The way static types can make development speed slower is not by forcing you to write the same code with extra types added in, but by forcing you to write different code. In this sense it isn’t really an argument about static types vs dynamic types (imagine a language with no type checks where programs have undefined behaviour at run time on anything that would have been a static type error) but rather a statically typed language vs a dynamic language.
There are patterns or language features in dynamically typed languages which are very hard to translate to static types. I think Peter Norvig has some slides somewhere comparing OOP patterns to lisp, a more dynamic language, and that is the sort of thing I’m thinking of here. For language features, consider implementing an n-ary list map in Haskell vs Common Lisp. Or the typed equivalent of delimited continuations.
That said, research on the differences between languages is difficult and inconclusive. I would like to see more of it.
It's very clunky to do in Haskell, but conceptually they can be built on top of a monadic pipeline.
In practice, the only case where statically typed languages force you to write different code is if:
(a) the type system is too restrictive and dumb (no generics, awkward function types, etc).
(b) you're trying to write a function that effectively returns monster types like Union[None, int, str, List[str]], which is common in dynamic languages but it's something you're not really supposed to do.
I agree with your broader point, that static types can in fact slow down development, but I have a different model.
Types are kind of two things at the same time: tags to tell the compiler how many bytes need to be allocated, and partial proofs of correctness of the program. The key observation of dynamic languages is that in most cases we can do without the first part: the runtime can handle that in the same way it handles memory management. I believe one of the mistakes of the original wave of dynamic typing proponents is that they kind of threw away the baby with the bathwater: the correctness part is important, and at least a form of optional typing should be provided.
The real trouble is that it's not easy to conjure type systems that are both expressive and ergonomic to use for the average developer. Sufficiently advanced type systems are indistinguishable from math riddles, and in a working environment that can be frustrating.
n-ary list map means a function that does map or map2 or map3 or …. (Where map2 is zipWith, not the one you get from the list monad.)
In Common Lisp the lambda-list looks something like:
(MAPCAR func list &rest lists)
And a slow implementation might look like:
(defun mapcar1 (f list &aux r)
(tagbody
g
(when list
(push (funcall f (pop list)) r)
(go g))
(nreverse r)))
(defun every (pred list)
(dolist (x list)
(unless (funcall pred x)
(return nil)))
t)
(defun mapcar (f list &rest lists)
(prog ((lists (cons list lists))
r)
loop
(when (every #'consp lists)
(push (apply f (mapcar1 #'car lists) r)
(setf lists (mapcar1 #'cdr lists))
(go loop))
(nreverse r)))
To implement such a function in ocaml you need to do something like:
type ('ftype,'output) t =
| [] : ('a,'a) t
| (::) : 'a list * ('b,'c) t -> ('a -> 'b,'c) t
val mapn : f:'f -> ('f,'x) t -> 'x
In haskell you would use a heterogenous list and type families. But note that I didn’t implement mapn in ocaml as it would be hard and unreadable…
One nice quality of dynamic languages is that you get very useful objects like lists or python/JavaScript style ‘dicts’ and lots of standard library functions to operate on them. You can implement them in a statically typed language to some extent but they are much less versatile and less first class. It means you often need to write different functions that do basically the same thing because they use different record or tulle types as input. I don’t think any mainstream languages are very close to static typing with structural subtyping.
For people for whom coding is an iterative, exploratory process, forcing type declarations upfront is not just slower… it’s a no-go.
There’s no way to bootstrap that process for them: you’re asking for the output as an input.
I think TypeScript has one of the best compromises.
Also, I think people are understating how easy dynamic typing makes substitutions — everyone else is writing their class twice (interface + class) just to get the same flexibility.
> For people for whom coding is an iterative, exploratory process, forcing type declarations upfront is not just slower... it’s a no-go.
Is there any kind of programming task where it's too much to ask the programmer to know the type of the function he's writing?
The only place I can imagine types being a burden is in a REPL (it's just boring to write them every time), but even then optional types like Python and TS are an unequivocal win (you're presumably going to move your code out of the notebook/repl and into proper source files at some point).
> understating how easy dynamic typing makes substitutions
Yeah so easy they're easily wrong. One of the worst problems of dynamic languages is that they don't have the concept of an interface, which makes it harder and not easier to write generic code.
> everyone else is writing their class twice (interface + class)
That's caveman Java right there. Interfaces with defaults (Java, Kotlin, even C# has something similar IIRC) can easily emulate dynamic patterns like mixins while preserving type safety and without the diamond problem.
> Is there any kind of programming task where it's too much to ask the programmer to know the type of the function he's writing?
I think the problem is that lots of type systems are inherently limited, so there may be patterns that you want to express that your type system simply can't, or that you have to write tons of boilerplate code to express. In languages like Java or C# that don't have Sum types, this can be as basic as "this value is a String or Integer". You have to write a class hierarchy to express this in Java. In JavaScript you simply accept a parameter and branch on `typeof value`. More powerful type systems allow you to express more patterns, but you still have to type them out. This can be done, but if you're doing exploratory work an unknown dataset then this can make the work much slower. And there be may little benefit.
In C# you just declare them as dynamic if you wish to emulate the ways of JavaScript.
People keep forgetting Basic is dynamic and .NET was designed to support it as well, additionally it got improved with DLR classes to support IronPython and IronRuby projects.
So you can co type crazy in C#, F# and C++/CLI, or just get hold of those dynamic features.
Oh sure. But at that point you are using dynamic typing. My dream language is one that allows for full gradual typing, with the ability to opt functions in (opt out could also work) to strict type checking which gives you the safety and performance improvements where you want them, but allows you the flexibility of dynamic typing where you don't. PHP-style runtime type checks would be automatically generated at the boundaries between dynamically typed and statically typed functions.
One problem with gradual typing is that it has comparable performance overhead to dynamic typing: You can only optimize out dynamic type checks, dispatching, marshaling etc. if you have sizeable, self-contained portions of your program (i.e. modules) that only use static types! So the "graduality" of it goes mostly unused in practice.
It can be a useful pattern when designing interface boundaries across modules (e.g. ABI's, data interchange formats etc.) to allow for the required flexibility in a limited context where the overhead isn't going to matter all that much.
Performance is only half the point. The other half is correctness. And I think this pattern of "important correctness and performance sensitive core, surrounded by simple but flexible glue code" is common enough that such a language makes sense. It describes pretty much all of the code that I write at work in fact. Being able to simply add type annotations and have the compiler generate the runtime type checks would be wonderful.
Genuine question: what does old PHP’s (not the newer stronger versions) interfaces and “type” hints count as? It’s very much a dynamic programming language from what I remember of it, but interfaces were crucial to it when I was using it in anger.
Is it one of those in-between systems, just more on the dynamic side?
I have never used PHP professionally, but if you're talking about PHP 5 style interfaces, they are definitely a form of type checking, and they were introduced with the specific purpose of facilitating object oriented programming in PHP.
The point I'm getting at is that using interfaces lets you make guarantees about what methods the objects must have. I'm not really picky about what specific technical means are being used to enforce those, but passing an object of the wrong type to a method is an entirely avoidable category of errors, and I prefer them to be flagged by the compiler/interpreter/runtime/tsc/mypy/whatever as soon as possible.
IME well designed interfaces are even more important than well designed classes, and dynamic languages unfortunately make it harder to enforce interfaces. It's not by chance that the first thing Typescript did was to introduce interfaces.
There are patterns or language features in dynamically typed languages which are very hard to translate to static types. I think Peter Norvig has some slides somewhere comparing OOP patterns to lisp, a more dynamic language, and that is the sort of thing I’m thinking of here. For language features, consider implementing an n-ary list map in Haskell vs Common Lisp. Or the typed equivalent of delimited continuations.
That said, research on the differences between languages is difficult and inconclusive. I would like to see more of it.