Forall.js: Idris-like static types and invariant checks for JavaScript | Hacker News Reader

Forall.js: Idris-like static types and invariant checks for JavaScript

github.com

63 pointsLightMachine17 comments

I would avoid using the terms "type" and "check" together for this, since "type checking" has a well established meaning (referred to as the "classical"/"logical" approach in this README).

Random testing like this, as WaxProlix says, is reminiscent of QuickCheck; in that case you might replace "type" with "property", since the phrase "property checking" is associated with this approach.

Alternatively, stick with "type" (or maybe "contract"? These terms are all very overloaded!), but use "testing" instead of "checking".

Note that this approach is should be possible in Idris too, since we have equality types. For example, assuming we have some QuickCheck-style functions available, I could say:

    someProp : Foo -> Bar -> Bool
    someProp x y = whateverExpressionWeLike

    somePropCheck : quickCheck someProp = True
    somePropCheck = refl

This will only type-check (in the traditional sense) if `quickCheck someProp` evaluates to `True`; otherwise `somePropCheck` will be ill-typed. `someProp` and `somePropCheck` have no impact at runtime, since they're never called from anywhere (they could be encapsulated privately in a module too).

`quickCheck` can be a normal function; it's simple enough to generate "random" values in a pure way (e.g. Mersenne Twister with some arbitrary seed), although that would lose one of the features of property checking: each time it runs, we explore more of the possible inputs, and hence gain more confidence. Since runtime guarantees don't depend on the particular values generated by `quickCheck`, it's still "morally pure" even if implemented in an impure way; hence we could use a mechanism similar to "type providers" to impurely generate random values, and hence regain that feature.

I agree with everything you said.

On doing that in Idris (which is my dream language - the only reason I don't use it is its JS compiler is too slow), would it be possible to randomly generate elements of any data type, or would we have to write an `Arbitrary` instance for every type you want to test? I don't know/remember much about generics on Idris, would love to learn more.

(Seriously, though, why is the JS compiler still so slow? Again - I'm willing to put substantial money into funding a better one...)

Hmm, I imagine there's probably a generic way to do it; but I was imagining a straight port of the Haskell version. Note that it might be difficult to generate values of some heavily-constrained dependent type; you'd effectively be implementing a proof search!

You probably could, but part of the purpose of Idris is to statically prove things with dependent types rather than test them.

Testing and proving aren't mutually exclusive; there's also a more fine-grained distinction, between proving things (say, with pen and paper, or an SMT solver, or whatever) and proving things using the type system of the language they're implemented in.

The nice thing about Idris is that it's specifically created with such "practical" issues in mind: it gives you powerful tools, but there's absolutely no obligation to use them. If you want to, you can avoid checking certain (or all) parts of a program. You need to define enough types to represent your computation/data, but not to prove its correctness; e.g. you can over-represent your domain using a bunch of sum types, stick infinite loops in the branches you know/think are unreachable, then turn off totality checking.

If we compare this to, say, Coq or Agda, not only must you prove everything-you-write-in-those-languages, you must also prove everything-you-write in those languages! (i.e. everything you write must be proved, and only proofs written in Coq/Agda will be accepted). You can assert axioms, but not their computational/runtime meaning, which is problematic when using them as a programming language.

There's a fair bit of work in this area that leverages Flow type syntax, e.g:

- https://github.com/gcanti/tcomb

- https://github.com/gcanti/io-ts (for typescript)

- https://github.com/codemix/flow-runtime (i wrote this)

And another alternative is to use Design by Contract, like Eiffel:

https://github.com/codemix/babel-plugin-contracts (also me)

Oh. Thank you very much for linking that. I wish I saw it 3 days ago, though. It is basically the same thing.

Closure Compiler already does this. Most JavaScript at Google is written with static types in JSDoc. Closure Compiler also has a DSL (sort of like template metaprogramming) where complicated generic type transitions can be defined with functional expressions.

Examples: https://github.com/google/closure-compiler/blob/d1c1673d1d89... and https://github.com/google/closure-compiler/blob/55cf43ee31e8...

Documentation: https://github.com/google/closure-compiler/blob/e7954defc514... and https://github.com/google/closure-compiler/blob/b8d8d1d6c6d2... and https://github.com/google/closure-compiler/blob/8829c165b30f...

Welcome to what happens when a company hires PhDs to write web pages.

> the catch is that it isn't doing classical (logical) type-checking, but, instead, it "type-checks" by inspecting with random samples.

So it's almost nothing like Idris, which is geared towards formal verification.

All of these unsound type systems make development and refactoring easier, but you don't really know what type checks are being done at a lower level.

TS* has (had) an untrusted type for input from potentially hostile adversaries. I would have preferred to do high-assurance client side code development in that, but it's no longer maintained.

This is really not like the Idris type system at all. Its not even really like the Haskell type system. This is basically doing the same thing as the Haskell quickcheck library but running it at compile time. AKA just testing a function with a few values.

If it quacks like a type and it types like a type...

Just kidding! A type system, in essence, is not much different from that, though - Pi, for example, is basically a type-level lambda that is applied to the typed function (and it is sufficient to express basically any real-world type system). In any case, I hope I made it very clear on the that it is in no means an actual type system, but my dirty way to get invariants and some refined types on plain JS. If we had a mature Idris->JS compiler this hack wouldn't be necessary. ):

Not trying to be a downer on the creator, because it's very cool, but I think the best thing that can come from libraries like this is adoption into Typescript. Typing is really better suited to being a part of the language and not a facet of libraries (or is it - an interesting thought there, the ability to import type systems).

that's (kind've) what facebooks flow tool does (in that you type header at top of the file to turn on type checking)

You might be interested in http://blog.racket-lang.org/2017/04/type-tailoring.html, which argues for customised type checking.

I'm confused. Does this actually statically checks, or is this testing?

Like, does it tag everything with type info and walk the code statically without executing it and checks that all connections respect their types?

It seems to me it does that for non dependent types, and then it does generative testing for the dependent types? And can also add runtime contracts for them?

So where is the famous Idris elaborator? Where are the universes? Where’s inductive types? Where are they?

An alternative that is much easier to adopt, is to use unexpected-check http://unexpected.js.org/unexpected-check/

Reply on news.ycombinator.com