Why Extending Through Subclassing (a framework's classes) is a Bad Idea

Ok, I’ll admit it, overly ambitious blog post. So I’ll refine it a little now, this is intended mainly as my thoughts on why as a tool developer (one who makes tools/frameworks that other programmers then use), its a bad idea to implement extensible objects via developer subclassing. This is actually how the web framework I wrote - Pylons - provides its extensibility to developers and lets them change how the framework functions.

Please excuse the short and possibly incomplete description, this is mainly a quick post to illustrate a tweet I recently made.

First, some background…

One of the things that Pylons 1.0 and prior is missing is a way to easily extend a Pylons project. While it can be done, its very ad-hoc, kludgy, and generally not very well thought-out. Or thought-out at all really. What was somewhat thought-out was how a developer was supposed to extend and customize the framework.

In a Pylons project, the project creates a PylonsApp WSGI object, and all the projects controllers subclass WSGIController. This seemed to work quite well, and indeed many users happily imported PylonsApp, subclassed it to extend/override methods they needed for customization, or changed how their WSGIController subclass worked to change how individual actions would be called.

Everything seemed just fine…. until…

Improving Pylons

When I had some free time a little while back, I set about looking into how to extend and improve Pylons to make up for where it was lacking, extensibility. I quickly realized that I’d need to change rather drastically how Pylons dispatch worked, and how controller methods were called to make them more easily extendable. But then with a certain feeling of dread, the subclassing issue nipped me. All my implementations of PylonsApp and WSGIController were effectively frozen.

Since every single developer using Pylons sub-classes WSGIController, and to a much lesser extent, PylonsApp, any change to any of the main methods would result in immediate breakage of every single Pylons users app that happened to customize them (the very reason subclassing was used!). This meant that I couldn’t very well change the implementation of the actual classes to fix their design, because that would just cause complete breakage. Ugh!

So after looking into it more, I’ve ended up with this short list of the obvious bad reasons this shouldn’t be done. BTW, in Pylons 2, controllers don’t subclass anything, and customization is all with hooks into the framework, no subclassing in sight!

Short List of Why It’s Bad

From a framework maintainers point of view…

  1. Implementations of the classes are effectively frozen, because all the class methods are the API.
  2. Correcting design flaws or implementation flaws are much more difficult if not impossible without major breakage due to point #1.
  3. Heavily sub-classed/large hierarchy classes can have performance penalties.

From a developer ‘extending’ the classes point of view…

  1. Figuring out how to unit-test is more difficult as the full implementation is not in your own code… its in the framework code.
  2. When using mix-in’s and other classes that also subclassed the framework, strange conflicts or overrides occur that aren’t at all obvious or easy to debug/troubleshoot.

I think there were a few more reasons I came across as well, but I can’t recall them at the moment. In short, I’m now of the rather firm opinion that the only classes you should ever subclass are your own classes. Preferably in the same package, or nearby.