lkitching

I've already shown that type hints do not constitute type checking:

  (defn f [^String s] (.length s))
  (f 3)

is a valid Clojure program that fails at runtime with a cast error.

  class X { public static int f(String s) { return s.length(); } }
  X.f(3)

is not a valid Java program at all. Clojure compilation generates bytecode to dispatch dynamically and all but the most basic checks are handled at runtime by the JVM. This is fundamentally different to the static type checking that languages like Java and Scala do. It's not that Clojure is hiding something from Java, but rather that it isn't doing the considerable amount of effort the Java type checker does to analyse the program before execution. This is by design - Clojure has deliberately avoided adding a static type system in favour of things like spec.

Of course Clojure has to ultimately be compiled into a native format for the host platform, bytecode in the case of the JVM implementation, but that doesn't require type checking in the same way Java does.

Clojure functions are compiled into implementations of clojure.lang.IFn - you can see from https://clojure.github.io/clojure/javadoc/clojure/lang/IFn.h... that this interface simply has a number of overloads of an invoke method taking variable numbers of Object parameters. Since all values can be converted to Object, either directly for reference types or via a boxing conversion, no type checking is required to dispatch a call. With a form like

  (some-fn 1, "abc", (Object.))

the some-fn symbol is resolved in the current context (to a Var for functions defined with defn), the result is cast (not checked!) to an instance of IFn and the call to the method with required arity is bound. This can go wrong in multiple ways: the some-fn symbol cannot be resolved, the bound object doesn't implement IFn, the bound IFn doesn't support the number of supplied arguments, the arguments are not of the expected type. Clojure doesn't check any of these, whereas the corresponding Java code would.

Protocol methods just get compiled into an implementation of IFn which searches for the implementation to dispatch to based on the runtime type of the first argument, so it doesn't introduce static type checking in any way.

Protocols do not work like Java interfaces or classes. Their methods are compiled into regular functions which lookup the implementation to use at runtime based on the runtime type of the receiver. Compilation will check for the named function but doesn't do any further checking. Given the following protocol and implementation:

  (defprotocol P
    (method [this ^Integer i]))

  (extend-protocol P
    String
    (method [s i] (.substring s i)))

both (method "test" "call") and (method 1 2) will be accepted by the compilation phase but will fail at runtime.

Of course there's no requirement for Clojure code to be AOT compiled anyway so in that case any name errors will still only be caught at runtime when the compilation happens.

Type hinted bindings are only converted into a cast and are not checked at compilation time either e.g.

  (defn hinted [^String s] (.length s))
  (hinted 3)

will be accepted but fail at runtime.

deftype is only used for Java interop an is also not a form of type checking. The methods will be compiled into Java classes and interfaces, but the implementations defer to regular Clojure functions which are not type checked. You can only make use of the type information by referencing the compiled class files in Java or another statically typed language, using them from Clojure will not perform type checking.

Neither defprotocol nor deftype introduce static typing into Clojure. Errors in their usage are not checked statically and are only discovered at runtime.

That just means the semantics of the language are defined by whatever the default implementation does. It's a big stretch to conclude that means Rust 'was' OCaml in some sense when the compiler was written with it. Especially now the Rust compiler is written in Rust itself.

I'm not convinced the implementation language of the compiler counts as a feature of the Rust language. If the argument is that Rust wouldn't have been invented without the original author wanting a 'systems OCaml' then fine. But it's possible Rust would still look similar to how it does now in a counterfactual world where the original inspiration was Haskell rather than OCaml, but removing the Haskell influence from Rust as it is now would result in something quite different.

Which OCaml features exist in Rust but not Haskell? The trait system looks very similar to Haskell typeclasses, but I'm not aware of any novel OCaml influence on the language.

The way in which monads are monoids isn't really helpful for understanding them for progamming.

Since this defines an interface, does this solve the null problem? e.g.

    Maybe<String> foo = null;
    foo.map(s -> "bar");

explicit pattern matching is usually discouraged anyway though, and you can do this now with Optional

    Optional<String> foo = Optional.empty();
    String message = foo.map(s -> "Hello " + s).orElse("Fine, leave me hanging");

Other languages like Scala also have an Option.fold method for this specific case.

How is this related to Brexit? Ireland have had a low corporation tax rate for the last 20 years, long before the Brexit vote.

Assuming you're referring to *1, *2, *3 and *e, these are only defined within the REPL and are never used in real programs.

The OP is contrasting between a 'validation' function with type e.g.

    validateEmail :: String -> IO ()

and a 'parsing' function

    validateEmail :: String -> Either EmailError ValidEmail

The property encoded by the ValidEmail type is available throughout the rest of the program, which is not the case if you only validate.

The post is suggesting that parsing and validation are different things, since the output of a parser captures the properties being checked in the type, and validation does not. Downstream consumers of validated input cannot rely on the properties that were validated since the representation type doesn't encode them e.g. the non-emptiness of a list.

Arguably the resource is the dataset of stock exchanges, and the CSV representation is forced to omit all the metadata but the HTML representation isn't.

There's been a fair amount of churn in the ecosystem, even if the language itself has been very stable. Leiningen was ubiquitous 5-6 years ago, but if you switched to using deps for dependency management you lost the ability to run tests or build uberjars and had to manually re-create these on a per-project basis. Now there's tools.build, but that also requires you to manually write essentially the same tasks for basic functionality in each project. Leiningen and tools.deps also seem to resolve depedencies differently, so you can run into issues simply by migrating from one to the other.

All those links are to the first game, not Wipeout 3.

Sorry for the confusion, but regardless of whether you call it casting, conversion, or something else, it's still trivial to do safely in one directly and impossible to do safely in the other. Given your fixation on the terminology I assume you've now accepted this.

> But via an optional type you guarantee that it's handled

It's not 'handled', which is the point. There's no default value so you have no option but to propagate the missing value. Callers that establish the precondition still have to deal with an Optional value even though it can't be empty. Callers which establish the precondtion through the types receive a more precise type and don't have this problem. Callers that establish the precondition in the more simply-typed version (Int, Int) -> Int also receive an Int. Only your version imposes the imprecision on the return type.

> There is nothing wrong with using a Maybe monad here. There is zero reason why you need to throw an exception.

The Optional[Value] returned from a map lookup is used to incidicate the possibility of a missing key, which is an expected outcome of the operation itself. Passing a null map represents a structural error in the construction of the program at the point the call is made. If you represent both in the same type you can't distinguish these two cases.

> You now have a hole in subtraction as 3 - 3 would be a singularity.

There's no hole here, subtraction on NonZero just has to return Int instead.

> If it comes from IO anything goes, you have to be prepared for Anything

Yes, if it comes from user input you have to establish the property dynamically. Encoding the property in the argument type forces you to actually do it (whether statically or dynamically), and help you push the constraint to the highest level. Putting the optionality in the return type doesn't force you do do this, and imposes a cost on every call that actually does.

> I said why does it even matter not why can't you do it

I've explained why it matters - the types are more precise in my version and if you start from that you can always throw away the extra precision if desired to get to your version. You can't go in the other direction, so starting from your version makes it impossible to safely recover an Int from the returned type of Optional[Int], even if you've already established the precondtion beforehand.

> There is 100% casting in your version

Creating an Optional[Int] from an Int is a conversion, not a cast. I thought it was obvious from the context but for the avoidance of any doubt, by 'casting' I mean an unsafe narrowing conversion. Optional[Int] is a larger type than Int, so it's trivial to create one from an Int:

    def pure(x: Int): Optional[Int] = Just(x)

you clearly can't safely go in the other direction, whether using pattern matching or otherwise. If you disagree, just complete the following definition:

    def fromOptional(o: Optional[Int]): Int =
        match o with
        | Some(i) => i
        | Nothing => ...

eventually you need to provide a default value for the case of no value.

> Imagine a map with RGB colors as keys.

Your example doesn't make sense, what would you expect (lookup Map.empty Red) to return? The optional return value is used to represent the key being missing in the map. Nonetheless the point I was making is that you wouldn't return Nothing from such a function in the event of a precondition failure e.g.

    def lookup(m: Map[K, V], v: K): Optional[V] =
        if m is None return Nothing
        ...

you would instead throw an exception if the input map is null and force the caller to handle it. The majority of static type systems are not powerful enough to encode arbitrary properties about values, so you have to decide which ones to check dynamically and which statically. Checking preconditions dynamically is reasonable if encodng them in the type system is too cumbersome.

> This prior point involves the creation of the type NonZero[Int] which involves: NonZero.fromInt

No, this is not necessarily the only way to create instances of NonZero. You could have a PosNat subtype with members one: PosNat and succ: PosNat -> PosNat. You could have a non-empty list type with a length member.

> Every other mathematical operation (+,-,x^y,/,) returns an Int not a NonZero[Int]

They don't return Optional[Int] either so I don't see how this is relevant. There's no reason the input has to come from some application of a different operator, it could come from configuration, user input, a property from some other type etc. The question is whether and how to model the constraints in the type. The constraint exists in the argument so it makes sense to constrain the input type, not widen the output.

> Notice how the above two sentences are the same?

Yes, if all you want to do is avoid establishing the property you care about and silently propagate some information-free 'failure' value to the top level, then you can do it either way. But the entire point of encoding properties in the types is to force you to establish them. These statements highlight the difference:

1. I've established the divisor is non-zero, called myDiv, received an Int and continue

2. I've established the divisor is non-zero, discarded that information to call yourDiv, recieved an Optional[Int] which cannot be empty, but which must be propagated. You could immediately unwrap the value but now you're just re-creating the dynamic behaviour of a function (Int, Int) -> Int which you've already rejected.

> First, Why does this even matter?

The reason you can't write my version using yours is that the types are less precise and you can't recover the imprecision in the output type after the fact. The only safe way to obtain an Int from an Optional[Int] is by providing a default value which doesn't exist in this case.

> The Optional[Int] doesn't exist so how do you create it?? You CAST

By casting I mean an unchecked narrowing conversion e.g. of the type Optional[Int] -> Int. There's no casting in my version.

> if you want to talk about "Well Established" then Optional is more well established

This is a false dichotomy, contracts are still used in static languages where you can't or don't want to try represent properties at the type level. You could for example define a function

    lookup: Map -> Key -> Optional[Value]

and still add preconditions that the map and key were non-null. The failure to uphold these represent a different kind of 'failure' than the key not being found so it wouldn't make sense to lift them into the return type.

> The safe version suffers from your same problem just moved

It didn't 'just' move, it moved to the point in the program you actually need to deal with the possibility of a zero divisor i.e. before calling div. Where does the divisor come from in the first place? You seem to be assuming there is necessarily some call to NonZero.fromInt at each call site to div but this is wrong. The non-zeroness of the divisor could be established at some prior point in the program and used in multiple places. In contrast your version has to deal with the possibility of returning None everwhere even if you've already established the property of the divisor beforehand.

> This is just your arbitrary preference

It's not arbitrary since it's possible to write your function using mine but not vice versa. If you disagree then please implementing the following function without casting:

    def convertDiv(f: (Int, Int) -> Optional[Int]): (Int, NonZero[Int]) -> Int

> You're just trying to justify a convention of doing this check before rather than later

The convention that callers are responsible for upholding the preconditions of the functions they call is well established: https://en.wikipedia.org/wiki/Design_by_contract. You obviously can't fix precondition violations by checking the result after the fact.

> Also Your unsafe version is again worse because it will trigger an exception on zero

That is the point of the unsafe version, yes. Sometimes you will statically know the argument is non-zero e.g. NonZero(3). If you want to avoid an exception then use the safe version.

> The structure of the code handling this type of div is identical to code handling an actual exception

You would never write an exception handler to handle such a failure from div. The divisor being non-zero is a precondition of calling div in the first place, which is something the caller is responsible for upholding. You shouldn't ever need to write an exception handler to catch precondition violations. Do you also write handlers to 'handle' null dereferences? Representing the partiality in the return type is just pushing the responsibility to some code that can't reasonably do anything.

> What then happens when I pass a zero?

I've already explained this, you obtain a NonZero[Int] from a function

    fromInt : Int -> Optional[NonZero[Int]]

and you can optionally add an unsafe version with type

    Int -> NonZero[Int]

> All you did is propagate the issue to somewhere else

Yes, the check has to be done somewhere since that is the point of encoding the property in the types. But encoding it in the argument type ensures the check is done before div is called which is where it needs to be done.

> There is nothing weak going on here

The type for your static version of div is:

    def div(x: Int, y: Int) -> Optional[Int]:

The precondition for the dynamic behaviour of div is that the divisor is non-zero. If you want to think of it in set terms this means y must be a member of Z - #{0}. But your encoding allows any member of Z. This is a larger set and therefore a weaker precondition.

Likewise the postcondition of the dynamic version is that an Int is returned, but yours only guarantees an Optional[Int]. Again in set terms this is something like Z + #{None} which is a larger set than Z and therefore a weakening of the post condtion.

Your version encodes the dynamic behaviour of throwing an exception in the return type, but I'm saying that

    def div(x: Int, y: NonZero[Int]) -> Int

is a more precise static definition of the behaviour of div, and if you're going to add types to the dynamic version you should prefer this approach. Our two definitions are not equivalent since your function can easily be implemented with mine:

    def yourDiv(x, y) = map (fun nz: myDiv x nz) (fromInt y)

but you can't (safely) implement my version using yours since it returns an Optional[Int] and an Int is required. This shouldn't be surprising since algebraically (Int, Int) -> Option Int is a larger type than (Int, NonZero[Int]) -> Int.

> You're not seeing the bigger picture. The only way to avoid a zero divisor is to use your made up type NonZero[Int]

Obviously I know this because I suggested using NonZero[Int] in the first place. Your version accepts an Int divisor and just encodes the partiality in the return type. But this doesn't force the user (i.e. caller) of div to handle a non-zero divisor at all, it forces the user of the return value to handle a potential missing value without any context for why it was missing. Propagaging this missing value isn't 'handling' the error at all, since that can only be resolved by ensuring the divisor is non-zero before calling div.

Div always returns an int as long as the precondition - that the divisor is non-zero - is satisfied. The caller is always responsible for ensuring the precondition is satisfied which means the caller must ensure the divisor is non-zero before making the call. Checking it dynamically and throwing an exception does have the drawbacks you state, but if you're going to represent the contract in the type system then you do that by constraining the type of the divisor, not widening the return type.

Your encoding has really changed the meaning of the div function - it now always accepts any two arbitrary ints and always allows the possibility of returning None. So you've weakened both the precondition and the postcondition, which is now easier for the caller but imposes a cost on every location where the return value is used. As the optionality encoded in the value propagates further from the call to div, the context for the source of the missing value is lost. This is 'safer' in the sense of avoiding crashes at runtime but doesn't actually help resolve the actual issue of avoiding a zero divisor when calling div.

My first preference is to change the type of the divisor to NonZero[Int], and throwing an exception within div is only my second preference. It doesn't make sense to change the return type of div to Optional[Int] because the success or failure or the operation is determined entirely by a property of one of the arguments, which the caller can always check. Making div partial and encoding it in the return type just moves the the problem away from the place it occured and can actually be handled.

The check has to go somewhere and the caller to div has the most context in the event the divisor is 0. If you return an optional from div then you either impose the check on all the callers, or just propagate None everywhere and some top-level function has to deal with mysteriously missing values.

The NonZero type should be responsible for checking the wrapped value is non-zero, you will probably want safe and unsafe constructor functions

    Int -> Optional[NonZero[Int]]
    Int -> NonZero[Int]

where the unsafe version throws.

If this is overkill for your application then I'd prefer throwing an exception in div rather than encoding the failure in the return type.

For division you really should put the requirement for the non-zero check in the type of the divisor instead of propagating failure to the caller e.g.

    div(x: Int, y: NonZero[Int]) -> Int

Exceptions at least have the benefit of retaining the location the context for the error occurred which gets lost without extra bookeeping when using optionals.

I think it's clearer to think of map as a function

    (a -> b) -> (f a -> f b)

i.e. one which lifts function application over values of the functor. This is then analagous to the role of `f` for mapping types, and corresponds to the concept of a functor in the category of types. This relationship is harder to see in the tupled version. There's a similar argument for e.g.

    =< :: (a -> m b) -> (m a -> m b)

> `Maybe a` is a type-lambda abstraction (still a type)

I don't know what you mean by this - lambdas are values not types. You could argue the type of the Maybe type constructor is Type -> Type but there is no Type type in Haskell so Maybe is not a 'type' in the Haskell type system in the way e.g. String is. The separation of types and values is AFAIK the defining characteristic of non-dependent type systems, which Haskell is (for now).

Maybe and Monad are not types - Maybe is a type constructor and Monad is a type class.

Your first two cases are equivalent to the `Some(o) => o` case; the syntax I used was a bit ad-hoc but the following F# typechecks:

    let join_option oo =
        match oo with
        | Some(o) -> o
        | None -> None