Essay on Inheritance - Popular, Attractive, and Not Very Good

Despite being one of the most popular code-reuse pattern, inheritance often leads to messy, less maintainable codebase in the long term. Why is it popular if it's bad? What can you do instead?

Inheritance is still a widely used tool in software design. I think there is something fundamental to inheritance that makes it attractive to software developers. But it’s not often the right tool to use.

Ten years ago, I was doing web development. The web application I was working on used Widget classes which produced HTML and JavaScript markup for the browsers. Unfortunately, we were still dealing with ancient quirky browsers, so we had to emit different markup for different browsers. We decided to use inheritance to solve this problem. We named each class after the browsers they support: IEWidget and FirefoxWidget, each inheriting from Widget.

The subclasses overrode relevant functions to adapt to different browser behaviours. This worked well for a while until a new popular browser called Chrome entered the market. The natural reaction to this was to create a new subclass for Chrome, except this couldn’t be done without duplicating a lot of code from the other classes. Clearly, the class design wasn’t working very well.

It became much worse as we moved into the wild world of mobile browsers where there were more than a handful of browsers with different quirks. It was obvious to me this design didn’t really scale but I didn’t know what to do back then.

Over time as I worked with more code, "composition over inheritance" suddenly clicked. I've read about it before, but it wasn’t obvious to me how to apply it well. I’ve worked at a few more companies since then, but I still see it misused all the time.

Popularity of Inheritance

In 2001, Joshua Bloch wrote in Effective Java why we should favour composition over inheritance. It’s been almost two decades since then. Why do we keep seeing this? I can think of a few reasons for this.

First, there is a whole generation of developers who were taught inheritance as the primary method of OOP design. Many earlier popular libraries have been written using inheritance. The famous unit test framework JUnit introduced the concept of unit test classes where you inherit from the base TestCase class. JUnit moved on to the new annotation-based style but it didn’t stop developers from structuring tests the old way. Newer frameworks like pytest use composition via dependency injection as the primary method for organizing test dependencies. My experience with pytest has been very positive. The framework naturally leads to more modular tests.

Fortunately, this seems to be going away. When I took a software design course in 2011, they didn’t teach composition over inheritance. Now it seems to be part of the curriculum.

Second, inheritance offers the path of the least resistance for implementation reuse. Mainstream languages like Java do not offer a convenient way to do composition, at least compared to inheritance. Inheritance gives you a special syntax to construct your parent class, for example. Compare that to composition where you have to pass the object via a constructor, add a member variable, and then every method call over that object has to be qualified with the name of the object. It feels wrong to have to go through so many steps! Combined with the tendency for Java developers to have verbose variable names, no wonder many people default to inheritance (just imagine having to write abstractCollection.x(), instead of x() for every delegated function x).

Another reason is that it takes a lot of experience and deliberate thinking about software design to understand and experience issues with inheritance. Let’s go back to the WebWidget example I mentioned earlier. The design worked fine for many years until the requirements changed (i.e., numerous new browsers). Once the requirements outgrew the design, the signs of a design breakdown like downcasting appeared in the codebase (i.e., instanceof in Java & dynamic_cast in C++). Unfortunately, by the time this happens, the original designers may not even be around to learn the consequences of their design. Even if they were around, they would have to know about the alternative design choices (like composition) to realize how it could have been done differently. Put it another way, you have to be at the right place at the right time to learn the lesson.

Deadly Attraction of Inheritance

Unlike the long-term downsides, there is an immediate upside to using inheritance. It gives developers nice, warm feelings right away. Seriously, developers have an occupational disease—genericitis—for both reusing code and categorizing objects, sometimes to their demise. Inheritance as a tool does both, so it’s insanely attractive for developers. It’s a dangerous trap, really. It feels good now, but it hurts you later. Many leaders in software design have been warning us about this particular trap through rules like composition over inheritance, the rule of three and YAGNI (you aren’t gonna need it). Unfortunately, they are not as well-known as principles like DRY (don’t repeat yourself). My explanation for this is that principles like the rule of three embody the next level of understanding above principles like DRY. This is something worth exploring more deeply.

Underneath the rule of three is the learning that we are not so good at predicting the future. This is well-known in the project management circles as the cone of uncertainty. Software design at its core is about making bets about the future. We predict what data belongs where, and how objects will interact with each other. When we get the design right, it pays off by making the code easier to modify and extend. On the other hand, when you make a wrong design decision, it bites us back with higher maintenance costs later. The stronger your prediction is, the more expensive it gets when you get it wrong. There is value in making a weaker prediction because it will cost you less when you get it wrong.

Let’s connect this back to inheritance. Inheritance is a very narrow prediction about the future. It defines strict subtyping relationships. Implicit in the relationship is the assumption that child classes do not overlap, and that they fit in a hierarchical category sharing implementation in a certain way. However, unlike mathematical objects, real-world entities can rarely be categorized hierarchically. The world of web browsers looked hierarchical until they weren’t. File systems were categorized into the "Unix-type" (/ as path separator, and case sensitive), the "Windows-type" (\ as path separator, and case insensitive), until they couldn’t be—HFS+ on MacOS uses / as path separator but it is case-insensitive. Evolution looked like a mathematical tree until we found out about the horizontal gene transfers. Hierarchical categorization is a good aid to understand the world, but it is often not the right tool to describe the truth. Inheritance is a narrow bet that is unlikely to pay off later.

One important thing to note is that the issues of the hierarchical categorization don’t apply when we use them to model an artificial world like mathematics. Every natural number is an integer. The set of all natural numbers don’t overlap with negative integers. We can be assured of this relationship not changing because we defined them that way. The troubles occur when you conflate prescriptive concepts like integers with real-world descriptive concepts like web browsers.

Difficulty of Advocating for Simpler Design

Advocating simpler designs at work could be challenging. It takes a lot of courage and conviction to say, “I don’t know what the future looks like”. Convincing others of this is a fundamentally asymmetric battle. Fancy designs, however wrong they may be in the future, sounds a lot cooler. On the other hand, enumerating all the ways the design could go wrong is much harder. The irony is that it’s harder because future prediction is hard.

This do something vs. do nothing asymmetry can be found in other fields. In medicine, iatrogenesis refers to the harms caused by medical professionals attempting to heal, rather than not doing anything. For example, the misuse of antibiotics brought us the lethal superbugs. Even though antibiotics don’t help with the common cold, many people still believe that it’s effective and demand their doctors give them the drugs. It’s much harder to explain to patients why antibiotics don’t work for colds than to write them a prescription and keep the patients happy. Never mind those micro bacterial superbugs kill thousands every year, unlike the common cold.

It’s human nature to do something than nothing even when it’s harmful to do things. Taleb talks about this problem in his book Antifragile.

What can we do about it

Here’s a very practical set of ideas for all of us to fight this problem.

As an individual developer, you can start building your modules using composition and show your co-workers that there are different, better ways to organize code. The maintenance benefit may not show immediately. One thing that will show immediately, though, is the ease of unit testing, as it is significantly easier to test delegation than different branches of inheritance hierarchy.

If you are using Python and are using the unittest module to write tests, consider switching to pytest. pytest supports the legacy unittest style tests as well so you can have an easy transition.

If you are using C++, you can make use of private inheritance over public when the subtyping relationship is not warranted.

As for Java, I think developers should consider using super-short single-letter member names for trivial compositions that would have been inheritances (e.g., a instead of abstractCollection). Some code reviewers may flinch when they see a single letter variable name, but I don’t think such a reaction is warranted. It’s still clearer than the inherited methods where method names are imported completely unqualified, and possibly from multiple ancestors. Such composition is supported by the same principle that recommends against import *, that it’s bad to import unspecified symbols into the scope. On the other hand, making variable names verbose when its meaning is unambiguous is not supported by any reason.

Finally, you can spread the word by sending your coworkers a link to this blog post.