真是个好问题。我也在探索的路上，互联网上的大多数答案带来的问题似乎和它们带来的解决方案一样多。

所以(冒着写一些我一年后不同意的东西的风险)，这里是我迄今为止的发现。

首先，我们喜欢丰富的领域模型，它为我们提供了高可发现性(我们可以用聚合做什么)和可读性(表达方法调用)。

// Entity
public class Invoice
{
    ...
    public void SetStatus(StatusCode statusCode, DateTime dateTime) { ... }
    public void CreateCreditNote(decimal amount) { ... }
    ...
}

我们希望在不向实体的构造函数中注入任何服务的情况下实现这一点，因为:

Introduction of a new behavior (that uses a new service) could lead to a constructor change, meaning the change affects every line that instantiates the entity! These services are not part of the model, but constructor-injection would suggest that they were. Often a service (even its interface) is an implementation detail rather than part of the domain. The domain model would have an outward-facing dependency. It can be confusing why the entity cannot exist without these dependencies. (A credit note service, you say? I am not even going to do anything with credit notes...) It would make it hard instantiate, thus hard to test. The problem spreads easily, because other entities containing this one would get the same dependencies - which on them may look like very unnatural dependencies.

那么，我们该怎么做呢?到目前为止，我的结论是方法依赖和双重分派提供了一个不错的解决方案。

public class Invoice
{
    ...

    // Simple method injection
    public void SetStatus(IInvoiceLogger logger, StatusCode statusCode, DateTime dateTime)
    { ... }

    // Double dispatch
    public void CreateCreditNote(ICreditNoteService creditNoteService, decimal amount)
    {
        creditNoteService.CreateCreditNote(this, amount);
    }

    ...
}

CreateCreditNote()现在需要一个负责创建信用票据的服务。它使用双重分派，将工作完全卸载到负责的服务，同时保持来自Invoice实体的可发现性。

SetStatus()现在对记录器有一个简单的依赖，显然记录器将执行部分工作。

对于后者，为了使客户端代码更简单，我们可以通过IInvoiceService进行日志记录。毕竟，发票日志似乎是发票的固有特性。这样一个单独的IInvoiceService有助于避免对各种操作的各种迷你服务的需求。缺点是，它变得模糊，该服务究竟将做什么。它甚至可能开始看起来像双重分派，而大部分工作实际上仍然在SetStatus()本身中完成。

我们仍然可以将参数命名为“logger”，希望能够揭示我们的意图。不过看起来有点弱。

相反，我将选择请求一个iinvoiceogger(正如我们在代码示例中已经做的那样)，并让IInvoiceService实现该接口。客户端代码可以简单地将它的单个IInvoiceService用于所有要求任何此类非常特殊的、发票固有的“迷你服务”的Invoice方法，而方法签名仍然非常清楚地表明它们要求的是什么。

我注意到我没有明确地谈到存储库。记录器是或使用存储库，但让我还提供一个更明确的示例。如果只在一两个方法中需要存储库，我们可以使用相同的方法。

public class Invoice
{
    public IEnumerable<CreditNote> GetCreditNotes(ICreditNoteRepository repository)
    { ... }
}

事实上，这为麻烦的惰性加载提供了另一种选择。

更新:出于历史原因，我把下面的文字留下来了，但我建议100%地避开惰性加载。

对于真正的、基于属性的惰性加载，我目前确实使用构造函数注入，但以一种不考虑持久性的方式。

public class Invoice
{
    // Lazy could use an interface (for contravariance if nothing else), but I digress
    public Lazy<IEnumerable<CreditNote>> CreditNotes { get; }

    // Give me something that will provide my credit notes
    public Invoice(Func<Invoice, IEnumerable<CreditNote>> lazyCreditNotes)
    {
        this.CreditNotes = new Lazy<IEnumerable<CreditNotes>>() => lazyCreditNotes(this));
    }
}

一方面，从数据库加载Invoice的存储库可以自由访问加载相应信用票据的函数，并将该函数注入到Invoice中。

另一方面，创建实际新Invoice的代码只会传递一个返回空列表的函数:

new Invoice(inv => new List<CreditNote>() as IEnumerable<CreditNote>)

(一个自定义的ILazy<out T>可以使我们摆脱难看的转换到IEnumerable，但这会使讨论复杂化。)

// Or just an empty IEnumerable
new Invoice(inv => IEnumerable.Empty<CreditNote>())

我很高兴听到你的意见，喜好和改进!

为什么要分离数据访问?

从书中，我认为模型驱动设计一章的前两页给出了一些理由，说明为什么要从领域模型的实现中抽象出技术实现细节。

您希望在领域模型和代码之间保持紧密的连接 分离技术关注点有助于证明模型对于实现是可行的 你希望无处不在的语言渗透到系统的设计中

这似乎都是为了避免独立的“分析模型”与系统的实际实现分离。

根据我对这本书的理解，它说这个“分析模型”可以在不考虑软件实现的情况下被设计出来。一旦开发人员试图实现业务方面理解的模型，他们就会由于需要而形成自己的抽象，从而在交流和理解中造成障碍。

在另一个方向上，开发人员在领域模型中引入太多的技术关注也会导致这种分歧。

因此，您可以考虑实践关注点分离，例如持久性，可以帮助防止这些设计和分析模型出现分歧。如果觉得有必要在模型中引入持久性之类的东西，那么这就是一个危险信号。也许这个模型对于实现来说并不实用。

引用:

“单一模型减少了出错的可能性，因为现在的设计是经过仔细考虑的模型的直接产物。设计，甚至代码本身，都具有模型的交流性。”

我对此的解释是，如果你最终要用更多行代码来处理数据库访问之类的事情，你就失去了交流能力。

如果需要访问数据库是为了检查唯一性之类的事情，请查看:

Udi Dahan:团队在应用DDD时所犯的最大错误

http://gojko.net/2010/06/11/udi-dahan-the-biggest-mistakes-teams-make-when-applying-ddd/

"所有规则都是不平等的"

and

使用领域模型模式

http://msdn.microsoft.com/en-us/magazine/ee236415.aspx#id0400119

在“不使用领域模型的场景”中，涉及到相同的主题。

如何分离数据访问

通过接口加载数据

“数据访问层”已经通过一个接口抽象出来，你可以调用它来检索所需的数据:

var orderLines = OrderRepository.GetOrderLines(orderId);

foreach (var line in orderLines)
{
     total += line.Price;
}

优点:该接口分离了“数据访问”管道代码，允许您仍然编写测试。数据访问可以在逐案处理的基础上，允许比通用策略更好的性能。

缺点:调用代码必须假定哪些已加载，哪些未加载。

出于性能原因，GetOrderLines返回带有空ProductInfo属性的OrderLine对象。开发人员必须非常了解接口背后的代码。

我在实际系统中尝试过这种方法。为了修复性能问题，您不得不一直更改加载内容的范围。最后，您需要查看接口背后的数据访问代码，以查看加载了什么，没有加载什么。

现在，关注点分离应该允许开发人员一次尽可能多地关注代码的一个方面。接口技术删除了数据是如何加载的，但没有删除加载了多少数据、何时加载以及在何处加载。

结论:分离度相当低!

延迟加载

数据按需加载。加载数据的调用隐藏在对象图本身中，其中访问属性可能导致在返回结果之前执行sql查询。

foreach (var line in order.OrderLines)
{
    total += line.Price;
}

优点:数据访问的“时间、地点和方式”对于专注于领域逻辑的开发人员来说是隐藏的。在聚合中没有处理加载数据的代码。加载的数据量可以是代码所需的确切数量。

缺点:当你遇到性能问题时，如果你有一个通用的“一刀切”的解决方案，就很难解决问题。延迟加载会导致整体性能变差，实现延迟加载可能很棘手。

角色接口/主动抓取

每个用例都是通过聚合类实现的角色接口显式实现的，允许每个用例处理数据加载策略。

抓取策略可能是这样的:

public class BillOrderFetchingStrategy : ILoadDataFor<IBillOrder, Order>
{
    Order Load(string aggregateId)
    {
        var order = new Order();

        order.Data = GetOrderLinesWithPrice(aggregateId);
    
        return order;
    }

}
   

然后你的集合看起来像这样:

public class Order : IBillOrder
{
    void BillOrder(BillOrderCommand command)
    {
        foreach (var line in this.Data.OrderLines)
        {
            total += line.Price;
        }

        etc...
    }
}

BillOrderFetchingStrategy用于构建聚合，然后聚合完成它的工作。

优点:允许为每个用例定制代码，从而实现最佳性能。符合界面隔离原则。没有复杂的代码需求。聚合单元测试不必模拟加载策略。一般的加载策略可以用于大多数情况下(例如“加载全部”策略)，特殊的加载策略可以在必要时实现。

缺点:开发者在修改域代码后仍然需要调整/审查抓取策略。

使用获取策略方法，您可能仍然会发现自己在更改业务规则时更改自定义获取代码。这不是一个完美的关注点分离，但最终会更易于维护，并且比第一种选择更好。抓取策略确实封装了如何、何时和何处加载数据。它具有更好的关注点分离，而不会像“一刀切”的惰性加载方法那样失去灵活性。

实体只捕获与其有效状态相关的规则。其中的数据有效吗?其中的数据可以这样改变吗?

聚合根对一组实体执行相同的操作。汇总的数据有效吗?总体数据能以这种方式改变吗?

域服务捕获关于实体或聚合之间更改的规则。我们可以这样改变X和Y吗?

None of this ever requires access to a repository or to infrastructure. What you do is that an application service will offer up a domain use case, for that use case, the application service will gather all the needed data from the repositories, that will return it your domain entities and/or aggregate roots and their value objects. The entities/aggregate roots and value objects would have validated that they are in a good state when created by the repository. Then the application service will use a combination of those entities (some of them could be aggregate roots), to perform the domain use case. If the domain use case requires changing X, Y and Z, the application service will ask X, Y and Z entities/aggregate roots if the current use case request of changes can be made to X, Y and Z, and if so, how should it be made. Finally, the application service will commit those changes back to the repository.

如果某些更改跨越实体或聚合，应用程序服务将使用域服务询问是否可以进行更改以及如何进行更改，并再次使用存储库提交这些更改。

如果一个域用例跨越多个有界上下文，这意味着它需要跨有界上下文的信息或更改，这被称为流程，并且您可以让一个流程服务管理整个流程生命周期，它将利用多个有界上下文的应用程序服务来跨所有有界上下文协调整个流程。

Finally, the application service can also use other application services, could be other micro-services in a shared bounded context, that would imply they share the same domain model, or it could do so across to application services in other bounded contexts, in which case you'd want to model those within your own bounded context's domain model as well, you'd treat those other bounded contexts much like a repository in a way. The application service communicates with another bounded context to get info about that other context, it then creates a representation of that info within its own domain model, using its own entities and VOs, and aggregates, which will again validate that state within their context. Similarly, you can commit changes to your domain model to other bounded contexts by asking them to change accordingly. All this can be implemented with direct method calls, remote API calls, async events, shared kernel, etc.

And to answer why it is like so, that's because the whole point is building software that can evolve over time without it becoming slower to make changes to it and add/modify its behavior while retaining its current correctness with regards to its current functionality. A good way to do this is by making it a change in one place doesn't break things elsewhere. This is why bounded contexts exist, already changes are restricted to each context, so a change in one is less likely to break another. This is also why the domain model validates all changes to the domain state, so you can't change part of the state in ways that breaks other usage of it. This is why aggregates are used, to maintain a change boundary between the things that need one, and clearly not have one where it doesn't need one. Finally, by having the whole domain layer, with domain model and domain services, not depend on any infrastructure, like the repository (and thus the DB), a change to the DB or repository will also not be able to break your domain model or services.

P.S.: Also note I use the term "state" loosely. It doesn't have to be a static value; state could be the application of some dynamic computation or rules that generates state when requested. You can have something like totalItemsCount on some entity which computes it when asked about what is the current totalItemsCount for the entity. Again, the entity will make sure to return you valid state, that means it will know how to correctly count the total and make sure that what is returned is the correct application of the domain rules for totalItemsCount.

起初，我相信允许我的一些实体访问存储库(即没有ORM的惰性加载)。后来我得出结论，我不应该这样做，我可以找到其他方法:

We should know our intentions in a request and what we want from the domain, therefore we can make repository calls before constructing or invoking Aggregate behavior. This also helps avoid the problem of inconsistent in-memory state and the need for lazy loading (see this article). The smell is that you cannot create an in memory instance of your entity anymore without worrying about data access. CQS can help reduce the need for wanting to call the repository for things in our entities. We can use a specification to encapsulate and communicate domain logic needs and pass that to the repository instead (a service can orchestrate these things for us). The specification can come from the entity that is in charge of maintaining that invariant. The repository will interpret parts of the specification into it's own query implementation and apply rules from the specification on query results. This aims to keep domain logic in the domain layer. It also serves the Ubiquitous Language and communication better. Imagine saying "overdue order specification" versus saying "filter order from tbl_order where placed_at is less than 30 minutes before sysdate" (see this answer). It makes reasoning about the behavior of entities more difficult since the Single-Responsibility Principle is violated. If you need to work out storage/persistence issues you know where to go and where not to go. It avoids the danger of giving an entity bi-directional access to global state (via the repository and domain services). You also don't want to break your transaction boundary.

据我所知，Vernon Vaughn在红皮书《实现领域驱动设计》中有两个地方提到了这个问题(注意:这本书完全得到了Evans的支持，你可以在前言中读到)。在第7章关于服务的章节中，他使用域服务和规范来解决聚合使用存储库和另一个聚合来确定用户是否经过身份验证的需求。引用他的话说:

根据经验，我们应该尽量避免使用存储库 (12)从聚合体内部，如果可能的话。

沃恩·弗农(2013-02-06)。实现领域驱动设计(Kindle位置6089)。培生教育。Kindle版。

在关于聚合的第10章中，在“模型导航”一节中，他说(就在他建议使用全局唯一id来引用其他聚合根之后):

通过标识引用并不完全阻止导航通过 该模型。有些人会在聚合中使用存储库(12) 查找。这种技术称为断开域模型 它实际上是惰性加载的一种形式。有一个不同的建议 但是，方法是:使用存储库或域服务(7)来查找 在调用聚合行为之前调用依赖对象。一个客户端 应用服务可以控制这个，然后分派到聚合:

他接着展示了一个代码示例:

public class ProductBacklogItemService ... { 
    ...
    @Transactional 
    public void assignTeamMemberToTask( 
        String aTenantId, 
        String aBacklogItemId, 
        String aTaskId, 
        String aTeamMemberId) { 

        BacklogItem backlogItem = backlogItemRepository.backlogItemOfId( 
            new TenantId(aTenantId), 
            new BacklogItemId(aBacklogItemId)); 

        Team ofTeam = teamRepository.teamOfId( 
            backlogItem.tenantId(), 
            backlogItem.teamId());

        backlogItem.assignTeamMemberToTask( 
            new TeamMemberId( aTeamMemberId), 
            ofTeam,
            new TaskId( aTaskId));
   } 
   ...
}

他接着还提到了另一种解决方案，即如何在聚合命令方法中使用域服务以及双重分派。(读他的书大有益处，我怎么推荐都不为过。在你厌倦了无休止地在网上翻找之后，拿出你应得的钱去读这本书。)

然后我和Marco Pivetta @Ocramius进行了一些讨论，他向我展示了一些从域中提取规范并使用它的代码:

不建议这样做:

$user->mountFriends(); // <-- has a repository call inside that loads friends? 

在域服务中，这是很好的:

public function mountYourFriends(MountFriendsCommand $mount) {
    $user = $this->users->get($mount->userId()); 
    $friends = $this->users->findBySpecification($user->getFriendsSpecification()); 
    array_map([$user, 'mount'], $friends); 
}

这里有点混乱。存储库访问聚合根。聚合根是实体。这样做的原因是关注点的分离和良好的分层。这在小型项目中没有意义，但如果您在一个大型团队中，您会说:“您通过产品存储库访问产品。Product是实体集合的聚合根，包括ProductCatalog对象。如果你想要更新ProductCatalog，你必须通过ProductRepository。”

通过这种方式，您在业务逻辑和内容更新的位置上有非常非常清晰的分离。你不会有一个孩子独自编写了整个程序，对产品目录做了所有这些复杂的事情，当涉及到将其集成到上游项目时，你坐在那里看着它，意识到这一切都必须抛弃。这也意味着当人们加入团队，添加新功能时，他们知道该去哪里以及如何组织程序。

But wait! Repository also refers to the persistence layer, as in the Repository Pattern. In a better world an Eric Evans' Repository and the Repository Pattern would have separate names, because they tend to overlap quite a bit. To get the repository pattern you have contrast with other ways in which data is accessed, with a service bus or an event model system. Usually when you get to this level, the Eric Evans' Repository definition goes by the way side and you start talking about a bounded context. Each bounded context is essentially its own application. You might have a sophisticated approval system for getting things into the product catalog. In your original design the product was the center piece but in this bounded context the product catalog is. You still might access product information and update product via a service bus, but you must realize that a product catalog outside the bounded context might mean something completely different.

Back to your original question. If you're accessing a repository from within an entity it means the entity is really not a business entity but probably something that should exist in a service layer. This is because entities are business object and should concern themselves with being as much like a DSL (domain specific language) as possible. Only have business information in this layer. If you're troubleshooting a performance issue, you'll know to look elsewhere since only business information should be here. If suddenly, you have application issues here, you're making it very hard to extend and maintain an application, which is really the heart of DDD: making maintainable software.

对评论1的回应:好问题。所以并不是所有的验证都发生在领域层。夏普有一个属性“DomainSignature”，它可以做你想要的。它是持久性感知的，但是作为一个属性可以使域层保持干净。它确保在您的示例中，没有名称相同的重复实体。

But let's talk about more complicated validation rules. Let's say you're Amazon.com. Have you ever ordered something with an expired credit card? I have, where I haven't updated the card and bought something. It accepts the order and the UI informs me that everything is peachy. About 15 minutes later, I'll get an e-mail saying there's a problem with my order, my credit card is invalid. What's happening here is that, ideally, there's some regex validation in the domain layer. Is this a correct credit card number? If yes, persist the order. However, there's additional validation at the application tasks layer, where an external service is queried to see if payment can be made on the credit card. If not, don't actually ship anything, suspend the order and wait for the customer. This should all take place in a service layer.

不要害怕在服务层创建可以访问存储库的验证对象。让它远离域层。

弗农·沃恩给出了一个解决方案:

使用存储库或域服务提前查找依赖对象 调用聚合行为。客户端应用程序服务可以 控制这个问题。