I know I'll likely get a biased answer here but what do you think the cons are of using types in this way? My current thinking is that types should should optimize for an easy learning curve so folks can just get on with writing code. This seems like it would require a deep knowledge of TS.
lodash does come to mind in fact. The last few shops where I've worked use a very small subset of lodash utilities and ignore most of the library. Their documentation is terse so it's often easier to just write your own code that is longer but more readable to the average engineer.
Conversely, and I don't mean to sound flippant, but are there any pros of (ab)using types in this way?
These projects are interesting in an amusing "oh look, another Turing complete type system" way, for the author to gain expertise, and as something to highlight in CVs and to chat about during interviews. But other than that, I would never actually choose to use any of this in a real-world project. Depending on the stability of this API (TPI?), and subjecting colleagues to esoteric libraries would be a nightmare to maintain and support.
The pros of any type-level logic all boil down to automatically enforcing something that you want to be enforced. The key word being _automatically_.
This can be frustrating in the moment sometimes, but other times a type can do two really powerful things:
1. prevent someone you’ve never spoken to from doing something legitimately dangerous
2. provide browsable documentation about the constraint its enforcing and how other places in the code resolve the constraint.
It doesn’t always work out that way, sure, but that’s the ideal types are striving for.
> provide browsable documentation about the constraint its enforcing and how other places in the code resolve the constraint.
It doesn't provide "browsable documentation". It provides complex unreadable types that you have to painstakingly manually "compile" in your head to figure out what the hell is going on
It's a fair characterization because it's a common enough issue to become a problem. It's a compounding of several issues, really, of which "types are documentation" is just one:
Note that there are very few languages that have those kinds of types. Usually it's enums and union types that usually tell you nothing about where to get those ranges, why those ranges exactly, or what those ranges mean.
Literally right now I'm looking at a GraphQL schema that tells me that for a collection of lists of data it returns the following list types: Generic, GenericV2, Onboarding, OnboardingActive... That's it. Actual call returns only a few of those with no apparent difference in actual data.
For errors that can be automatically fixed by IDE, yes. It doesn't make this documentation ;)
I've seen quite a few cases where IDE's type expansion (whether on hover or in type errors) would result in a dozen lines of nested types. Worse than no documentation at all.
A dozen lines of nested types are like a dozen lines of source code, and I generally recommend treating them the same if they aren't legible: poke it with a REPL (in this case, try to declare a variable or new type that should work if your understanding of the types are correct) and see what the tools say.
(And, of course, as with other source code, illegibility implies more documentation needed. For type systems in particular, I think people tend to under-declare their types [i.e. they'll do "Record<KeyType, Array<{name: string, address: string}>" instead of "Record<KeyType, EmployeeRecords>"], which gives the compiler fewer hooks to shortcut talking about a complex type).
There is one thing consistently true about type systems that frustrates me: I write in procedural code all day, but basically all type systems are functional code. It is frustrating to have to spin my reasoning around to attack a problem in a different language structure when I've been doing straight-line statement-by-statement all day. I don't have a solution to that, but I observe the challenge.
> A dozen lines of nested types are like a dozen lines of source code, and I generally recommend treating them the same if they aren't legible: poke it with a REPL (in this case, try to declare a variable or new type that should work if your understanding of the types are correct) and see what the tools say.
None of these actions make types documentation. The necessity to do this makes them a hindrance at best.
> I think people tend to under-declare their types [i.e. they'll do "Record<KeyType, Array<{name: string, address: string}>" instead of "Record<KeyType, EmployeeRecords>"
By the same argument, code isn't documentation. And yet it is; all manner of style guides will tell us "Document the non-obvious, but let the code speak for itself when it's obvious." The challenge, of course, is obviousness is subjective. And I've definitely encountered type stacks where the author thought they were being perfectly clear and the reader didn't. But that makes it no different from other forms of docs.
I think the truth is types (and code) are a little of both documentation and implementation. After all, if code alone were sufficient, we'd just be hand-rolling 1 and 0 patterns to tickle the CPU directly; the existence of programming languages themselves are to bridge the gap between the minimum information a computer needs and human comprehension of what we're trying to get the computer to do.
(Whether types are docs at all, let me put it this way: every modern style guide I've found for languages where strong static typing isn't available declares I have to put the types in the documentation, especially for function calls and the like. That seems to heavily imply that types are a form of auto-checked documentation).
Can you clarify your belief that code isn't documentation? I'm far more familiar with the alternate philosophy that code is just another way humans communicate with each other (the computer doesn't care; the computer can take a machine-compatible binary from anywhere and execute it with no care as to whether humans understand the how or why of the program). Consider the "Literate programming" approach and tooling that operates in that paradigm (such as Jupyter notebooks).
Comments clearly aren't code, in the sense that they're a specific form in a language dedicated to being thrown out by the compiler / interpreter / etc (usually; I'm handwaving over the metaprogramming approaches like Doxygen that parse comments). But there's a reason we name variables `time_to_completion` instead of just `a`, or name functions (or have functions instead of jump pointers crammed into global variables), etc.
Documentation: "is any communicable material that is used to describe, explain or instruct regarding some attributes of an object, system or procedure, such as its parts, assembly, installation, maintenance and use."
That's a pretty [good definition I think](https://en.wikipedia.org/wiki/Documentation). I also think a type does provide explanation and instructions about how to use typed code. So, I think by definition, types are undeniably documentation.
But projects like these go far beyond just constraints. They're a library of utility functions for doing arithmetic, working with lists, and even sorting. All of this can be equally accomplished by a regular and well-established library that doesn't depend on exotic, and possibly undocumented, features of a type system, and without incurring build time penalties.
If you push all your code into the type system, you're basically just writing your program in a clunkier, less readable language. So, you will just end up having to debug your type system, which defeats the whole purpose. I think expressive type systems are great, but like most things in engineering, it's a trade-off; the complexity of having to manually manage invariants vs. the complexity of managing the type system.
The correct balance depends on how critical the piece of code you are working on is, and how much you want to write unit tests for it.
In cases where the type system can encode the invariants you care about, it can be worth the extra complexity so that you have code that has compiler enforced correctness (at least in some aspects). Of course, you can't do this for any invariant in a given type system (eg lifetimes are not representable in Java). When it is possible though, debugging a type system has the advantage that it's done on your local machine, away from prod.
lodash does come to mind in fact. The last few shops where I've worked use a very small subset of lodash utilities and ignore most of the library. Their documentation is terse so it's often easier to just write your own code that is longer but more readable to the average engineer.