Subclassing

In this article

This article is a little bit different from the usual articles I write. It’s more a collection of thoughts and patterns than a guide. Almost all the patterns I will describe are found out the hard way: by making mistakes. By no means do I consider myself an authority on subclassing, but I did want to share a few things I’ve learned. Don’t read this as a definitive guide, but rather as a collection of examples.

When asked about OOP (object-oriented programming), Alan Kay (the inventor) wrote that it’s not about classes, but rather about messaging.1 Still, a lot of people focus on creating class hierarchies. In this article, we’ll look at some cases where it’s useful, but we’ll mostly pay attention to alternatives to creating complicated class hierarchies. In our experience, this leads to code that’s simpler and easier to maintain. A lot of things have been written on this topic, which you can find in books like Clean Code and Code Complete, both of which are recommended reading.

When to Subclass

First, let’s talk about some cases where it makes sense to create subclasses. If you’re building a UITableViewCell with custom layout, create a subclass. The same holds for almost every view; once you start doing layout, it makes sense to move this into a subclass, so that you not only have your code nicely bundled up, but also have a reusable object that you can share across projects.

Suppose you’re targeting multiple platforms and versions from your code, and you need to somehow write custom bits for every platform and version. It might then make sense to create an OBJDevice class, which has subclasses like OBJIPhoneDevice and OBJIPadDevice, and maybe even deeper subclasses like OBJIPhone5Device, which override specific methods. For example, your OBJDevice could contain a method applyRoundedCornersToView:withRadius:. It has a default implementation, but can be overridden by specific subclasses.

Another case where subclassing might be very helpful is in model objects. Most of the time, my model objects inherit from a class that implements isEqual:, hash, copyWithZone:, and description. These methods are implemented once by iterating over the properties, making it a lot harder to make mistakes. (If you’re looking for a base class like this, you can consider using Mantle, which does exactly this, and more.)

When Not to Subclass

In a lot of projects that I’ve worked on, I’ve seen deep hierarchies of subclasses. I am guilty of doing this as well. Unless the hierarchies are very shallow, you very quickly tend to hit limits.2

Luckily, if you find yourself in a deep hierarchy like that, there are lot of alternatives. In the sections below, we’ll go into each in more detail. If your subclasses merely share the same interface, protocols can be a very good alternative. If you know an object needs to be modified a lot, you might want to use delegates to dynamically change and configure it. When you want to extend an existing object with some simple functionality, categories might be an option. When you have a set of subclasses that each override the same methods, you might instead use configuration objects. And finally, when you want to reuse some functionality, it might be better to compose multiple objects instead of extending them.

Alternatives

Alternative: Protocols

Often, a reason to use subclassing is when you want to make sure that an object responds to certain messages. Consider an app where you have a player object, that can play videos. Now, if you want to add YouTube support, you want the same interface, but a different implementation. One way you can achieve this with subclassing is like this:

@interface Player : NSObject

- (void)play;
- (void)pause;

@end


@interface YouTubePlayer : Player

@end

Most likely, the two classes don’t share a lot of code, just the same interface. When that’s the case, it might be a good solution to use protocols instead. Using protocols, you would write the code like this:

@protocol VideoPlayer <NSObject>

- (void)play;
- (void)pause;

@end


@interface Player : NSObject <VideoPlayer>

@end


@interface YouTubePlayer : NSObject <VideoPlayer>

@end

This way, the YouTubePlayer doesn’t need to know about the Player internals.

Alternative: Delegation

Again, suppose you have a Player class like in the example above. Now, at one place, you might want to perform a custom action on play. Doing this is relatively easy: you can create a custom subclass, override the play method, call [super play], and then do your custom work. This is one way to deal with it. Another way is to change your Player object and give it a delegate. For example:

@class Player;

@protocol PlayerDelegate

- (void)playerDidStartPlaying:(Player *)player;

@end


@interface Player : NSObject

@property (nonatomic,weak) id<PlayerDelegate> delegate;

- (void)play;
- (void)pause;

@end

Now, in the player’s play method, the delegate gets sent the playerDidStartPlaying: message. Any consumers of this class can now just implement the delegate protocol instead of having to subclass, and the Player object can stay very generic. This is a very powerful technique, which Apple uses abundantly in its own frameworks. Think of classes like UITextField, but also NSLayoutManager. Sometimes you want to group different methods together in separate protocols, such as UITableView – which has not only a delegate but also a data source – does.

Alternative: Categories

Sometimes, you might want to extend an object with a little bit of extra functionality. Suppose you want to extend NSArray by adding a method arrayByRemovingFirstObject. Instead of subclassing, you can put this into a category. It works like this:

@interface NSArray (OBJExtras)

- (void)obj_arrayByRemovingFirstObject;

@end

When using categories and extending a class that’s not your own, it’s good practice to prefix your methods. If you don’t, there is a chance that somebody else might implement the same method using the same technique. Then, if the behavior doesn’t match, unexpected things could happen.

One of the dangers of using categories is that you might end up with a large amount of categories, and you can lose your overview. In that case, it’s probably easier to create custom classes.

Alternative: Configuration Objects

One of the mistakes I keep on making (but can now recognize fairly quickly), is having a class with some abstract functionality, and then a lot of subclasses that override one specific method. For example, in a presentation app, you might have a class Theme, which has a couple of properties, such as backgroundColor and font, and some logic for laying things out on a slide.

Then, for each theme, you create a subclass of Theme, override a method (e.g. setup), and configure the properties. Using the superclass directly wouldn’t make sense. In this case, you can make your code a bit simpler by using configuration objects. You can keep the shared logic in the Theme class (e.g. slide layout), but move the configuration into simpler objects that only have properties.

For example, a ThemeConfiguration class would have the backgroundColor and font properties, and the Theme class gets a value of this class in the initializer.

Alternative: Composition

The most powerful alternative to subclassing is composition. If you want to reuse existing code but you’re not sharing the same interface, composition can be your weapon of choice. For example, suppose you are designing a caching class:

@interface OBJCache : NSObject

- (void)cacheValue:(id)value forKey:(NSString *)key;
- (void)removeCachedValueForKey:(NSString *)key;

@end

One easy way to do this is by subclassing NSDictionary and implementing the two methods by calling to dictionary methods:

@interface OBJCache : NSDictionary

However, there are a couple of drawbacks. The fact that it is implemented with a dictionary should be an implementation detail. Now, everywhere where you expect an NSDictionary parameter, you could provide an OBJCache value. If you would ever want to switch to something completely different (e.g. your own library), you might need to refactor a lot of code.

A better approach is to store this dictionary in a private property (or instance variable), and only expose the two cache methods. Now, you maintain the flexibility to change the implementation as you gain more insight, and consumers of your class don’t need to be refactored.


  1. http://c2.com/cgi/wiki?AlanKayOnMessaging 

  2. http://c2.com/cgi/wiki?LimitsOfHierarchies