MassTransit | 基于StateMachine实现Saga编排式分布式事务

Posted 2023-01-06 dotNET跨平台

什么是状态机

状态机作为一种程序开发范例，在实际的应用开发中有很多的应用场景，其中.NET 中的async/await 的核心底层实现就是基于状态机机制。状态机分为两种：有限状态机和无限状态机，本文介绍的就是有限状态机，有限状态机在任何时候都可以准确地处于有限状态中的一种，其可以根据一些输入从一个状态转换到另一个状态。一个有限状态机是由其状态列表、初始状态和触发每个转换的输入来定义的。如下图展示的就是一个闸机的状态机示意图：

从上图可以看出，状态机主要有以下核心概念：

1. State：状态，闸机有已开启（opened）和已关闭（closed）状态。

2. Transition：转移，即闸机从一个状态转移到另一个状态的过程。

3. Transition Condition：转移条件，也可理解为事件，即闸机在某一状态下只有触发了某个转移条件，才会执行状态转移。比如，闸机处于已关闭状态时，只有接收到开启事件才会执行转移动作，进而转移到开启状态。

4. Action：动作，即完成状态转移要执行的动作。比如要从关闭状态转移到开启状态，则需要执行开闸动作。

在.NET中，dotnet-state-machine/stateless和MassTransit都提供了开箱即用的状态机实现。本文将重点介绍MassTransit中的状态机在Saga 模式中的应用。

MassTransit StateMachine

在MassTransit 中MassTransitStateMachine就是状态机的具体抽象，可以用其编排一系列事件来实现状态的流转，也可以用来实现Saga模式的分布式事务。并支持与EF Core和Dapper集成将状态持久化到关系型数据库，也支持将状态持久化到MongoDB、Redis等数据库。是以简单的下单流程：创建订单->扣减库存->支付订单举例而言，其示意图如下所示。

基于状态机实现编排式Saga事务

那具体如何使用MassTransitStateMachine来应用编排式Saga 模式呢，接下来就来创建解决方案来实现以上下单流程示例。依次创建以下项目，除共享类库项目外，均安装MassTransit和MassTransit.RabbitMQNuGet包。

项目	项目名	项目类型
订单服务	MassTransit.SmDemo.OrderService	ASP.NET Core Web API
库存服务	MassTransit.SmDemo.InventoryService	Worker Service
支付服务	MassTransit.SmDemo.PaymentService	Worker Service
共享类库	MassTransit.SmDemo.Shared	Class Library

三个服务都添加扩展类MassTransitServiceExtensions，并在Program.cs类中调用services.AddMassTransitWithRabbitMq();注册服务。

using System.Reflection;
using MassTransit.CourierDemo.Shared.Models;

namespace MassTransit.CourierDemo.InventoryService;

public static class MassTransitServiceExtensions

    public static IServiceCollection AddMassTransitWithRabbitMq(this IServiceCollection services)
    
        return services.AddMassTransit(x =>
        
            x.SetKebabCaseEndpointNameFormatter();

            // By default, sagas are in-memory, but should be changed to a durable
            // saga repository.
            x.SetInMemorySagaRepositoryProvider();

            var entryAssembly = Assembly.GetEntryAssembly();
            x.AddConsumers(entryAssembly);
            x.AddSagaStateMachines(entryAssembly);
            x.AddSagas(entryAssembly);
            x.AddActivities(entryAssembly);
            x.UsingRabbitMq((context, busConfig) =>
            
                busConfig.Host(
                    host: "localhost",
                    port: 5672,
                    virtualHost: "masstransit",
                    configure: hostConfig =>
                    
                        hostConfig.Username("guest");
                        hostConfig.Password("guest");
                    );

                busConfig.ConfigureEndpoints(context);
            );
        );
    

订单服务

订单服务作为下单流程中的核心服务，主要职责包含接收创建订单请求和订单状态机的实现。先来定义OrderController如下：

namespace MassTransit.SmDemo.OrderService.Controllers;
[ApiController]
[Route("[controller]")]
public class OrderController : ControllerBase

    private readonly IBus _bus;
    public OrderController(IBus bus)
    
        _bus = bus;
    

    [HttpPost]
    public async Task<IActionResult> CreateOrder(CreateOrderDto createOrderDto)
    
        await _bus.Publish<ICreateOrderCommand>(new
        
            createOrderDto.CustomerId,
            createOrderDto.ShoppingCartItems
        );
        return Ok();
    

紧接着，订阅ICreateOrderCommand，执行订单创建逻辑，订单创建完毕后会发布ICreateOrderSucceed事件。

public class CreateOrderConsumer : IConsumer<ICreateOrderCommand>

    private readonly ILogger<CreateOrderConsumer> _logger;

    public CreateOrderConsumer(ILogger<CreateOrderConsumer> logger)
    
        _logger = logger;
    
    public async Task Consume(ConsumeContext<ICreateOrderCommand> context)
    
        var shoppingItems =
            context.Message.ShoppingCartItems.Select(item => new ShoppingCartItem(item.SkuId, item.Price, item.Qty));
        var order = new Order(context.Message.CustomerId).NewOrder(shoppingItems.ToArray());
        await OrderRepository.Insert(order);
        
        _logger.LogInformation($"Order order.OrderId created successfully");
        await context.Publish<ICreateOrderSucceed>(new
        
            order.OrderId,
            order.OrderItems
        );
    

最后来实现订单状态机，主要包含以下几步：

1. 定义状态机状态：一个状态机从启动到结束可能会经历各种异常，包括程序异常或物理故障，为确保状态机能从异常中恢复，因此必须保存状态机的状态。本例中，定义OrderState以保存状态机实例状态数据：

using MassTransit.SmDemo.OrderService.Domains;

namespace MassTransit.SmDemo.OrderService;

public class OrderState : SagaStateMachineInstance

    public Guid CorrelationId  get; set; 
    public string CurrentState  get; set; 
    public Guid OrderId  get; set; 
    public decimal Amount  get; set; 
    public List<OrderItem> OrderItems  get; set; 

1. 定义状态机：直接继承自MassTransitStateMachine并同时指定状态实例即可：

namespace MassTransit.SmDemo.OrderService;

public class OrderStateMachine : MassTransitStateMachine<OrderState>

1. 注册状态机：这里指定内存持久化方式来持久化状态，也可指定诸如MongoDb、mysql等数据库进行状态持久化：

return services.AddMassTransit(x =>

    //...
    x.AddSagaStateMachine<OrderStateMachine, OrderState>()
        .InMemoryRepository();

1. 定义状态列表：即状态机涉及到的系列状态，并通过State类型定义，本例中为：

   1. 已创建：public State Created get; private set;

   2. 库存已扣减：public State InventoryDeducted get; private set;

   3. 已支付：public State Paid get; private set;

   4. 已取消：public State Canceled get; private set;

2. 定义转移条件：即推动状态流转的事件，通过Event<T>类型定义，本例涉及有：

   1. 订单成功创建事件：public Event<ICreateOrderSucceed> OrderCreated get; private set;

   2. 库存扣减成功事件：public Event<IDeduceInventorySucceed> DeduceInventorySucceed get; private set;

   3. 库存扣减失败事件：public Event<IDeduceInventoryFailed> DeduceInventoryFailed get; private set;

   4. 订单支付成功事件：public Event<IPayOrderSucceed> PayOrderSucceed get; private set;

   5. 订单支付失败事件：public Event<IPayOrderFailed> PayOrderFailed get; private set;

   6. 库存已返还事件：public Event<IReturnInventorySucceed> ReturnInventorySucceed get; private set;

   7. 订单取消事件：public Event<ICancelOrderSucceed> OrderCanceled get; private set;

3. 定义关联关系：由于每个事件都是孤立的，但相关联的事件终会作用到某个具体的状态机实例上，如何关联事件以推动状态机的转移呢？配置关联Id。以下就是将事件消息中的传递的OrderId作为关联ID。

   1. Event(() => OrderCreated, x => x.CorrelateById(m => m.Message.OrderId));

   2. Event(() => DeduceInventorySucceed, x => x.CorrelateById(m => m.Message.OrderId));

   3. Event(() => DeduceInventoryFailed, x => x.CorrelateById(m => m.Message.OrderId));

   4. Event(() => PayOrderSucceed, x => x.CorrelateById(m => m.Message.OrderId));

4. 定义状态转移：即状态在什么条件下做怎样的动作完成状态的转移，本例中涉及的正向状态转移有：

   (1) 初始状态->已创建：触发条件为OrderCreated事件，同时要发送IDeduceInventoryCommand推动库存服务执行库存扣减。

Initially(
    When(OrderCreated)
        .Then(context =>
        
            context.Saga.OrderId = context.Message.OrderId;
            context.Saga.OrderItems = context.Message.OrderItems;
            context.Saga.Amount = context.Message.OrderItems.Sum(x => x.Price * x.Qty);
        )
        .PublishAsync(context => context.Init<IDeduceInventoryCommand>(new
        
            context.Saga.OrderId,
            DeduceInventoryItems =
                context.Saga.OrderItems.Select(x => new DeduceInventoryItem(x.SkuId, x.Qty)).ToList()
        ))
        .TransitionTo(Created));

(2) 已创建-> 库存已扣减：触发条件为DeduceInventorySucceed事件，同时要发送IPayOrderCommand推动支付服务执行订单支付。

During(Created,
    When(DeduceInventorySucceed)
        .Then(context =>
        
            context.Publish<IPayOrderCommand>(new
            
                context.Saga.OrderId,
                context.Saga.Amount
            );
        ).TransitionTo(InventoryDeducted),
    When(DeduceInventoryFailed).Then(context =>
    
        context.Publish<ICancelOrderCommand>(new
        
            context.Saga.OrderId
        );
    )
);

(3) 库存已扣减->已支付：触发条件为PayOrderSucceed事件，转移到已支付后，流程结束。

During(InventoryDeducted,
    When(PayOrderFailed).Then(context =>
    
        context.Publish<IReturnInventoryCommand>(new
        
            context.Message.OrderId,
            ReturnInventoryItems =
                context.Saga.OrderItems.Select(x => new ReturnInventoryItem(x.SkuId, x.Qty)).ToList()
        );
    ),
    When(PayOrderSucceed).TransitionTo(Paid).Then(context => context.SetCompleted()));

最终完整版的OrderStateMachine如下所示：

using MassTransit.SmDemo.OrderService.Events;
using MassTransit.SmDemo.Shared.Contracts;

namespace MassTransit.SmDemo.OrderService;

public class OrderStateMachine : MassTransitStateMachine<OrderState>

    public State Created  get; private set; 
    public State InventoryDeducted  get; private set; 
    public State Paid  get; private set; 
    public State Canceled  get; private set; 

    public Event<ICreateOrderSucceed> OrderCreated  get; private set; 
    public Event<IDeduceInventorySucceed> DeduceInventorySucceed  get; private set; 
    public Event<IDeduceInventoryFailed> DeduceInventoryFailed  get; private set; 
    public Event<ICancelOrderSucceed> OrderCanceled  get; private set; 
    public Event<IPayOrderSucceed> PayOrderSucceed  get; private set; 
    public Event<IPayOrderFailed> PayOrderFailed  get; private set; 
    public Event<IReturnInventorySucceed> ReturnInventorySucceed  get; private set; 
    public Event<IOrderStateRequest> OrderStateRequested  get; private set; 
    
 public OrderStateMachine()
    
        Event(() => OrderCreated, x => x.CorrelateById(m => m.Message.OrderId));
        Event(() => DeduceInventorySucceed, x => x.CorrelateById(m => m.Message.OrderId));
        Event(() => DeduceInventoryFailed, x => x.CorrelateById(m => m.Message.OrderId));
        Event(() => ReturnInventorySucceed, x => x.CorrelateById(m => m.Message.OrderId));
        Event(() => PayOrderSucceed, x => x.CorrelateById(m => m.Message.OrderId));
        Event(() => PayOrderFailed, x => x.CorrelateById(m => m.Message.OrderId));
        Event(() => OrderCanceled, x => x.CorrelateById(m => m.Message.OrderId));
        Event(() => OrderStateRequested, x =>
        
            x.CorrelateById(m => m.Message.OrderId);
            x.OnMissingInstance(m =>
            
                return m.ExecuteAsync(x => x.RespondAsync<IOrderNotFoundOrCompleted>(new  x.Message.OrderId ));
            );
        );

        InstanceState(x => x.CurrentState);

        Initially(
            When(OrderCreated)
                .Then(context =>
                
                    context.Saga.OrderId = context.Message.OrderId;
                    context.Saga.OrderItems = context.Message.OrderItems;
     var amount = context.Message.OrderItems.Sum(x => x.Price * x.Qty);
                    context.Saga.Amount = amount;
                )
                .PublishAsync(context => context.Init<IDeduceInventoryCommand>(new
                
                    context.Saga.OrderId,
                    DeduceInventoryItems =
                        context.Saga.OrderItems.Select(x => new DeduceInventoryItem(x.SkuId, x.Qty)).ToList()
                ))
                .TransitionTo(Created));

        During(Created,
            When(DeduceInventorySucceed)
                .Then(context =>
                
                    context.Publish<IPayOrderCommand>(new
                    
                        context.Saga.OrderId,
                        context.Saga.Amount
                    );
                ).TransitionTo(InventoryDeducted),
            When(DeduceInventoryFailed).Then(context =>
            
                context.Publish<ICancelOrderCommand>(new
                
                    context.Saga.OrderId
                );
            )
        );

        During(InventoryDeducted,
            When(PayOrderFailed).Then(context =>
            
                context.Publish<IReturnInventoryCommand>(new
                
                    context.Message.OrderId,
                    ReturnInventoryItems =
                        context.Saga.OrderItems.Select(x => new ReturnInventoryItem(x.SkuId, x.Qty)).ToList()
                );
            ),
            When(PayOrderSucceed).TransitionTo(Paid).Then(context => context.SetCompleted()),
            When(ReturnInventorySucceed)
                .ThenAsync(context => context.Publish<ICancelOrderCommand>(new
                
                    context.Saga.OrderId
                )).TransitionTo(Created));

        DuringAny(When(OrderCanceled).TransitionTo(Canceled).ThenAsync(async context =>
        
            await Task.Delay(TimeSpan.FromSeconds(10));
            await context.SetCompleted();
        ));


        DuringAny(
            When(OrderStateRequested)
                .RespondAsync(x => x.Init<IOrderStateResponse>(new
                
                    x.Saga.OrderId,
                    State = x.Saga.CurrentState
                ))
        );
    

库存服务

库存服务在整个下单流程的职责主要是库存的扣减和返还，其仅需要订阅IDeduceInventoryCommand和IReturnInventoryCommand两个命令并实现即可。代码如下所示：

using MassTransit.SmDemo.InventoryService.Repositories;
using MassTransit.SmDemo.Shared.Contracts;

namespace MassTransit.SmDemo.InventoryService.Consumers;

public class DeduceInventoryConsumer : IConsumer<IDeduceInventoryCommand>

    private readonly ILogger<DeduceInventoryConsumer> _logger;

    public DeduceInventoryConsumer(ILogger<DeduceInventoryConsumer> logger)
    
        _logger = logger;
    

    public async Task Consume(ConsumeContext<IDeduceInventoryCommand> context)
    
        if (!CheckStock(context.Message.DeduceInventoryItems))
        
            _logger.LogWarning($"Insufficient stock for order [context.Message.OrderId]!");
            await context.Publish<IDeduceInventoryFailed>(
                new  context.Message.OrderId, Reason = "insufficient stock" );
        
        else
        
            _logger.LogInformation($"Inventory has been deducted for order [context.Message.OrderId]!");
            DeduceStocks(context.Message.DeduceInventoryItems);
            await context.Publish<IDeduceInventorySucceed>(new  context.Message.OrderId );
        
    


    private bool CheckStock(List<DeduceInventoryItem> deduceItems)
    
        foreach (var stockItem in deduceItems)
        
            if (InventoryRepository.GetStock(stockItem.SkuId) < stockItem.Qty) return false;
        

        return true;
    

    private void DeduceStocks(List<DeduceInventoryItem> deduceItems)
    
        foreach (var stockItem in deduceItems)
        
            InventoryRepository.TryDeduceStock(stockItem.SkuId, stockItem.Qty);
        
    

namespace MassTransit.SmDemo.InventoryService.Consumers;

public class ReturnInventoryConsumer : IConsumer<IReturnInventoryCommand>

    private readonly ILogger<ReturnInventoryConsumer> _logger;

    public ReturnInventoryConsumer(ILogger<ReturnInventoryConsumer> logger)
    
        _logger = logger;
    

    public async Task Consume(ConsumeContext<IReturnInventoryCommand> context)
    
        foreach (var returnInventoryItem in context.Message.ReturnInventoryItems)
        
            InventoryRepository.ReturnStock(returnInventoryItem.SkuId, returnInventoryItem.Qty);
        

        _logger.LogInformation($"Inventory has been returned for order [context.Message.OrderId]!");
        await context.Publish<IReturnInventorySucceed>(new  context.Message.OrderId );
    

支付服务

对于下单流程的支付用例来说，要么成功要么失败，因此仅需要订阅IPayOrderCommand命令即可，具体PayOrderConsumer实现如下：

using MassTransit.SmDemo.Shared.Contracts;

namespace MassTransit.SmDemo.PaymentService.Consumers;

public class PayOrderConsumer : IConsumer<IPayOrderCommand>

    private readonly ILogger<PayOrderConsumer> _logger;

    public PayOrderConsumer(ILogger<PayOrderConsumer> logger)
    
        _logger = logger;
    
    public async Task Consume(ConsumeContext<IPayOrderCommand> context)
    
        await Task.Delay(TimeSpan.FromSeconds(10));
        if (context.Message.Amount % 2 == 0)
        _logger.LogInformation($"Order [context.Message.OrderId] paid successfully!");
            await context.Publish<IPayOrderSucceed>(new  context.Message.OrderId );
        
        else
        
            _logger.LogWarning($"Order [context.Message.OrderId] payment failed!");
            await context.Publish<IPayOrderFailed>(new
            
                context.Message.OrderId,
                Reason = "Insufficient account balance"
            );
        
    

运行结果

启动三个项目，并在Swagger中发起订单创建请求，如下图所示：

由于订单总额为奇数，因此支付会失败，最终控制台输出如下图所示：

打开RabbitMQ后台，可以看见MassTransit按照约定创建了以下队列用于服务间的消息传递：

其中order-state队列绑定到类型为fanout的同名order-stateExchange，其绑定关系如下图所示，该Exchange负责从其他同名事件的Exchange转发事件。

总结

通过以上示例的讲解，相信了解到MassTransit StateMachine的强大之处。StateMachine充当着事务编排器的角色，通过集中定义状态、转移条件和状态转移的执行顺序，实现高内聚的事务流转控制，也确保了其他伴生服务仅需关注自己的业务逻辑，而无需关心事务的流转，真正实现了关注点分离。