When to use interfaces or abstract classes? When to use both?

Question

While certain guidelines state that you should use an interface when you want to define a contract for a class where inheritance is not clear (IDomesticated) and inheritance when the class is an extension of another (Cat : Mammal, Snake : Reptile), there are cases when (in my opinion) these guidelines enter a gray area.

For example, say my implementation was Cat : Pet. Pet is an abstract class. Should that be expanded to Cat : Mammal, IDomesticated where Mammal is an abstract class and IDomesticated is an interface? Or am I in conflict with the KISS/YAGNI principles (even though I'm not sure whether there will be a Wolf class in the future, which would not be able to inherit from Pet)?

Moving away from the metaphorical Cats and Pets, let's say I have some classes that represent sources for incoming data. They all need to implement the same base somehow. I could implement some generic code in an abstract Source class and inherit from it. I could also just make an ISource interface (which feels more "right" to me) and re-implement the generic code in each class (which is less intuitive.) Finally, I could "have the cake and eat it" by making both the abstract class and the interface. What's best?

These two cases bring up points for using only an abstract class, only an interface and using both an abstract class and an interface. Are these all valid choices, or are there "rules" for when one should be used over another?

---

I'd like to clarify that by "using both an abstract class and an interface" that includes the case when they essentially represent the same thing (Source and ISource both have the same members), but the class adds generic functionality while the interface specifies the contract.

Also worth noting is that this question is mostly for languages that do not support multiple inheritance (such as .NET and Java.)

Answer 1

I always use these guidelines: - use Interfaces for multiple TYPE inheritance (as .NET/Java don't use multiple inheritance) - use abstract classes for a re-usable implementation of a type

The rule of the dominant concern dictates that a class always has a main concern and 0 or more others. (see: http://citeseer.ist.psu.edu/tarr99degrees.html ). Those 0 or more others you then implement through interfaces, as the class then implements all the types it has to implement (its own, and all interfaces it implements).

In a world of multiple implementation inheritance (e.g. C++/Eiffel), one would inherit from classes which implement the interfaces. (in theory. In practise it might not work that well).

Answer 2

If you want to provide the option of replacing your implementation completely, use an interface. This applies especially for interactions between major components, these should always be decoupled by interfaces.

There may also be technical reasons for prefering an interface, for example to enable mocking in unit tests.

Internally in a component it may be fine to just use an abstract class directly to access a hierarchy of classes.

If you use an interface and have a hierarchy of implementing classes then it is good practice to have an abstract classe which contain the common parts of the implementation. E.g.

interface Foo
abstract class FooBase implements Foo
class FunnyFoo extends FooBase
class SeriousFoo extends FooBase

You could also have more abstract classes inheriting from each other for a more complicated hierarchy.

Answer 3

There is also something called DRY principle - Don't Repeat Yourself.

In your example of data sources you say there is some generic code that is common between different implementations. To me it seems that the best way to handle that would be to have an abstract class with the generic code in it and some concrete classes extending it.

The advantage is that every bug fix in generic code benefits all concrete implementations.

If you go interface only you will have to maintain several copies of same code which is asking for trouble.

Regarding abstract + interface if there is no immediate justification for it I would not do it. Extracting interface from abstract class is an easy refactoring so I would do it only when it is actually needed.

Answer 4

I tend to use base classes (abstract or not) to describe what something is, while I use interfaces to describe the capabilities of an object.

A Cat is a Mammal but one of it's capabilities is that it is Pettable.

Or, to put it a different way, classes are nouns, while interfaces map closer to adjectives.

Answer 5

As a first rule of thumb, I prefer abstract classes over interfaces, based on the .NET Design Guidelines. The reasoning applies much wider than .NET, but is better explained in the book Framework Design Guidelines.

The main reasoning behind the preference for abstract base classes is versioning, because you can always add a new virtual member to an abstract base class without breaking existing clients. That's not possible with interfaces.

There are scenarios where an interface is still the correct choice (particularly when you don't care about versioning), but being aware of the advantages and disadvantages enables you to make the correct decision.

So as a partial answer before I continue: Having both an interface and a base class only makes sense if you decide to code against an interface in the first place. If you allow an interface, you must code against that interface only, since otherwise you would be violating the Liskov Substitution Principle. In other words, even if you provide a base class that implements the interface, you cannot let your code consume that base class.

If you decide to code against a base class, having an interface makes no sense.

If you decide to code against an interface, having a base class that provides default functionality is optional. It is not necessary, but may speed up things for implementers, so you can provide one as a courtesy.

An example that springs to mind is in ASP.NET MVC. The request pipeline works on IController, but there's a Controller base class that you typically use to implement behavior.

Final answer: If using an abstract base class, use only that. If using an interface, a base class is on optional courtesy to implementers.