成。通过继承基类PipeBase<PopulationContext>,我们定义了三个Pipe类来完成针对人口统计数据的三项基本处理任务。
public sealed class FooPopulationPipe : PipeBase<PopulationContext> { public override string Description => "Global PopulationProcessor Foo"; protected override void Invoke(PopulationContext context) =>Console.WriteLine($"{nameof(FooPopulationPipe)} is invoked."); } public sealed class BarPopulationPipe : PipeBase<PopulationContext> { public override string Description => "Global PopulationProcessor Bar"; protected override void Invoke(PopulationContext context) => Console.WriteLine($"{nameof(BarPopulationPipe)} is invoked."); } public sealed class BazPopulationPipe : PipeBase<PopulationContext> { public override string Description => "Global PopulationProcessor Baz"; protected override void Invoke(PopulationContext context) => Console.WriteLine($"{nameof(BazPopulationPipe)} is invoked."); }
我设计Pipelines的初衷是让每个参与者(包含非技术人员)在代码的频繁迭代过程中,可以清晰地了解当前的处理流程,所以我会将当前应用构建的所有Pipeline的处理流程导出来。基于这个目的,每个Pipe类型都需要利用其Description属性提供一段描述当前处理逻辑的文本。Pipe具体的处理逻辑实现在重写的Invoke方法中。如果涉及异步处理,需要继承更上层的基类Pipe<TContext>(PipeBase<TContext>的基类)并重写异步的InvokeAsync方法。
Pipeline的构建实现在如下所示的BuildPipelines方法中,我们利用该方法提供的IPipelineProvider对象注册了一个命名为“PopulationProcessor”的Pipeline。具体来说,我们调用的是它的AddPipeline<TContext>方法,该方法提供的第一个参数为Pipeline的注册名称,另一个参数是一个类型为Action<IPipelineBuilder<TContext>>的委托,它利用提供的IPipelineBuilder<TContext>对象完成了上面定义的三个Pipe的注册。
using App; using Artech.Pipelines; var builder = WebApplication.CreateBuilder(args); builder.Services.AddPipelines(BuildPipelines); var app = builder.Build(); app.MapGet("/test", async (IPipelineProvider provider, HttpResponse response) => { Console.WriteLine("Execute PopulationProcessor pipeline"); var context = new PopulationContext(new PopulationData()); var pipeline = provider.GetPipeline<PopulationContext>("PopulationProcessor"); await pipeline.ProcessAsync(context); return Results.Ok(); }); app.Run(); static void BuildPipelines(IPipelineProvider pipelineProvider) { pipelineProvider.AddPipeline<PopulationContext>( name: "PopulationProcessor", setup: builder => builder .Use<PopulationContext, FooPopulationPipe>() .Use<PopulationContext, BarPopulationPipe>() .Use<PopulationContext, BazPopulationPipe>()); }
Pipelines框架涉及的服务通过IServiceCollection接口的AddPipelines方法进行注册,BuildPipelines方法转换成委托作为该方法的参数。我们注册了一个指向“/test” 的路由终结点来演示针对管道的执行。如代码片段所示,我们利用注入的IPipelineProvider对象根据注册名称得到具体的Pipeline对象,并创建出相应的PopulationContext上下文作为参数来执行此Pipeline对象。程序执行后,请求路径”/pipelines”可以得到一个Pipeline的列表,点击具体的链接,对应Pipeline体现的流程就会呈现出来。
如果请求路径“/test”来执行构建的管道,管道执行的轨迹将会体现在控制台的输出结果上。
二、Pipeline的“内部中断”
构成Pipeline的每个Pipe都可以根据处理逻辑的需要立即中断管道的执行。在如下这个重写的BarPopulationPipe类型的Invoke方法中,如果生成的随机数为偶数,它会调用上下文对象的Abort方法立即终止Pipeline的执行。
public sealed class BarPopulationPipe : PipeBase<PopulationContext> { private readonly Random _random = new(); public override string Description => "Global PopulationProcessor Bar"; protected override void Invoke(PopulationContext context) { Console.WriteLine($"{nameof(BarPopulationPipe)} is invoked."); if (_random.Next() % 2 == 0) { context.Abort(); } } }
这样的化,当我们构建的Pipeline在执行过程中,有一半的几率BazPopulationPipe将不会执行,如下所示的输出结果体现了这一点。
对于继承自Pipe<TContext>的Pipe类型,其实现的InvokeAsync方法可以采用如下的方式中止当前Pipeline的执行,因为参数next返回的委托用于调用后续Pipe。如果不执行此委托,就意味着针对Pipeline的执行到此为止。
public sealed class BarPopulationPipe : Pipe<PopulationContext> { private readonly Random _random = new(); public override string Description => "Global PopulationProcessor Bar"; public override ValueTask InvokeAsync(PopulationContext context, Func<PopulationContext, ValueTask> next) { Console.WriteLine($"{nameof(BarPopulationPipe)} is invoked."); if (_random.Next() % 2 != 0) { return next(context); } return ValueTask.CompletedTask; } }
三、Pipeline的“外部中断”
在调用Pipeline时,我们可以利用执行上下文提供的CancellationToken中止Pipeline的执行。我们按照如下的方式再次改写了BarPopulationPipe的执行逻辑,如下面的代码片段所示,我们不再调用Abort方法,而是选择延迟2秒执行后续操作。
public sealed class BarPopulationPipe : Pipe<PopulationContext> { private readonly Random _random = new(); public override string Description => "Global PopulationProcessor Bar"; public override async ValueTask InvokeAsync(PopulationContext context, Func<PopulationContext, ValueTask> next) { Console.WriteLine($"{nameof(BarPopulationPipe)} is invoked."); if (_random.Next() % 2 != 0) { await Task.Delay(2000); } await next(context); } }
我们按照如下的方式重写了PopulationContext的CancellationToken属性。我们为构造函数添加了两个参数,一个代表当前HttpContext上下文,另一个表示设置的超时时限。CancellationToken根据这两个参数创建而成,意味着管道不仅具有默认的超时时间,也可以通过HTTP调用方中止执行。
public sealed class PopulationContext: ContextBase { public PopulationContext(PopulationData data, HttpContext httpContext, TimeSpan timeout) { Data = data; CancellationToken = CancellationTokenSource.CreateLinkedTokenSource(httpContext.RequestAborted, new CancellationTokenSource(timeout).Token).Token; } public PopulationData Data { get; } public override CancellationToken CancellationToken { get; } }
在注册的终结点处理器中,我们在执行Pipeline之前,将作为参数传入的PopulationContext上下文的超时时间设置为1秒。
var builder = WebApplication.CreateBuilder(args); builder.Services.AddPipelines(BuildPipelines); var app = builder.Build(); app.MapGet("/test", async (HttpContext httpContext,IPipelineProvider provider, HttpResponse response) => { Console.WriteLine("Execute PopulationProcessor pipeline"); var context = new PopulationContext(new PopulationData(), httpContext, TimeSpan.FromSeconds(1)); var pipeline = provider.GetPipeline<PopulationContext>("PopulationProcessor"); await pipeline.ProcessAsync(context); return Results.Ok(); }); app.Run();
根据BarPopulationPipe的执行逻辑,Pipeline的执行具有一半的几率会超时,一旦超时就会立即抛出一个OperationCancellationToken异常。
四、处理层次化数据结构
Pipelines设计的主要目的是用来处理层次化的数据结构,这涉及到子Pipeline的应用。目前我们处理的人口数据体现为一个简单的数据类型,现在我们让它变得更复杂一些。假设我们需要处理国家、省份和城市三个等级的人口数据,其中StatePopulationData代表整个国家的人口数据,它的Provinces属性承载了每个省份的数据。ProvincePopulationData代表具体某个省份的人口数据,其Cities属性承载了每个城市的人口数据。
public class PopulationData { public object Statistics { get; set; } = default!; } public class StatePopulationData { public IDictionary<string, ProvincePopulationData> Provinces { get; set; } = default!; } public class ProvincePopulationData { public IDictionary<string, PopulationData> Cities { get; set; } = default!; }
现在我们需要构建一个Pipeline来处理通过StatePopulationData类型表示的整个国家的人口数据,具体的处理流程为:
- 利用FooStatePipe处理国家人口数据
- 利用BarStatePipe处理国家人口数据
- 构建子Pipeline处理每个省份人口数据,子Pipeline处理逻辑:
- 利用FooProvincePipe处理省份人口数据
- 利用BarProvincePipe处理省份人口数据、
- 构建子Pipeline处理每个城市人口数据,子Pipeline处理逻辑
- 利用FooCityPipe处理城市人口数据
- 利用BarCityPipe处理城市人口数据
- 利用BazCityPipe处理城市人口数据
- 利用BazProvincePipe处理省份人口数据
- 利用BazStatePipe处理国家人口数据
为此我们需要定义9个Pipe类型,以及3个执行上下文。如下所示的是三个执行上下文类型的具体定义:
public sealed class StatePopulationContext: ContextBase { public StatePopulationData PopulationData { get; } public StatePopulationContext(StatePopulationData populationData) => PopulationData = populationData; } public sealed class ProvincePopulationContext : SubContextBase<StatePopulationContext, KeyValuePair<string, ProvincePopulationData>> { public string Province { get; private set; } = default!; public IDictionary<string, PopulationData> Cities { get; private set; } = default!; public override void Initialize(StatePopulationContext parent, KeyValuePair<string, ProvincePopulationData> item) { Province = item.Key; Cities = item.Value.Cities; base.Initialize(parent, item); } } public sealed class CityPopulationContext: SubContextBase<ProvincePopulationContext, KeyValuePair<string, PopulationData>> { public string City { get; private set; } = default!; public PopulationData PopulationData { get; private set; } = default!; public override void Initialize(ProvincePopulationContext parent, KeyValuePair<string, PopulationData> item) { City = item.Key; PopulationData = item.Value; base.Initialize(parent, item); } }
9个对应的Pipe类型定义如下。每个类型利用重写的Description提供一个简单的描述,重写的Invoke方法输出当前怎样的数据(那个省/市的人口数据)。
public sealed class FooStatePipe : PipeBase<StatePopulationContext> { public override string Description => "State Population Processor Foo"; protected override void Invoke(StatePopulationContext context)=>Console.WriteLine("Foo: Process state population"); } public sealed class BarStatePipe : PipeBase<StatePopulationContext> { public override string Description => "State Population Processor Bar"; protected override void Invoke(StatePopulationContext context) => Console.WriteLine("Bar: Process state population"); } public sealed class BazStatePipe : PipeBase<StatePopulationContext> { public override string Description => "State Population Processor Baz"; protected override void Invoke(StatePopulationContext context) => Console.WriteLine("Baz: Process state population"); } public sealed class FooProvincePipe : PipeBase<ProvincePopulationContext> { public override string Description => "Province Population Processor Foo"; protected override void Invoke(ProvincePopulationContext context) => Console.WriteLine($"\tFoo: Process population of the province {context.Province}"); } public sealed class BarProvincePipe : PipeBase<ProvincePopulationContext> { public override string Description => "Province Population Processor Bar"; protected override void Invoke(ProvincePopulationContext context) => Console.WriteLine($"\tBar: Process population of the province {context.Province}"); } public sealed class BazProvincePipe : PipeBase<ProvincePopulationContext> { public override string Description => "Province Population Processor Baz"; protected override void Invoke(ProvincePopulationContext context) => Console.WriteLine($"\tBaz: Process population of the province {context.Province}"); } public sealed class FooCityPipe : PipeBase<CityPopulationContext> { public override string Description => "City Population Processor Foo"; protected override void Invoke(CityPopulationContext context) => Console.WriteLine($"\t\tFoo: Process population of the city {context.City}"); } public sealed class BarCityPipe : PipeBase<CityPopulationContext> { public override string Description => "City Population Processor Bar"; protected override void Invoke(CityPopulationContext context) => Console.WriteLine($"\t\tBar: Process population of the city {context.City}"); } public sealed class BazCityPipe : PipeBase<CityPopulationContext> { public override string Description => "City Population Processor Baz"; protected override void Invoke(CityPopulationContext context) => Console.WriteLine($"\t\tBaz: Process population of the city {context.City}"); }
static void BuildPipelines(IPipelineProvider pipelineProvider) { pipelineProvider.AddPipeline<StatePopulationContext>(name: "PopulationProcessor", setup: builder => builder .Use<StatePopulationContext, FooStatePipe>() .Use<StatePopulationContext, BarStatePipe>() .ForEach<StatePopulationContext, ProvincePopulationContext, KeyValuePair<string, ProvincePopulationData>>( description: "For each province", collectionAccessor: context => context.PopulationData.Provinces, subPipelineSetup: provinceBuilder => provinceBuilder .Use<ProvincePopulationContext, FooProvincePipe>() .Use<ProvincePopulationContext, BarProvincePipe>() .ForEach<ProvincePopulationContext, CityPopulationContext, KeyValuePair<string, PopulationData>>( description: "For each city", collectionAccessor: context => context.Cities, subPipelineSetup: cityBuilder => cityBuilder .Use<CityPopulationContext, FooCityPipe>() .Use<CityPopulationContext, BarCityPipe>() .Use<CityPopulationContext, BazCityPipe>()) .Use<ProvincePopulationContext, BazProvincePipe>()) .Use<StatePopulationContext, BazStatePipe>()); }
终结点处理程序在执行新的Pipeline时,会按照如下的形式将StatePopulationContext上下文构建出来。处理人口数据涉及三个省份(江苏、山东和浙江),每个省份包含三个城市的人口数据。
var builder = WebApplication.CreateBuilder(args); builder.Services.AddPipelines(BuildPipelines); var app = builder.Build(); app.MapGet("/test", async (HttpContext httpContext, IPipelineProvider provider, HttpResponse response) => { Console.WriteLine("Execute PopulationProcessor pipeline"); var data = new StatePopulationData { Provinces = new Dictionary<string, ProvincePopulationData>() }; data.Provinces.Add("Jiangsu", new ProvincePopulationData { Cities = new Dictionary<string, PopulationData> { {"Suzhou", new PopulationData() }, {"Wuxi", new PopulationData() }, {"Changezhou", new PopulationData() }, } }); data.Provinces.Add("Shandong", new ProvincePopulationData { Cities = new Dictionary<string, PopulationData> { {"Qingdao", new PopulationData() }, {"Jinan", new PopulationData() }, {"Yantai", new PopulationData() }, } }); data.Provinces.Add("Zhejiang", new ProvincePopulationData { Cities = new Dictionary<string, PopulationData> { {"Hangzhou", new PopulationData() }, {"Ningbo", new PopulationData() }, {"Wenzhou", new PopulationData() }, } }); var context = new StatePopulationContext(data); var pipeline = provider.GetPipeline<StatePopulationContext>("PopulationProcessor"); await pipeline.ProcessAsync(context); return Results.Ok(); }); app.Run();
应用启动后,我们依然可以从Pipeline导出页面看到整个Pipeline的处理流程。
当我们请求“/test”,Pipeline针对国家人口数据的执行流程体现在控制台输出上。
五、利用扩展方法使Pipeline构建更简洁
Pipeline的构建过程体现了完整的处理流程,所以我们应该构建代码尽可能地简洁,最理想的状态就是让非技术人员也能看懂。Pipelines提供的用于注册Pipe的API均为泛型方法,并且会涉及两到三个必须显式指定的泛型参数,使用起来还不是很方便。不过这个问题可以通过自定义扩展方法来解决。
public static class Extensions { public static IPipelineBuilder<StatePopulationContext> UseStatePipe<TPipe>(this IPipelineBuilder<StatePopulationContext> builder) where TPipe : Pipe<StatePopulationContext> => builder.Use<StatePopulationContext, TPipe>(); public static IPipelineBuilder<ProvincePopulationContext> UseProvincePipe<TPipe>(this IPipelineBuilder<ProvincePopulationContext> builder) where TPipe : Pipe<ProvincePopulationContext> => builder.Use<ProvincePopulationContext, TPipe>(); public static IPipelineBuilder<CityPopulationContext> UseCityPipe<TPipe>(this IPipelineBuilder<CityPopulationContext> builder) where TPipe : Pipe<CityPopulationContext> => builder.Use<CityPopulationContext, TPipe>(); public static IPipelineBuilder<StatePopulationContext> ForEachProvince(this IPipelineBuilder<StatePopulationContext> builder, Action<IPipelineBuilder<ProvincePopulationContext>> setup) => builder.ForEach("For each province", it => it.PopulationData.Provinces, (_, _) => true, setup); public static IPipelineBuilder<ProvincePopulationContext> ForEachCity(this IPipelineBuilder<ProvincePopulationContext> builder, Action<IPipelineBuilder<CityPopulationContext>> setup) => builder.ForEach("For each city", it => it.Cities, (_, _) => true, setup); }
如上面的代码片段所示,我们针对三个数据层次(国家、省份、城市)定义了注册对应Pipe的扩展方法UseStatePipe、UseProvincePipe和UseCityPipe。还分别定义了ForEachProvince和ForEachCity这两个扩展方法来注册构建处理省份/城市人口数据的子Pipeline。有了这5个扩展方法,构建整个Pipeline的代码就可以变得非常简单而清晰,即使不写任何的注释,相信每个人(包括非开发人员)都能读懂涉及的处理流程。
static void BuildPipelines(IPipelineProvider pipelineProvider) { pipelineProvider.AddPipeline<StatePopulationContext>(name: "PopulationProcessor", setup: builder => builder .UseStatePipe<FooStatePipe>() .UseStatePipe<BarStatePipe>() .ForEachProvince(provinceBuilder => provinceBuilder .UseProvincePipe<FooProvincePipe>() .UseProvincePipe<BarProvincePipe>() .ForEachCity(cityBuilder => cityBuilder .UseCityPipe<FooCityPipe>() .UseCityPipe<BarCityPipe>() .UseCityPipe<BazCityPipe>()) .UseProvincePipe<BazProvincePipe>()) .UseStatePipe<BazStatePipe>()); }