I mentioned this concept on Twitter in a conversation with Miles Sabin and Eiríkr Åsheim last night, and it occurred to me I’ve never written up this idea. I call it my quadratic theory of software projects.
It’s one I originally formulated in the context of program languages, but I’ve since decided that that’s over-simplistic and really it’s more about the whole project of development. It probably even applies perfectly well to things that are not software, but I’m going to be focusing on the software case.
The idea is this: Consider two properties. Call them “effort” and “achievement”, say. If I wanted to attach a more concrete meaning to those, we could say that “effort” is the number of person hours you’ve put into the project and “achievement” is the number of person hours it could have taken an optimal team behaving optimally to get to this point, but the exact meaning of them doesn’t matter – I only mention to give you an idea of what I’m thinking of with these terms.
The idea is this: If you plot on a graph, with achievement on the x axis and the amount of effort it took you to get there on the y axis, what you will get is roughly quadratic.
This isn’t actually true, because often there will be back-tracking – the “oh shit that feature is wrong” bit where you do a bunch of work and then realise it wasn’t necessary. But I’m going to count that as achievement too: You developed some stuff, and you learned something from it.
It’s also probably not empirically true. Empirically the graph is likely to be way more complicated, with bits where it goes surprisingly fast and bits where it goes surprisingly slow, but the quadratic is a useful thought tool for thinking about this problem.
Why a quadratic?
Well, a quadratic has three parts. We’ve got \(y = A + B x + C x^2\). In my model, \(A, B, C \geq 0\).
And in this context, each of those three parts have a specific meaning:
The constant component (A) is the overhead that you had to get started in the first place – planning, familiarising yourself with the toolchain, setting up servers, etc.
The linear factor (B) is how hard it is to actually make progress – for example, if you’re developing a web application in C++ there’s an awful lot of difficulty in performing basic operations, so this factor could be quite high. Other factors that might make it high are requiring a detailed planning phase for every line of code, requiring a 10:1 lines of test code to lines of application code, etc.
The quadratic factor (C) is the interesting one – constant and linear overhead are in some sense “obvious” features, but the quadratic part is something that people fail to take into account when planning. The quadratic overhead is how much you have to think about interactions with what you’ve already done. If every line of code in my code-base affects every other line in my code-base, then I have to deal with that: every line I write, I have to pay an overhead for every line I’ve already written. If on average a line of code interacts with only 10% of the other lines in the project, then I have to pay 10% of that cost, but it’s still linear in the size of the code base (note: I’m implicitly assuming here that lines of code is a linear function of achievement. I think in reality it’s going to be more complicated than that, but this whole thing is an oversimplification so I’m going to ignore that). When you have to pay a cost that is linear in your current progress, the result is that the total amount of cost you’ve paid by a given point is quadratic in your current progress (this is because calculus).
To use an overused word, the quadratic factor is essentially a function of the modularity of your work. In a highly modular code base where you can safely work on part of it without having any knowledge of most of the rest, the quadratic factor is likely to be very low (as long as these parts are well correlated with the bits you’re going to need to touch to make progress! If you’ve got a highly modular code base where in order to develop a simple feature you have to touch half the modules, you’re not winning).
There are also other things that can contribute to this quadratic factor. e.g. the amount that you have to take into account historical context: If a lot of the reasons why things are done is historical, then you have a linear amount of history you need to take into account to do new work. These all essentially work out as the same sort of thing though: The fraction of what you’ve already done you need to take into account in order to do new things.
So here’s the thing: Your approach to development of a project essentially determines these values. A lot of different aspects will influence them – who your team members are, what language you’re working in, how many of you there are, how you communicate, whether you’re doing pair programming, whether you’re doing test driven development, how you’re doing planning, etc and etc. Almost everything you could call “development methodology” factors in somehow.
And if you compare two development methodologies you’d find in active use for a given problem, it’s going to be pretty rare that one of them is going to do better (i.e. have a lower value) on all three of these coefficients: Some approaches might have a lower constant and linear overhead but a remarkably high quadratic overhead, some approaches might have a very high constant overhead but a rather low linear and quadratic, etc. Generally speaking, something which is worse on all three is so obviously worse that people will just stop doing it and move on.
So what you end up doing is picking a methodology based on which constants you think are important. This can be good or bad.
The good way to do this is to look at your project size and pick some constants that make sense for you. For small projects, the constant costs will dominate, for medium projects the linear costs will dominate, and large projects the quadratic costs will dominate. So if you know your project is just a quick experiment, it makes sense to pick something with low linear and constant costs and high quadratic costs, because you’re going to throw it away later (of course, if you don’t throw it away later you’re going to suffer for that). If you know your project is going to be last a while, it makes sense to front-load on the constant costs if you you can reduce the quadratic cost. In between, you can trade these off against eachother at different rates – maybe gain a bit on linear costs by increasing the quadratic cost slightly, etc.
The bad way to do it is to automatically discount some of these as unimportant. If you just ignore the quadratic cost as not a thing you will find that your projects get mysteriously bogged down once you hit a certain point. If you’re too impatient to pay constant costs and just leap in and start hacking, you may find that the people who sat down and thought about it for a couple hours first end up sailing past you. If you think that scalability and the long-term stability of the project is all that matters then people who decided that day to day productivity also mattered will probably be years ahead of you.
Sadly, I think the bad way to do it is by far the more common. I’ll grant it’s hard to predict the future of a project, and that the relationship between methodologies and these values can be highly opaque, which can make this trade-off hard to analyse, but I think things would go a lot better if people at least tried.