Discover more from 12 Grams of Carbon
Theory of Programming
Written February, 2023. Published March, 2023.
What is Programming?
The word "programming" has a fascinating etymology. It derives from the word 'program', which in the 19th century meant 'a plan or scheme announced beforehand' and in the 17th century meant 'written notice'. That in turn derived from programma (Latin) and prographein (Greek), meaning 'to write publicly'.
Programming means 'the act of writing publicly'. Seems weird?
I've recently been thinking about Theory-building and why employee churn is lethal to software companies, a blog post from one Baldur Bjarnason. To paraphrase slightly, the post extends on the idea that programming is not about writing code; rather, programming is about solving problems; the code is merely a description of the solution. Without a programmer who understands the code – who has essential context, a model of how the code is meant to function – the code is meaningless. This concept was first proposed by Peter Naur in a 1985 essay called Programming as Theory Building:
A main claim of the Theory Building View of programming is that an essential part of any program, the theory of it, is something that could not conceivably be expressed, but is inextricably bound to human beings. It follows that in describing the state of the program it is important to indicate the extent to which programmers having its theory remain in charge of it. As a way in which to emphasize this circumstance one might extend the notion of program building by notions of program life, death, and revival. The building of the program is the same as the building of the theory of it by and in the team of programmers. During the program life a programmer team possessing its theory remains in active control of the program, and in particular retains control over all modifications. The death of a program happens when the programmer team possessing its theory is dissolved. A dead program may continue to be used for execution in a computer and to produce useful results. The actual state of death becomes visible when demands for modifications of the program cannot be intelligently answered. Revival of a program is the rebuilding of its theory by a new programmer team.
I think this is actually quite insightful, and I wish I had learned about it sooner. This framing helps explain many things about programming that I personally found unexpectedly difficult, like:
Why is understanding a new code base so difficult, even though all of the code is freely present and available?
Why is 'code discovery' a seemingly unsolved problem?
How do you decide between 'learning a piece of code' and 'throwing it out to start from scratch' (especially as regards to Chesterton's fence)?
I've come to believe that many of these difficulties arise from cultural and internal practices that reflect the belief that code is the most important output of a programmer's career. It's not; the model is.
How might we organize our engineering team with this new understanding? What cultural practices should we adopt or change?
An efficient system is one that doesn't do the same work twice. Engineering teams can be modeled as computational systems (see: Organizations As Systems). If programmers solve problems by building mental models, then:
it is vitally important to share this knowledge, because doing so allows our engineers to minimize single points of failure while maximizing parallelism;
no two programmers have to build the same mental model independently.
In an ideal world, once the work of model-building has been done once, the updated understanding of the problem space is immediately shared and equally understood by everyone else on the team.
In practice, this isn't possible – we aren't Borg. Programmers have to look over each other's work to share mental models, and there will always be some information that doesn't transfer. But to avoid duplicating work, we should be making this process as easy as possible. In Corkscrew Development, I talk about how it's often more useful to write helpful documentation than it is to write code:
There are steps you can take to increase the amount of compound value code will provide. Code that is well written, well documented, and has a clean interface is more likely to be reused – not just by you, but also by other people. Going a step further, code that is easily packaged into a binary that requires no technical knowledge to run is, potentially, even more valuable. These steps increase the number of people who can or will actually use that code later. The more people who can use your code, the more likely it will be reused, the more its compounding value.
As in any system optimization, we aim to target the area with the biggest bottleneck to achieve the most improvement with the least work. Same here. We can maximize our efficiency as a system by making the model-update process as low-cost as possible.
Going back to our etymological definition, programming means 'the act of writing publicly'. The goal of programming is not to write code, it is to convey a solution to others as efficiently as possible. A piece of code is a formalized solution, akin to a mathematical proof. But just like academia, you need to provide a lot of context to make sure your proof can be well understood. The philosophers who got the most play in the Greek fora were those who could convey their point most effectively, actual logic be damned.
Unfortunately, it is insufficient to treat model-updates as static, one-time processes. Codebases with multiple collaborators end up behaving like shifting, organic things. Each editor is likely to have a slightly outdated model; each additional update causes everyone else's model to drift. This process eventually (inevitably) leads to staleness – a point where a developer's internal model no longer matches the actual code, and the developer must spend some amount of time updating their model before they can continue their work.
We want to prevent staleness as much as possible, without necessarily incurring the additional cost of simply having everyone keep themselves updated. Below I provide an incomplete list of process improvements (with the intent of growing this list over time). Note: the most efficient form of model transfer is from the original developer to everyone else; perhaps unsurprisingly, many of the approaches listed below take advantage of this asymmetry by loading more work on the original developer, while everyone else can consume the updates passively.
In-line and external documentation both help solve this problem in an a la carte manner. When someone notices staleness, they can use first external and then in-line documentation to help resolve their questions, before diving into the code themselves. The ability for a developer to choose how deep they need to go makes this a particularly effective method for reducing staleness, especially when paired with logical/coherent abstractions.
Code review guarantees that at least one other person on the team has an opportunity to explicitly QA the developer. As per On Code Review, it also provides a chance to evaluate whether the model has been sufficiently externalized.
Regular 'presentations' that allow engineers to share things they have been working on, with an eye towards exposing and clarifying complexity.
Descriptions of changes appearing in Slack or email as they go in, ensuring that everyone is at least passively aware of what changes are being made and who is responsible for making them.
Testing, especially integration testing, allows developers to see a library in action by providing a wide range of possible use cases.