In my opinion, one of the most insidious form of technical debt is what I like to call future-coding. You might also call it over-engineering, second-system syndrome, etc. It’s code written in an overly elaborate and general way in order to, the future-coder reasons, pre-emptively handle use cases that we’ll “probably see in the future”.
I find it more treacherous than plain “stupid code” because it predominantly affects smart people. It’s mostly smart people with limited experience (once you see the downsides of future coding a few times, you tend to knock it off), but still, these are bright people you’d want to hire and keep around in your organization.
Let me illustrate my point using a hypothetical example with two equally smart programmers, where one just has a little bit more experience than the other and is less prone to future-coding.
The Tale of Two Street Lights
Let’s say you’re making a game where a large portion of the time the player is navigating an urban environment. One day the artists ask for a feature to make it easier to place street lights in a level. They say: “Can we have a way to click a path through the level and have the game place lights at an even spacing along it?” This seems like an easy enough request so you start implementing it.
Here’s where our story forks into two divergent paths.
First option – the pragmatic coder
The pragmatic coder writes some quick UI in the editor where the artist clicks on points in the level that gets stored into an array.
On the runtime side he writes a dozen-line for-loop that simply walks the path segments at equal spacing and spawns a street light each time. Job done, the pragmatic coder moves on to the next task.
Second option – the future-coder
The future-coder writes the same UI as the pragmatic coder, but for the runtime he decides upon a different approach. After all, he’s read books about design patterns and large scale software design, and he knows about the importance of the Single Responsibility Principle, and keeping code factored for changes that might occur in the future.
Thus, he decides to abstract over the exact type of path used. Sure, the artists only asked for a simple path today, but in the future they might want a B-spline or some other kind of curve, so he pulls out the logic for the path into its own Path object that can step along the curve in some kind of iterator fashion. Of course, that’s not enough, in order to change the behavior of the curve he puts this Path object behind an interface and adds a factory to spawn the right kind of path based on some kind of configuration parameter. He also updates the UI to allow for future kinds of paths. Maybe he even implements a B-spline option just so that he can test this properly with more than one option.
Now, the code to spawn street lights first creates an abstract path from a factory, but how does he determine the spacing of the objects? The artists asked for simple equal spacing today, but in the future they might want something different, so he writes another abstract class that handles the spacing policy (with a factory to go with it).
Street lights in your game happens to be reasonably symmetric, so today just spawning them with their default orientation works fine, but in the future the artists might want to use this system to spawn objects where the orientation needs to depend on the curve, so our future-coder updates the path object to also supply tangents and bi-tangents, and writes an abstract class to handle the specific placement policy in order to take this new information into account.
Phew, that was a lot of work – he’s added a lot of new classes and glue code, and the code that actually uses it is a lot more opaque and hard to read than what the pragmatic programmer came up with. It will be worth it in the future, however, when this extra flexibility will pay off. The future-coder goes home late today, but feels good that he’s written an elegant system that will stand the test of time, rather than just a dumb one-off implementation.
Reality Kicks In!
Let’s go through a few scenarios of what might happen next.
Good news, no changes!
Turns out this simple system was all the artists ever needed. They never ask for any changes, and the code is shipped unchanged from the original implementation.
The pragmatic coder feels good that his code did exactly what was needed and never caused any issues or technical debt in the project.
The future-coder feels bad that he spent all this time on an extensible system and nobody ever made the most of the flexibility in his elegant design, but at least it was used at all and the artists liked it, so it’s not too bad.
The feature gets cut
A week later the artists decide that they don’t mind placing street lights manually after all, since it gives them extra control. The automatic street light placing feature goes unused, and is eventually deleted.
The pragmatic coder barely remembers the feature and couldn’t give a damn.
The future-coder feels bad that his elegant system was completely deleted. He grumbles to the other programmers about the awful artists and designers who always keep changing the requirements and wasting so much of his time asking for stuff that they end up not using.
The artists are happy with the feature, but they have one minor request: “Yes, I know we asked for equal spacing of the street lights, but actually we want to make sure that there’s always a street light on each corner of the path, and that the spacing within each path segment is slightly adjusted to ensure this”. To the artists this feels like a simple change so they don’t feel too bad about asking for it.
For the pragmatic coder this is a trivial change. He updates his dozen-liner to first compute the number of street lights required for each segment by dividing the length by the desired spacing and rounding it to the nearest integer, then computing an updated spacing by dividing the segment length by this count. This takes a few minutes of his time and he moves on to other things.
The future-coder starts thinking about how to accommodate this new request into his streetlight-placing framework. Shit-shit-shit, he thinks, a frustrated panic starts to spread in his mind. He never anticipated this change and can’t really see how to update his system to handle it. What does it even mean to place a light at “each corner” when the path is completely abstract? After all, if a future path is defined with a spline, the “corner” may not have a control point on it.
And even if you did have a way to define what a “corner” means, how do you feed that data to the object that governs the spacing policy, now that it’s completely decoupled from the path object? This will require some major re-architecting he thinks, and postpones the task until he can find a larger slot in his schedule. Later that night the future-coder sits in a bar with a programmer coworker, complaining about those damn artists that keep changing the specs on him, oblivious to the irony that it’s precisely his overly complicated attempt to be robust to spec-changes that is causing him problems in the first place.
Ah, a solution! He decides to define “corners” in terms of the curvature of the path, with a configurable threshold, and then feeds this data to the spacing policy object so it can use it to compute the appropriate spacing. The code gets kind of messy since the path and spacing classes are now getting more coupled, and there’s a lot of tunable parameters involved, but it does the job so he checks it in and goes home.
The next day the artists informs him that his solution doesn’t work at all. They want to use this as way to ensure exact placement of streetlights even on straight-line stretches of road, by placing dummy control points in the path. So this notion of using curvature to define street light positions just doesn’t work.
Let’s leave the future-coder to work out a solution to this self-inflicted problem. I have no idea what he could be doing to rescue his system at this point, but whatever he comes up with it’s unlikely to be pretty, leading to even more technical debt and complexity.
Requirements change - redemption for our future-coder!
In this scenario the requirements change in a different way. The artists now want to have the option of using either a straight path with sharp corners, or a smooth B-spline path.
The pragmatic coder adds a flag to the UI, and a switch in his runtime code that either calls the old placement code, or calls some new code that interprets the control points as a B-spline and steps through them using his math library’s B-spline code. The code is still easy to understand, almost certainly contains no bugs, and it didn’t take too long to update.
The future-coder feels redeemed! He may even have had B-spline option already implemented so he can simply point the artist at the appropriate UI option. But even if he has to implement the B-spline option from scratch it’s a simple matter of implementing a new Path object that iterates over a B-spline, and update all the relevant factories and configuration options. Yes it’s more code than what the pragmatic coder needed to write to make this change, and it’s still much harder to read and navigate, and it’s not at all as clear that it doesn’t contain bugs. But look at the UML-diagram! It’s so elegant, and follows the single responsibility principle, and dammit we’ve already seen that we’ve needed one additional option already, so in the future when the number of options grows to a dozen or so we’ll really see the benefit of this elegant framework!
The future-coder wins the lottery
This time the scenario is that the artists keep asking for more and more kinds of paths over time, with increasingly elaborate spacing options, and it becomes one of the most reused systems in the whole project.
After the a few different path options have been added, each having a few different spacing options, the pragmatic coder starts seeing the limits of his original design, and decides that a more abstract and generic design will have greater benefits than downsides. So he refactors his code so that the path is specified by an abstract path object, and pulls the spacing policy out into its own object too. At the end of the day he ends up with a system much like the future-coder, but doesn’t feel too bad because he spent almost no time on the original systems in the first place, and the simplicity of them meant he could get the initial system going quickly, and respond to change-requests faster than the future-coder could’ve.
The future-coder feels like a goddamn programming genius. He totally predicted this and it paid off! He is now more certain than ever that just spending some time up front predicting the future will pay off down the line.
He’s blind to the fact that for every time he happens to get it right, there are a hundred other instances where his failed predictions led to unnecessarily complicated code, time wasted debugging because you can’t tell that the code obviously has no deficiencies, and additional development effort because the extra scaffolding required by his general systems made it harder to address changes that weren’t anticipated.
It’s like he’s won at the roulette table in Vegas – he’s convinced that he’s just that good, that he’s found_ a system_. He’ll keep spending his money at the casino, without ever really doing the math on whether or not his overall winnings is higher than his losses. Everyone says they’re up on average in Vegas right? It’s a mystery that Casinos still make money!
We probably all have a bit of the future-coder in us, but we must resist him! The bottom line is that the future-coder almost never wins. In reality, predicting the future is virtually impossible, so don’t even try. It hardly ever pays off. Best case, you’ve just wasted a lot of time, but in reality abstraction has cognitive overhead throughout the project when reading or debugging code, and can make it harder to deal with unanticipated changes (ironically), as well as cause performance issues. And even when it does pay off it doesn’t actually buy you all that much and is actually a net negative because it skews your intuition into thinking that future-coding is a good thing.
The human brain didn’t evolve to think in abstract terms. You will probably have noticed this when trying to understand a new mathematical concept – it’s much easier if you start with concrete examples rather than the fully general definitions and proofs! Simple code that just does one thing in the dumbest way possible takes very little time for someone to understand, whereas if you have to jump through abstract classes and factories (each named in suitably abstract ways) it can take an absurd amount of time to understand something that really isn’t all that complicated.
All these extra virtual calls act as barriers to inlining, in addition to the overhead of the virtual function calls themselves, which cause performance issues. Yes, I know you’ve been told that you shouldn’t worry about performance for code that isn’t in a hotspot, but the reality is that for many applications there aren’t any real hotspots anymore. It’s depressing to look at the profiler for gameplay code and see that your warmest “hotspot” takes 5% of your frame time, and has already been optimized to death. By that time it’s often too late – performance has been siphoned off by tens of thousands of small inefficiencies added over many years, all because people didn’t want to worry about minor performance hits for code that wasn’t on the mythical “hot-path”. Simple code tends to be cheaper to execute, most of the time, in addition to being better in almost every other way as well.
So am I saying that we should just hack things together in the quickest way possible, with no regard for future maintainability? No, of course not. I’m advocating doing things in the simplest way possible, not the quickest way possible. Often, the simplest way requires removing a bunch of cruft that has accumulated over time, deleting stateful flags, pulling logic out into clearly named functions etc. Sometimes, it even requires an abstract factory.
You should optimize for simplicity, readability and modifiability. If you do, you will also end up with code that’s more maintainable, easy to modify and optimize in the future. Take the example above where our heroes needed to add support for a B-spline. Let’s assume that there was no library for interpolating a B-spline already so that they had to add it. I’d expect both types of programmers to pull that code out into its own function. My motivation for doing this isn’t that I’m trying to predict the future, though! You pull it out into its own function because doing so makes the code easier to read _today, _by making sure the B-spline functionality is hidden behind a sensibly named function instead of cluttering up the “flow” of the main function with unnecessary details. The minor cognitive cost of the extra function call is vastly outweighed by the increased readability of hiding unimportant details. It’s not about predicting the future, it’s about increasing readability.
Sensibly named function calls without side effects almost always add more to readability then they cost, if they’re virtual less so, if they have side effects even less so, if they’re in a class even less so, and so on. It’s all a tradeoff, and the break-even point for any given abstraction will change depending on the specifics of what you’re doing.
Very few abstraction techniques have no merit at all, but the road to programmer hell is paved with best-practices, applied too early. You should postpone applying these abstraction techniques until the cognitive and other costs of using them are clearly outweighed by their benefits (e.g. when you have demonstrated reusability for a component). There are often genuine systems even in game engines where it’s perfectly reasonable to apply many of the abstraction techniques you’ve read about in books about design patterns and large scale software design etc., but the vast majority of code is better off if you keep things simple, and avoid trying to predict the future.