Summary:

1. The code that gets written is the code that’s easier to write.
2. Anything not forbidden by the language semantics will be done as a “temporary fix”.
3. Codebases decay along the gradient of expedient hacks.

Programming languages have syntax, semantics, and pragmatics: how the language is used in practice. The latter is harder to design for. Language pragmatics is tooling, best practices, and the code you see in the wild. Pragmatics matter because they determine the shape of the efficient frontier between expedience and engineering quality, and from then on it’s gradient descent.

Contents

1. Types
2. Shotgun I/O
3. Tooling
4. Tedium

Types

Summary: make types easier to define.

One fundamental principle of language pragmatics: the code that gets written is the code that’s easier to write. Consider Java and Haskell.

In Java, defining a class is a monumental task. The language certainly makes it feel that way. You have to create a whole new file just for that class. The fields must be written in quadriplicate: the fields themselves then the constructor, the getters, and setters. equals and hashCode methods must be implemented. The consequence: Java projects have fewer classes than they otherwise would have. Classes tend towards God classes: large classes with multiple incompatible uses and responsibilities. They accumulate, for example, fields that start out as null and become non-null after certain operations are performed on the class.

Whereas in Haskell a type is a few lines of code. You derive the equivalents of equals and hashCode. The consequence: Haskell projects have more types than equivalent Java projects have classes. Those types are narrower, more specific. Instead of mutating fields on the same class, you can have distinct types to represent the different stages that a value goes through as it traverses the codebase. null goes away.

In Java the problem is compounded by the fact that the simulationist school of OOP endorses the idea that each concept in the domain of discourse object should correspond to a class. So while in a Haskell project you may have fiften types to represent a purchase order (representing different stages of validation/enrichment, like the intermediate representations in a compiler), in Java you’d have one omnipresent Order class that is tied to every component in the system. In Haskell the types are anemic and proudly so, in Java this would be called an antipattern.

Shotgun I/O

Summary: ban “quick fixes” at the level of language semantics, or make them hard to write.

By analogy to shotgun parsers, shotgun IO is when effectful code is smeared all over the codebase.

Consider Haskell vs. everything else. Many language communities have some best practice about moving IO to the edges. But there is no strict language-level enforcement of this rule. Because of expedience, because you have to release this tonight or else, someone puts some database access code deep inside otherwise pure data transform code, along with a FIXME comment.

Gradually you go down the slope of pragmatics: I/O is spread across the codebase. In Haskell, there is no way out: you have to use the IO monad. And so if you’re mixing IO and pure code, you know it, because it’s right there in the function signature, and that knowledge gives you an opposite gradient, you know the direction you have to move to make the code cleaner.

Or consider something like the Django ORM: you can write some database access code that returns a QuerySet. Then you can either read from that QuerySet, or pass it on to some other function, which may perform further queries on it.

A consequence of this is that anything that touches a QuerySet can kick off database access. There is no requirement that database access be centralized—in a DAO class, for example—and so because of the pressures of expedience database queries crawl, hack after hack, like hydrogen molecules through the interstitial spaces of a crystal, over the entire codebase.

Tooling

Summary: bad tooling goes unused but is hard to displace, lowering economic productivity.

In C and C++ there are many build systems and package managers, but none is the obvious focal point in the way cargo is for Rust. So people vendor their dependencies, sometimes downloading them by hand. Smaller libraries are often made header-only, this is advetised as a feature because they require less build system overhead to integrate into your projects. This discourages dependencies, which is basically the same as discouraging the division of labour, with obvious costs.

And even bad tooling is hard to replace, again, because of coordination costs and the permanence of focal points.

Consider Common Lisp: ASDF is the standard build system, and Quicklisp the standard package manager. ASDF doesn’t support specifying version bounds, and even if it did, Quicklisp doesn’t support using them. And Quicklisp, despite being over 10 years old, is still in beta and feels like alpha quality software: Quicklisp downloads packages over HTTP, and you are at the mercy of God and the feds. So while the rest of the world has moved on to reproducible builds, determinism, and SAT solving for dependency resolution, in the Common Lisp world you can’t even fix dependency versions.

And who’s going to replace this? You’re going to rewrite everyone’s ASDF system definition files? There’s an alternative to Quicklisp, called CLPM. It’s abandonware. We’ll fix Common Lisp tooling the day after we all move to Mastodon.

And somehow Common Lisp tooling is a pleasure to use compared to the tooling for Python, a language that has multiple orders of magnitude more headcount and investment.

Fixing entrenched tooling requires building an alternative that’s 10x better, making the transition process as seamless as possible, and then reaching out and opening PRs in every repository under the sun making the necessary changes. People could hate the old tooling and it’s still costly for them to change.

Tedium

Summary: programmers will go to great lengths to avoid writing tedious code, to the detriment of other things.

Erik Naggum and the parable of the catapult:

People use C because it feels faster. Like, if you build a catapult strong enough that it can hurl a bathtub with someone crouching inside it from London to New York, it will feel very fast both on take-off and landing, and probably during the ride, too, while a comfortable seat in business class on a transatlantic airliner would probably take less time (except for getting to and from the actual plane, of course, what with all the “security”¹) but you would not feel the speed nearly as much.

The argument here is there’s two kinds of productivity: productivity in the small (immediately) and productivity in the large (when you add up the cost of testing, bugfixes, refactoring, onboarding etc.). C is productive in the small and unproductive in the large.

You observe this with text editing. Moving with the arrow keys, erasing text by holding backspace for an eternity feels slow, but it is predictable and reliable. Vim or Emacs-style editing, where you fly through the buffer with keybindings, feels faster, but a single wrong keypress puts your editor in some unpredictable state it can be hard to recover from. Sometimes the midwit with Notepad gets to the destination faster because they’re not fiddling with their psychotically-optimized Emacs neuralink-mode bindings.

You observe this with types. Dynamic types feel faster, at the REPL, when you’re coding at 1Hz, because you’re not factoring in the (unseen) cost of future bugs, the cost of refactors you won’t do because you don’t have the confidence to refactor which static types give you, the cost of legacy software that can’t be replaced because it can’t be understood by anyone, the cost of a Python server doing four requests per second while you pay five figures to AWS every month.

You observe this with macros. To avoid writing four lines of almost-but-not-quite duplicated code programmers will reach for a macro system. One-tenth of every OCaml or Rust codebase by mass is #[attributes()] to tell the macro system to generate e.g. JSON serialization or pretty-printing code or database interfaces. I do it too. In the grand scheme of things this doesn’t save that much time: tedious code is fast to write but it only feels slow. What is the cost? The cost is the codebase becomes harder and harder to understand, because all of the interesting behaviour is outside the source code.

It’s like those photographs of distant quasars: the “photograph” is the output of a pipeline of Fortran codes written to interpret the data from six different instruments written between 1960 and 2005.

With metaprogramming, your codebase is an input to a vast and unseen system: a pipeline of build-time macroexpansion that spits out something that may very well be unrecognizable. And for what? To save a minute of typing? Because we are programmers, not typists, and it is below our dignity to type?

Some languages do macros right, like Common Lisp, and you should take advantage of them. For essentially every other language, metaprogramming should produce outputs you can check into source control. That is: generated code is fine, but you must be able to see it.

The central, underlying fallacy here is salience bias: measuring what is seen and not what is unseen.