🔗 Dynamic type systems aren’t even simpler

Alexis King just published a great blog post titled “No, dynamic type systems are not inherently more open”.

That reminded me of the talk I gave last year at FOSDEM, titled “Minimalism versus types”, where I advocated for static types from a slightly different angle. I tried to convince people that types in dynamically-typed programs are often more complicated than people realize. And often more complicated than in typical statically-typed languages.

People often say the opposite, that static type systems are more complicated, and dynamically-typed languages are simpler. At the surface level, this seems true: in a dynamic world you just go merrily declaring variables, assigning values and doing things with them, without ever having to write down any types, no matter how trivial or complex they are. Things can’t get any simpler in the typing department than “doing nothing”, right?

Well, types are nothing more than the shapes and allowed behaviors of your data. So it’s not like you don’t have shapes and behaviors in any program, regardless of the language… so, you have types, whether you write them or not. They are there, even in assembly language, even if at a conceptual level, as the sets of “valid values” your program can manipulate. And you have to think about them, and they have to make sense, and they have to do the right thing. So, in short, in a correct dynamically-typed program the types need to be just as correct as they are in a statically-typed one (or else you’ll get runtime errors).

In other words, the types are there, but you have to run the type checker in your head. And you know what’s funny? When people don’t write down the types, they often end up with types that are often more complicated than the types from people who do write them. The complexity just goes under the radar, so it piles up.

One day you open that module which you haven’t touched in six months, and you see a function call where the third argument is null. You need to remember what kinds of variables you can pass to that third argument, or read the whole source code to figure it out. You follow through the code to see all places that third argument is used and realize the accepted type of the third argument depends on what you give to the second argument. Congratulations, you’re dealing with a dependent type, which means you’ve just surpassed Haskell in terms of type system complexity. Compilers that deal with this kind of type system are so complex they are effectively proof assistants (and are at the forefront of programming language research), and here you are dealing with those data types with your brain (and your faith in your ability to come up with sufficient tests) alone.

Given that there is no mechanical type checker to prescribe what is expressible, and that the dynamic runtime will accept anything as long as the program doesn’t crash, when doing typechecking in your head you essentially have the world’s most powerful and complicated type checker at your disposal. And once you start making use of this power, you end up dealing with the world’s most complicated type system.

And when you give people expressive power, they use it. In my experience, people go wild constructing complicated structures in dynamic languages that they normally wouldn’t in static languages. It’s not that static languages are less powerful (Turing equivalence, blah blah), but they make the things you’re doing more obvious to you (Alexis’s post has some great examples). In a dynamically-typed program people are all to keen to make variables and data structures perform double or triple duty (“this is a X but also a Y under circumstances Z”), but when they have to write down what they’re doing as types, it’s like a little conscience check, and they think twice before creating a more complex type for something that could be described in a simpler way (simple example: they’ll probably make two plain functions instead of making one function that takes a string argument that changes the behavior of other arguments). Static types nudge you towards simpler, less “clever” solutions (and we all know what kind of solution is more maintainable in the long run).

But okay, let’s assume we avoid “clever” and pick the same solutions in either. Writing the same program in a static or a dynamic language to process the same data in the same way, you will end up with values of roughly the same types in both. The fact that the variables have static types or not does not change that.

“But in a dynamic language I don’t have to write the types! It’s less work!”

“Not having to” write types but having to think about them anyway is like doing math “not having to” write anything down and doing all calculations in your head. How is it an advantage to not use pen and paper to track down your train of thought as you do a complex calculation, and instead be restricted to do it only in your head? And how is an advantage to not have a mechanical tool — like a calculator, which can actually compute the things you wrote down — to check whether what you wrote with pen and paper makes sense?

I’m lazy, so I hate doing any math in my head. I’ll take writing things down and have a machine check it for me any day of the week. Why wouldn’t I want the same when programming? That’s what computers are for, right? To save us from computing things in our head. So I’ll write my types, and have the compiler check whether they make sense, thank you very much. It’s less work.