Achieving decoupling in Model classes

Question

I am trying to test-drive (or at least write unit tests) my Model classes but I noticed that my classes end up being too coupled. Since I can't break this coupling, writing unit tests is becoming harder and harder.

To be more specific:

Model Classes: These are the classes that hold the data in my application. They resemble pretty much the POJO (plain old Java objects), but they also have some methods. The application is not too big so I have around 15 model classes.

Coupling: Just to give an example, think of a simple case of Order Header -> Order Item. The header knows the item and the item knows the header (needs some information from the header for performing certain operations). Then, let's say there is the relationship between Order Item -> Item Report. The item report needs the item as well. At this point, imagine writing tests for Item Report; you need have a Order Header to carry out the tests. This is a simple case with 3 classes; things get more complicated with more classes.

I can come up with decoupled classes when I design algorithms, persistence layers, UI interactions, etc... but with model classes, I can't think of a way to separate them. They currently sit as one big chunk of classes that depend on each other.

Here are some workarounds that I can think of:

1. Data Generators: I have a package that generates sample data for my model classes. For example, the OrderHeaderGenerator class creates OrderHeaders with some basic data in it. I use the OrderHeaderGenerator from my ItemReport unit-tests so that I get an instance to OrderHeader class. The problem is these generators get complicated pretty fast and then I also need to test these generators; defeating the purpose a little bit.
2. Interfaces instead of dependencies: I can come up with interfaces to get rid of the hard dependencies. For example, the OrderItem class would depend on the IOrderHeader interface. So, in my unit tests, I can easily mock the behaviour of an OrderHeader with a FakeOrderHeader class that implements the IOrderHeader interface. The problem with this approach is the complexity that the Model classes would end up having.

Would you have other ideas on how to break this coupling in the model classes? Or, how to make it easier to unit-test the model classes?

---

Update after going through the answers:

• I have done a little more thinking on this problem and noticed that some of these dependencies were not really necessary (as most answers pointed out). The reason why they existed is because my UI layer needed it. Basically, I have inserted these dependencies so that UI layer can easily navigate from child to the header and display some header fields at the child level views. Looking at it now, it looks so simple; however it took some time for me to figure out why these dependencies existed in the first place. Now, the UI will basically keep the header in its context when navigating to the child level views.
• On the other hand, I have removed some of the dependencies via interfaces. These dependencies were mostly one-to-one; for example between Header and HeaderReport. Previously, HeaderReport depended on Header for retrieving information (items, date, etc...) and this made HeaderReport very hard to test. Now, the HeaderReport depends on the HeaderDataProvider interface and I am able to achieve complete isolation. HeaderReport is very easy to test with stub classes for the HeaderDataProvider interface.
• It looks easy to create these abstractions (UI Layer - Model Layer - Backend Layer), however in reality, it really takes a lot of practice!

Answer 1

Do the classes really need to be decoupled? In your example, from business point of view, the order header and order item are entities, that are really close. They probably form an aggregate. So trying to decouple those will create unnecessary complexity and will make reading the model confusing.

To solve this, I would try to locate aggregates in your code and then unit test whole aggregates.

Answer 2

I know you're trying to simplify your scenario to make it easier to discuss, but why does the OrderItem care about the Header? The Header is Aggregator (called the "Root" in DDD) and it is responsible for knowing about its components (OrderItem being one of them).

If you're discussing special cases like Discounts being applied to Orders containing for example 10 or more of a certain Item that would be a job of say a discount service to find which discounts apply as the order is being processed.

The same thing would apply to ItemReport...it is responsible for knowing the rules for how the items are selected, filtered, ordered, etc. The Item shouldn't be concerned with reporting, that's not its job.

Answer 3

Would you have other ideas on how to break this coupling in the model classes? Or, how to make it easier to unit-test the model classes?

In my opinion, coding against the interfaces is the way to go. In another words, de-coupling your code by design is the approach that modern projects do follow. That greatly helps with the needs of unit-testing.

In another words, it is recommended to define your formal, precise and verifiable interface specifications for software components that you are planning to develop or trying to de-couple to remove dependencies.

Answer 4

One way to do it is, as you said, to create interfaces between the classes. Each class would take its dependencies as interface references that can be mocked or stubbed in a unit test context. This might make you think a bit more about the essential parts of the design that are needed by each class. For example, you might find that the order item does not need the full order header. This process might lead you to discover latent concepts in the design, which will increase the chances of reuse and create elements more inline with their intent, or raison d'être.

On the other hand, maybe creating interfaces does seem odd for such synergistic and coupled concepts. If that's the case, you can use fluent test data builders to create a domain specific and direct way to easily create dependent objects. You don't need to test these builders directly, as they'll be used by test code to create the dependent objects.