All Code is an API

Many years ago I read an article whose key premise was that all code is in one way or another an API. The original article is lost in the depths of the Internet, and I can’t for the life of me find it (leave a comment if you do manage to find it, I’d be interested in re-reading it!). The article changed how I thought about the code I was writing, so I thought it would be nice to examine the idea here.

An Application Programming Interface (API) is a defined mechanism for programmatically controlling the behaviour of some component. This term is most often used to describe methods by which third parties can integrate with another application - for example Facebook provide an API through which third parties can build applications on top of their platform, Slack expose one that lets you post messages, write bots and more. At Pattern Analytics we provided APIs for the various libraries, and at Allinea we had an API for adding custom performance metrics to the MAP profiler. As a programmer, APIs are incredibly useful - anything I can’t do natively in JIRA I know I can probably get to work with a bit of Python or Node.js. When Rhodecode was lacking a view that let me see all open pull requests (even ones I wasn’t a reviewer on) a few lines of HTML and AngularJS made one (although it isn’t the prettiest thing I’ve ever done). The theme here is that if your product is targeted at developers, you sure as hell need an API if you want them to be invested in your solution.

So what makes an API different from normal code? It’s really only two things, firstly, APIs should be well documented interfaces to some underlying logic. They usually form explicit contracts with their users about how they can be used and how they will behave. Secondly, it is often not possible to look at the implementations - and even when it is possible, it may not be practical to look into the implementation in order to simply use the API. In even more cases it is either not possible or not practical to change the implementation.

Everything is an API

On any reasonably sized project, no matter how many developers there are, every piece of code is an API. Every module, class, or function is a piece of code that a developer might want to reuse in the future, so it needs to be documented, otherwise the developer will be wasting their time digging through the internals to figure out how to avoid duplicating the code. In many cases, the developer might find it easier to just re-write the code (assuming they even realise the undocumented code exists!) Looking at the implementation is impractical - so we need documentation, but what about change? Well, every piece of code in the system might well be used somewhere, so if we change a piece of code we need to make sure we preserve the contracts otherwise we run the risk of breaking dependent code elsewhere in the system.

So, in practice, every trait we want of an API we probably want in normal code. If you’re writing an API you probably put a reasonable amount of thought into your design, because once it’s in the wild you can’t take it back (and because you probably have to support the thing) - if we want all the same traits in normal code, then we should be putting the same amount of thought into our internal modules, classes and functions.

What should an API be?

So what should we be thinking about when we’re designing an API? The Qt project has a nice set of API design principles that I think everyone should know.

  • Minimal: a minimal API exposes only the minimum number of functions and classes. It makes it easier to use because there is less to remember. It is tied to the Single Responsibility Principle and the idea that the component should only do one thing.
  • Complete: a complete API exposes everything a user would expect it to provide, and is about ensuring the component is usable in a variety of contexts.
  • Clear and simple semantics: The Principle of Least Astonishment is a simple idea that prevents nasty surprises for users of your API. Common tasks need to be easy (I’m looking at you Boost date-time library string formatting, and unexpected side effects should be avoided.
  • Intuitive: an intuitive component is one that is easy to use correctly and difficult to use incorrectly. As a guideline, imagine trying to figure out how to use the component for common tasks without looking at the documentation or the implementation. A good API is one where an intermediate programmer can get away with this.
  • Easy to memorise: no-one wants to constantly cross reference documentation when using a library to remember common tasks. A good library has strong naming conventions to ease the cognitive burden, and shares as much as it can with recognisable patterns and concepts (as an example, we have a naming convention where any function that does something analogous to a standard library function will have the same name).
  • Readable: some people might not believe in self documenting code, but APIs should strive to be so. Even worse than having to reference documentation to write code is having to reference documentation to read it. What code does needs to be obvious from the names of the classes and functions alone. A classic example is avoiding the use of boolean function arguments, because it is often necessary to read the documentation to figure out what true and false actually mean.

It’s easy to see that all of the above make for better code, so the sooner we start thinking about all code we write as an API for someone else the higher quality our code bases will be.

Posted on November 10, 2016
Want to see the edit history? Check the source on Github.

More posts