> If you don't trust a type system, it's hard to justify using it at all, apart from the documentation it provides to other devs.

100% agree, although I have to wonder, why don't you see that Typescript is NOT a type system you can depend on.

Fact of the matter is that Typescript has serious soundness issues and will always have. Those problems haven't disappeared.

What this translates to in practice is that Typescript annotations are little more than documentation. And it simply doesn't compare with any sane static language where the first line of defense is the compiler.

What is the practical benefit of soundness? I offer that soundness is a disease that makes programming languages far more complex than they have any right to be.

A static type system exists to prevent some classes of bugs and enable code completion.

However whenever "soundness" becomes a design goal, it enevitably usurps everything else (because it is an objective binary goal) and leads to languages that cease to value other things like simplicity or directness.

A type system can never catch all bugs and reliable programs ultimately depend on a human being understanding and verifying the code. The most important goal - if you value reliable programs - is that the code is intelligible to people.

> A type system can never catch all bugs

True, but that doesn't mean we shouldn't try to improve the type system. We've already seen large improvements in the past in TypeScript and flow. For example, nullable types were not there in the beginning but were added to the type system because they are immensely useful. Making the type system sound would be another such improvement.

If you start with that premise, why not "A type system can never catch all bugs, therefore let's not use it at all"? Sounds stupid, doesn't it?

> The most important goal - if you value reliable programs - is that the code is intelligible to people.

The code that's running the airplane I'm sitting in, I don't give a fuck whether it's intelligible. Though I do care very much that it has been formally proven to be correct. Having a sound type system is pretty much a prerequisite for formal proofs.

Also, intelligible code and soundness are not exclusive. What makes you think that? I can write unreadable looking code in JavaScript and Haskell. And I can write beautiful, readable, concise code in both languages, too.

I'm not convinced that formal provability as a goal for a language generally used to write web front-ends makes a lot of sense.

The "gaps" in TypeScript's type system are actually escape hatches which allow clean-ish interoperability with the code that people in the real world actually use. Like, dammit, JQuery and all the other filthy things that make web pages go. That code is already written and no one is going to change it to make the TypeScript typings "work". Such code by the way is usually incompletely (or even incompetently!) typed but perfectly servicable in practice.

> and leads to languages that cease to value other things like simplicity

Simplicity is subjective. Some people think that TypeScript is overly complex compared to simply using JavaScript. Other people feel like Scala's type system doesn't go far enough.

I'd say that in a type system, what is "simple" is what you're used to and understand, and what's "complicated" is what you don't.

Now, friction is another story, and some type systems are very friction heavy (Hello Java!), while some very sound type systems are not (Elm). While Flow has issues/bugs and an extreme lack of proper tooling, the type system itself arguably has less friction than TS. If the implementation was on par (that is, you changed nothing to the type system at all, but just improved documentation, tooling and stability), I'd be able to get my programs in a better state faster, with less typing and catching more bugs.

Unfortunately that isn't the case, and in the last couple of releases, TypeScript at least got decent enough to use.

Reading from the comments here, I believe that people don't realize how bad the unsoundness is in Typescript:

    class Dog { name: String }
    class Engine { name: String }

    function start(ref: Engine) {
      if (!(ref instanceof Engine)) 
        throw Error("oops")
    }

    // Typescript does not complain
    start(new Dog())

This sample is due to Typescript doing "structural typing". According to its rules, if Engine and Dog both have the "name" property, then they must have the same behaviour as well. This is also inconsistent with the behaviour of classes in Javascript (e.g. see the behaviour of `instanceof`).

The biggest problem is that generics are what they call "bivariant", which is lingo for "completely fucked up". Here's a classic gotcha that caught people by surprise with Java's array since forever:

    class Animal { isAnimal: Boolean = true }
    class Dog extends Animal { isDog: Boolean = true}
    class Cat extends Animal { isCat: Boolean = true }

    const dogs: Dog[] = [ new Dog() ]
    const animals: Animal[] = dogs
    animals.push(new Cat)

    for (const dog of dogs) console.log(dog.isDog)
    //=> true
    //=> undefined

Well, it's unfair to pick on this, since they give a similar example in their own documentation, right? But this extends to inheritance as well, because input parameters in functions are not contravariant in Typescript:

    abstract class FlatMap<A> {
      abstract flatMap<B>(f: (a:A) => FlatMap<B>): FlatMap<B>
    }

    class Box<A> extends FlatMap<A> {
      flatMap<B>(f: (a: A) => Box<B>): Box<B> {
        throw new Error("Not implemented!")
      }
    }

It should be obvious to everybody why this is wrong:

    class AnotherBox<A> extends FlatMap<A> { ... }

    const box: FlatMap<number> = new Box<number>()
    box.flatMap(x => new AnotherBox<number>())

In case you're wondering, that interface is the beginning of the Monad pattern and the problem here is that monads are incompatible with each other. So in order to compose instances together, you do rely on the compiler to protect you. Think Promises being combined with Arrays, both of which are monadic types. Not going to work, is it?

In less expressive static language lacking higher kinded types, which you need to express such interfaces, there's a current trick that people do to work around it: https://www.cl.cam.ac.uk/~jdy22/papers/lightweight-higher-ki... ; This has been used in Elm and it's being used in Kotlin as well, see for example: https://github.com/FineCinnamon/Katz

But Typescript is on a whole new level of wrong, because you can't trust the compiler for correctness, so you can't trust it for protection, at all. You see, this is not pie in the sky academics, but actual code that can happen in your app, mistakes that the compiler should have prevented you from making. In the above case Typescript really is just documentation and just like documentation in many cases, it can be wrong documentation. To make this even more frustrating, Microsoft's other language, C#, does not have the problems that I enumerated.

And I know what people say - oh, this is still useful. Well, guess what, writing JSDoc + having Google Closure to validate is actually more useful, while not being non-standard ;-)

You're treating this specific case as a fundamental issue, when it's a specific and superficial one that can be reasonably addressed in a future TS release:

See for example:

- Nominal types proposal https://github.com/Microsoft/TypeScript/issues/202

- Co/contravariance annotations proposal https://github.com/Microsoft/TypeScript/issues/10717

These choices are fundamental to the language as it was designed. There's nothing more fundamental than how variance works.

Want to bet that in 2 years from now we are still going to talk about the same language with the same issues, and that those two proposals are going to be rejected?

I'm not sure I understand why you think that. TypeScript explicitly prefers real world usability to theoretical soundness, so if the community is vocal about the need for these features, I don't see why TS would turn them down.

> real world usability to theoretical soundness

The soundness we are talking about in this case isn't theoretical at all and in fact such an argument is anti-intellectual. This is anti-intellectualism.

As for the "community", we are talking about one where the majority has never worked in sane static languages.

You are misunderstanding what I said.

Either one of the 2 issues I linked would fix the specific problems you outlined. You don't need a globally sound type system, H-M inference, or anything like that to fix those specific problems; local inference and opt-in nominal typing would work just as well.

Having used TS daily for a few years, I can vouch that your cases are pathological, and don't come up in practice often, if at all.

Even if they enable it, it might be through a switch like strict nulls (which was a huge mistake to implement so late).

That's not great.

I agree they waited too long for strict nulls!

However, both features I linked at currently proposed seem to be backwards compatible, so hopefully no switch.

For sure that's great. The issue are still:

- it's not there today - it still has to be bolted on (eg: adding new keywords, as opposed to, let say, Flow automatically doing nominal typing for Classes, which was the right move). - The ecosystem did not get to take advantage of it for years, so existing types, libraries, etc, are not written that way. That's a lot of tech debt.

That's why new languages keep popping up all the time. The later you bring in critical features, the more bloated you force the system to be.

I will admit TypeScript has been doing a decent job at containing the bloat (most of the crap features were actually what they implemented super early on, like the namespaces, enums, etc...the new stuff is really cool). But I think having proper competition enabled that: see the original strict null issue, where people were all but saying it was not possible/practical. Good thing we had Flow as a counterargument. Now with Flow dying a slow death, we'll have less fragmentation (great!), but less competition (possibly stagnating language. Boo!)

I agree on both points. An interesting facet of your 1st point is Stroustrup's Rule [0], which gives more context for an explicit `nominal` keyword.

[0] https://thefeedbackloop.xyz/stroustrups-rule-and-layering-ov...

Yeah I can totally see that. I liked the classes are nominal, plain objects are structural because its close to what people expect without even knowing the difference.

But yes, TypeScript was very much designed for a target audience (all of the C#/Java-style class features and namespace stuff early on, which less and less people want as time go and people learn about better ways of doing things).

It seems inevitable to me that eventually (might take decades), Haskell-level type systems will be the norm and people will learn them in highschool. The world is just not ready (and neither am I)

> It seems inevitable to me that eventually (might take decades), Haskell-level type systems will be the norm

That would be exciting, but I can't help but wonder if FP is just a matter of taste. We had Smalltalk and ML in the 70s, C++ and Erlang in the 80s, Java and Haskell in the 90s, and now it feels like all the best languages at least nod to a multi-paradigm approach: Scala, Swift, Python, JavaScript, C#, Rust, etc. The right tool for the right job, you know?

> and people will learn them in highschool

I'm actually working on a book along these lines! Look for it on HN in a few months :)

> Typescript annotations are little more than documentation

This is true at the boundaries of your codebase, but if your application is complex enough, tsc is a life saver. Typescript saves me from bugs literally every day. Not to mention the benefits around tooling and refactoring.

Agreed that there are some big footguns with respect to soundness, though.

Nit: TypeScript predated flow by 2 years.