https://www.manning.com/books/type-driven-development-with-i...
https://www.manning.com/books/type-driven-development-with-i...
As an aside, isn't type driven development how everyone codes naturally in statically typed languages? Whether it's Ocaml or Java, I feel like I try to get the types defined, then put on music and turn my brain off and just line the types up (much moreso in Ocaml than in Java).
It is not ideal, but pretty good in terms of benefits.
At this point I was even quite eager to update my JS skills with TS knowledge. All in all it was fun to write this project. But my error was to trust the types like I would when programming on my main project in Scala. Big error…
Since than I think TS is much oversold.
You can do fun things with the types (even some things that I miss in Scala, and there is not much to miss in Scala in general!), but you just can't trust the TS types. So this makes them superfluous in some sense. (Yes, some types are better than no types; especially when it comes to the IDE experience; but as long as any lib call (due to trivial bugs in "typing-files"), or some "impossible" input can break things this is not as helpful as one would like; especially when coming form really strongly typed languages).
Later versions are very strict even about dynamic types, treating them as Union[] rather than Any, forcing you to check the type before using it. Likewise with strict non nullable.
The same, but at "lint time" (not runtime) will happen with mypy or pyright, although I am sure there are rough edges.
This misses all the advantages of static types.
Dynamic languages are in many ways simply a recognition of that fact, namely that data is sparse, incremental, changes and accretes and so on. There's a reason languages like Idris, Haskell and so forth are much more popular in solipsistic domains like theorem-proving than they are in a business application.
Scala developers would disagree.
Especially in business applications strong static types are extremely helpful to prevent bugs and model the domains of interest in a proper way.
One may argue that the proof-strength formality can be a hindrance: in such domains, inconsistencies and contradictions abound, and some of them cannot be predicted in advance. But I believe that languages like Idris can accommodate this. Maybe I am wrong, or maybe they can, but not more efficiently than more free-form languages.
...but ok, I am failing to see what I can do with COBOL that I cannot do with Idris, business-logic-wise. lol
Also, like the other guy mentioned, the types lend themselves very well to auto-completion. It's no coincidence than in Hoogle, you can often find a function by its type. E.g., what could "[a] -> [b] -> [(a,b)]" possibly do?
You're coding a card game, say poker. Games should only start with a shuffled deck.
1. A reasonable, common solution would be to have a type like Deck, and a function shuffle, taking a Deck and returning a Deck (where taking can mean it's an instance's member function, and returning can mean inplace mutation)
2. A type driven approach can be having two distinct types: Deck and ShuffledDeck, and a function that takes a Deck and returns a ShuffledDeck, and all procedures henceforth in the game's logic require ShuffledDeck - _not_ Deck.
In the type driven approach the possibility of playing with an unshuffled deck has been removed solely through the usage of types, and successful compilation is a proof no cheaters can sneak in a bad deck.
But it's not just that. There's exactly one function that takes Deck and returns ShuffledDeck, so a language like idris knows this has to fill a hole somewhere, in the one place that starts with Deck and then need ShuffledDeck. So in some sense part of the code has been written automatically.
I rarely see the second approach used, though I see analogies of this example all the time.
In your example, how would you qualify using an enum "State" with the case "shuffled" and "ordered" ? I feel like this would help me check all cases in places where it's needed, sometimes with the help of the compiler. But does that qualify as "type-driven" ?
In particular, do languages like idriss require encoding all possible states of a state machine as a unique type to leverage the full power of the checking mechanisms ? This would seem a bit unmanageable in practice..
I've never used idris, but as far as I understand it encoding state machines in the type system is exactly the kind of thing type-dependent languages in general and idris in particular shine at. This is the prototypical example given, and it's the one Mathew Farwell himself gives in a podcast interview [1]
[1] software engineering radio, episode 296, http://feedproxy.google.com/~r/se-radio/~5/Kg1Py4rd2i0/SE-Ra...
Its type system is fairly complex. To accomplish the goal you describe, functions must be declared as deterministic, non-deterministic, or multi-deterministic. Meaning, they always have a solution, sometimes have a solution, or multiple solutions for the same input.
Armed with that information, the compiler will throw an error if you haven't thought of all possibilities in your logic flow. The fact that the type system required you to be very descriptive in many different ways compounded the number of bugs that the compiler would catch before your program ever runs... I think someone said "if you can get your code past the compiler, it will work on the first try."
I wish Mercury were more popular, it's an awesome project!
pub fn parse(s: String) -> SubscriberName {
let is_empty_or_whitespace = s.trim().is_empty();
let is_too_long = s.graphemes(true).count() > 256;
let forbidden_characters = ['/', '(', ')', '"', '<', '>', '\\', '{', '}'];
let contains_forbidden_characters = s
.chars()
.any(|g| forbidden_characters.contains(&g));
if is_empty_or_whitespace || is_too_long || contains_forbidden_characters {
panic!("{} is not a valid subscriber name.", s)
} else {
Self(s)
}
}
And let's say your code to add a new subscriber to the email list only accepts a SubscriberName, not a String: pub struct NewSubscriber {
pub email: SubscriberEmail,
pub name: SubscriberName,
}
Because `parse` gives back a SubscriberName if and only if all of the parse checks pass, you simply cannot add a new subscriber without all checks passing; the function will simply fail otherwise.This example comes from the excellent book Zero To Production In Rust, which has a chapter on type-driven development [0].
I'm quite sure nobody would do something like that "in production"…
The problem starts with the signature of this function: There is—of course—no function which could take an arbitrary String and return a SubscriberName.
You need to return at least an Option, or better a Result.
But at this point the fun starts. You need to thread this wrapped value throughout your whole program.
Crashing your program (by a panic in the above function, or by a unwrap somewhere later) because someone sent you an invalid String is not an option in real life.
It's easy to validate something "and just throw an exception".
The fun starts when you can't do that (or don't want to do it like that).
But I would need to look this up also myself. I'm not 100% sure. (But I've never heard of any implicits implemented in Rust; they discussed this feature here and there, but it's not part of anything in the language or the compiler, afik).
OT: I've never seen you on the Heise forums ever since. But I guess there isn't much to miss anyway by now…
I've recognized your handle as I remember you as being one of the few people there who actually know what they're talking about when it comes to software engineering.
But I've also moved on. The overall level has dropped to a bottomless pit by now. It's no fun anymore.
Fair enough and thank you for the kind words. :)
Does that work in anything but boilerplate code? Like, hole filling won't be able to decide between then and else clauses in a conditional, for instance, because both clauses return the same type. And conditionals are one of the most basic data flow structures.
It depends on the problem and the language.
The most impressive I've ever seen was implementing "call with current continuation" as a monad, given nothing but the type signature.
Generally, it works best in strongly-typed, functional languages.
That sounds cool. Where to find this? I get only irrelevant search results.
I'm curious about this. Can you provide some examples, please? Which languages can do this?
https://www.youtube.com/watch?v=DRq2NgeFcO0
The book also walks you through lots of examples. The key part is the type needs to constrain the possible implementations down quite a bit.
you can sort of do this with ghci and :hoogle.
the simplest example is
foo : A -> A
since we know absolutely NOTHING about A, all we can really do is give it back to you, so foo turns out to be identity, or id.
another might be bar: [A] -> Nat we still don't know anything about A, but we know a little about lists, so a good suggestion might be length.
All of the environments are pretty interactive, and you pick and choose. but from time to time there is exactly one possible implementation that meets all the constraints, so the compiler can fill that in.
When the solution is ambiguous, Idris (and I believe wingman) will supply suggestions that you can pick from.
In your `bar` case, `bar xs = 42` is also a solution. Idris suggests 0 1 2 3 ... I think, it's going off of the constructors for the return type. I expect this technology to improve over time (try to use arguments, etc.)
Idris and Agda will also do stuff short of "fill in the answer" like "case split an argument for me" or "insert the appropriate constructor or lambda and leave a hole inside". This is quite helpful in cases where it's not gonna find the answer, but can type out the next step or two for you. (Wingman might have this, I haven't used it much. I don't know Coq at all.)
On the Agda side, the first couple of sections of "Programming Language Foundations in Agda" will walk you through using that functionality. Idris has a couple of videos that others have pointed out. It's fun stuff.
and yes programming with holes is almost magical. I really like "run as much as you can" then complain when you get stuck. kind of a mind blowing change for me.
Not if you add the condition that the type must be as general as possible.
> since we know absolutely NOTHING about A, all we can really do is give it back to you, so foo turns out to be identity, or id.
> another might be bar: [A] -> Nat we still don't know anything about A, but we know a little about lists, so a good suggestion might be length.
Certainly a good suggestion, but, unlike `foo`, not the only one—e.g., `bar xs = length xs + 1` or `bar xs = 2*(length xs)` also work.
For some type signatures there is (are) only one (or only a few) meaningful implementation(s).