Types Are for People, not Computers

Types—in the static-typing sense—are useful because they help people, not computers. Oh sure, we use them, in part, to subdue the compiler or meet some need peculiarly arising from our computer. But types are valuable because they are a way of communicating.

Type systems are a way of communicating.

Type systems are a way of announcing what you understand, expect, or intend. Good type systems let you do so at the level of abstraction you choose. And they let you describe these things based on what things are or what they do.

So if types are a way of communicating, who or what are you talking to? Several audiences. You are talking to yourself—leaving notes. You are also talking to the computer—but often for your sake rather than its. You are asking it to hold you accountable to what you told it you understood, expected, or intended. Finally, you are talking to other developers (including future you).

We can match how we use the type system with what we understand, expect, or intend of our objects.

Consider this JavaScript code:

const mathItForMe = (a, b) => {
  return { added: a + b, subtracted: a - b, multiplied: a * b, divided: a / b };

mathItForMe(5, "3");
// => { added: '53', subtracted: 2, multiplied: 15, divided: 1.6666666666666667 }

mathItForMe(true, 7);
// => { added: 8, subtracted: -6, multiplied: 7, divided: 0.14285714285714285 }

mathItForMe("3", true);
// => { added: '3true', subtracted: 2, multiplied: 3, divided: 3 }

In what programming context is the desired behavior for using + with a string and a boolean to coerce the boolean into the string true or false and concatenate it to the string? It’s madness. But then again, all implicit coercion is the same kind of insanity; some instances are just more florid than others. Using types imposes some discipline:

#[derive(Debug, PartialEq)]
struct GotMathed {
    added: i32,
    subtracted: i32,
    divided: i32,
    multiplied: i32,

fn math_it_for_me(num_1: i32, num_2: i32) -> GotMathed {
    GotMathed {
        added: num_1 + num_2,
        subtracted: num_1 - num_2,
        divided: num_1 / num_2,
        multiplied: num_1 * num_2,

fn main() {
    let x: i32 = 7;
    let y: i32 = 5;

        math_it_for_me(x, y),
        GotMathed {
            added: 12,
            subtracted: 2,
            divided: 1,
            multiplied: 35,

But notice that we aren’t necessarily capturing our intent. What if, instead of a signed 32-bit integer (an i32), we use an unsigned, 64-bit integer (a u64)? Or a 32-bit float (an f32)? Our function doesn’t work—but it probably should. This function isn’t about doing math on 32-bit integers, it’s about doing math on numbers.

The type system has helped us figure out what our function ought to do by requiring us to describe it explicitly. Here we’ve gone too narrow. This is worlds better than a function that takes types we didn’t even consider, like the JavaScript example. The behavior there was pathological because we were too accepting.

But now the type system is putting pressure on us to think about exactly what this function ought to do. If it were part of a private API—a private method in a class or a non pub function in an impl block in Rust—we might accept that it places arbitrary limits on the types it accepts in order to streamline implementation or avoid premature optimization. Plus, we would have additional context and the method’s restricted visibility means we wouldn’t be beholden to others who rely on the code.

use core::ops::{Add, Div, Mul, Sub};

#[derive(Debug, PartialEq)]
struct GotMathed<T> {
    added: T,
    subtracted: T,
    divided: T,
    multiplied: T,

fn math_it_for_me<T>(num_1: T, num_2: T) -> GotMathed<T>
    T: Add<Output = T>    // Specify "trait bounds": A trait is like an
        + Sub<Output = T> // interface. It specifies the behavior (method
        + Mul<Output = T> // signatures) a type must implement. Here we're
        + Div<Output = T> // saying `T` is a type that must implement addition,
        + Copy            // subtraction, multiplication, and division, in each
        + Clone           // case returning the same type. It must also
        + PartialEq,      // allow the value to be copied and cloned, and
{                         // compared for equality (used by `assert_eq!` ↓).
    GotMathed {
        added: num_1 + num_2,
        subtracted: num_1 - num_2,
        divided: num_1 / num_2,
        multiplied: num_1 * num_2,

fn main() {
    let x: f32 = 7.0;
    let y: f32 = 5.0;

        math_it_for_me(x, y),
        GotMathed {
            added: 12.0,
            subtracted: 2.0,
            divided: 1.4,
            multiplied: 35.0,

Here we have brought our intent, expectations, and understanding into alignment. We are communicating to the computer what our function needs to be able to do and how we intend to do it; the computer will now make sure that we are able to do what we promised with the types our function will accept. We have put ourselves in a position where the computer can help us. We are also communicating to other people what our function is for based on what kinds of arguments are acceptable and what they can expect it to return.

Note that we are now describing what kinds of arguments are appropriate in precisely the terms that make sense. We’re no longer talking about arbitrary concretions: we are talking about the arguments at exactly the level of abstraction that ought to define them.1

We have specified the arguments that math_it_for_me takes in terms of behavior, and in so doing have aligned our intent with our understanding and expectations. We could alter the code so that num_1 and num_2 could be different numeric types, but that would force us to consider edge cases and typecasts, and so it is reasonable to limit our code the way we have, out to the limit of undefined or tricky behavior that we’re not interested in dealing with. And all of these positive changes were driven by the type system’s pressure to define things explicitly, together with its allowing us to describe acceptable arguments at the right level of abstraction.

Good type systems remind us to think about things we might otherwise forget.

Type systems with Option/Maybe and Result/Either are a good example of how type systems can force us to handle all possible results. For example, Rust defines the Option type:

enum Option<T> {
    // Enums in Rust can contain values. Here the `Some` variant contains a
    // value of the generic type `T`.

There is no null in Rust. A function that can fail to return a value of type T must return an Option<T> (or similar) rather than a T. If you write a function that is supposed to return a T but there is any code path in the function that would fail to return a T, the compiler will yell at you for not satisfying the function’s type signature. If the caller of that function tries to treat its return type like a T rather than an Option<T>, you will likely get a type error.2

So unless you follow the antipattern of calling unwrap() or expect() with your Option type,3 the compiler will hound you until you have dealt with the null case. Libertine Rubyists and Pythonistas may object and say that they always handle their nils. But that’s not what the runtime logs say for any of the applications I’ve worked on in any of my programming jobs.4

Specifying types in terms of interfaces, traits, or typeclasses is the same thing as duck typing, except it’s harder to screw up.

In dynamically typed languages, duck types

are abstract; this gives them strength as a design tool but this very abstraction makes the duck type less than obvious in code. [¶] When you create duck types you must document and test their public interfaces. Fortunately, good tests are the best documentation, so you are already halfway done; you need only write the tests.5

Pace the brilliant Sandi Metz, types are better documentation than tests, at least for the kinds of things types help us with. Unlike tests, types are nearby, facilitate our tooling, and provide a consistent means of communication. Tests and types are both better than documentation, but types are documentation attached to our code like PostIt notes, while tests are like books down in the stacks that we have to go looking for, hoping that the card catalog will help us find them.

Elsewhere, Metz writes,6

[I]n my code, I don’t get run-time type errors. Because my experience is that dynamic typing is perfectly safe, I find myself resenting having to enter type annotations when I work in statically typed languages. While I appreciate the fact that static typing means that I don’t have to write some tests, I hate having to add this extra code.

Having said that, I realize that many folks have had different experiences with dynamic typing. I write trustworthy code where objects behave like you’d expect. This means that I can trust that any object with which I’m interacting just works. This, in turn, means that I don’t have to check if objects behave the right way. Sadly, I’ve seen many OO applications where these things were not true. Folks fall into the trap of writing code that’s not trustworthy. Because they can’t trust message sends to return objects that behave correctly, they have to check the type of the return of messages sends. This leads to a descending spiral of manually adding type checking, and code which ultimately breaks in confusing and painful ways.

I’m not as good a programmer as Sandi Metz is, and you probably aren’t, either. It may be like the case of the great athlete who can’t coach inferior talents because she didn’t have to learn to work with the same limitations.

But even if your code is Metzian in its brilliance, implicit interfaces are like verbal contracts: a fine way to get confused about what has been promised. Worse, the absence of a rigorous and explicit definition for your interface means your computer is powerless to help you get it right. Remember, types are for people.

Types are not a panacea.

There are some limitations to type systems, and I should acknowledge them to show that I am not enamored of a caricature of types. The problems are real, but they are worth it.

Sometimes types aren’t for people; sometimes they are about the programmer helping the computer.

In Rust, if you have a function that can return more than one error type, you must in some cases put the error in a Box. That is because different errors are different sizes, and the return type of a Rust function must be of a size known at compile-time. So the indirection of a Box (which is a pointer) means that the compiler knows how big the return value is. You have to help the compiler know what’s going on—you’re helping the computer rather than the reverse.

This is annoying. But even Ruby and Python aren’t written in English sentences. Semicolons and curly braces are mostly there to help the computer, too. Just because types are sometimes for computers doesn’t mean than types aren’t mostly useful for people.7

Types can’t help you with all your business logic.

I once used a u8 (an 8-bit, unsigned integer) to represent people’s ages. Because people haven’t lived past 122, 255 seemed like plenty of room. Fair enough. But I also saw it as a way of embedding business logic into the type. It just so happens that 255 is not an insane limit for validating a person’s age. If you input -50 or 3000, you have probably made a mistake.

That got me thinking: What if I could somehow define a 7-bit unsigned integer. Then I could use the type system to do a kind of validation.

I now see that this was colossally wrongheaded. This is mixing levels of abstraction in the wrong way. It’s like noticing that motor oil and coffee are about the same color, so they should be interchangeable to keep your car running and to perk you up in the morning.

Types are not for fine-grained validation. They’re not a replacement for tests.

But when you use them for their proper purposes, types can help you talk to yourself, your computer, and other people; they can catch certain kinds of errors; and they can put pressure on you to think, design, and communicate more clearly. That’s enough to make them a great tool that most programmers should use.

  1. Cf. the Dependency Inversion Principle.] ↩︎

  2. If you just so happened to treat your Option<T> in a way that is fully compatible with T, the type system would not catch it. ↩︎

  3. If you call either with a None type, you get a panic, which is like an unhandled exception. ↩︎

  4. Rich Hickey doesn’t like this pattern, and TypeScript handles nullability a different way, but at bottom any alternative approach should require handling the null case. ↩︎

  5. Sandi Metz, Practical Object-oriented Design in Ruby: An Agile Primer 98 (2013). ↩︎

  6. I’m Sandi Metz, Ask Me Anything!, (Jan. 24, 2018). ↩︎

  7. The title “Types Are Mostly for People, Though Occasionally for Computers, Which can Be a Pain” seemed less punchy. ↩︎