Is Go Duck-Typed?
bionic.fullstory.com
bionic.fullstory.com
Interestingly, python has included structural subtyping in 3.8[1] as part of the typing module.
Structural vs. nominal typing is pretty well known in programming language design circles, but circulating that knowledge is a constant uphill challenge. One of the reasons many of us are frustrated with Go's designers were how quickly they threw this work under the bus as part of their initial marketing strategy, which kind of knocks away the ladder for anyone else who is curious to learn about this stuff.
AFAIK SML is nominal. OCaml has a structural subsystem in that its object system is structural. The vast majority of the language is nominally typed.
> It's probably more that most mainstream typed languages, like C, C++, Java, C#, etc. have mainly gone down the nominal route
It's not just mainstream typed languages. Almost all statically typed languages use nominative typing. Tt doesn't have any real drawback and it's just simpler to understand and work with.
A language like typescript would go with a structural system because it's a much easier bridge and sell from an object-oriented dynamically typed world.
I guess I was referring to how records are structural. Although granted tagged unions are nominal, and you can get nominal typing through modules. I was probably wrong in posing it as 'one or the other' - seeing as many languages have a mix of both. I'm definitely not saying that nominal typing is bad, it's just that it's nice to have the option to go structural if you want, and many have not known that this option exists.
functor F(S : sig val x : int end) = struct ... end
structure Foo = F(struct val x = 3 val y = 4 end)
datatype foo = Bar | Baz
val foo1 = Bar
val foo2 = Baz
You might be thinking instead about how some languages implement algebraic data types over a language like Java by replacing "foo" with a supertype and implementing the variants as subtypes. Scala and Kotlin are examples of this. Doing things this way allows you to do interesting things that are not possible in SML, such as typing for exactly the variant you expect.
Ocaml has always had, in addition to normal records, structurally typed records and polymorphic variants. It also has SML's structurally typed module system, though goes much further by allowing modules as runtime values.
Duck typing as a concept was re(?)-popularized around the advent of ruby, so might make sense to talk from a ruby perspective: http://rubylearning.com/satishtalim/duck_typing.html
With ruby, most everything you interact with is a true object, ie. you can modify behaviour of almost all classes/objects by adding/removing methods, mixin, etc. It's not unheard of monkey patching Integer or Array classes, core parts of standard library and language. Thus, you could dynamically invoke any object with any methods, without regards to compile-time constraints or "types". Errorhandling thus delegated almost entirely to runtime beyond basic syntax parsing.
Golang furthers the notion of capabilities as shared method signatures, by loosely binding this into interfaces, and enforcing many errors by static type checks at compile time. It's interesting as a development away from inheritance and towards composition and loose coupling of static program components.
"Duck typing" can be implicit, if you can call quack() on this() without error, you (the programmer) assume it's a Duck. The language may make it explicit (loosely structural typed like in golang, or other variations), or implicit (like in ruby). There are varying degrees, but you should be able to have a collection of the "Ducks" and be able to invoke quack() on them all.
Can't really say I see the connection with reflection. If quack() fails, you can get an error, either at runtime or compile-time. So no reflection required, though using reflection you may avoid errors dynamically.
Structural Typing as a concept is much better defined though.
> Structural typing is a static typing system that determines type compatibility and equivalence by a type's structure, whereas duck typing is dynamic and determines type compatibility by only that part of a type's structure that is accessed during run time.
That sounds like static vs. dynamic implementations of the same thing.
Consider a generic "proxy" object that can be arbitrarily disconnected and reconnected to objects of different types. These can exist in any duck-typed runtime system. What is such an object's "type", even in the sense of its "runtime type"? There isn't one. There might be a sense in which it has an instantaneous type—a type it has as of a given program world-state—but that information is inaccessible to the runtime, since any probing it might do to ascertain this might also cause the object's instantaneous typing to change in the middle of the probing procedure.
(See also: the Universal Server in Erlang [ https://joearms.github.io/published/2013-11-21-My-favorite-e... ]. What is this server's "type"?)
A type system is something that can make guarantees about the behavior of a program in advance. In that sense, duck-typing isn't really a type system. It's just asking objects questions and then blindly trusting the responses you get. There's nothing but convention guaranteeing that an object that e.g. in Ruby implements `respond_to?` a certain way, will _actually_ respond to those methods when they're called. The object can lie. Which means you don't have a guarantee, and so you don't have a type system.
This, I think, is the main point of the parent comment.
I think the distinction (and overlap) come into focus when you think computationally: how is a type computed? Type semantics are determined by the type checker, and Go's interface types are checked at compile time. Duck type checking happens at runtime. There's no language support in Go for this, but the closest thing would be passing a bunch of empty interfaces around and using the stdlib's reflect package everywhere so that you never panic, but sometimes return a user-generated TypeError.
Sure, Go's interfaces allow clients to accept multiple concrete types, but that's checked at compile time, not runtime, and so it will never be duck typed.
package main
import "fmt"
type Dog struct{}
func (d Dog) NumberOfLegs() int {
return 4
}
func (d Dog) Genus() string {
return "Canin"
}
func main() {
var dog interface{} = Dog{}
fmt.Println(dog.(interface{ NumberOfLegs() int }).NumberOfLegs())
fmt.Println(dog.(interface{ Genus() string }).Genus())
fmt.Println(dog.(interface{ Happy() bool }).Happy())
}
https://play.golang.org/p/d0TkGTCZa2S
Runtime results in panic:
4
Canin
panic: interface conversion: main.Dog is not interface {
Happy() bool }: missing method Happy
goroutine 1 [running]: main.main() /tmp/sandbox251002434/prog.go:18 +0x200
So golang is somewhat dynamic as well, although using interface{} is kludgy and one would want to avoid that and reflection whenever possible.
With duck typing, the interface is implicit - if `f` only calls `a` and `b`, then that’s exactly what `o` needs to support.
Duck typing also allows more dynamic interfaces, like “`o` must support `a`, and if calling `a` returns true, it must also support `b`”.
# let f obj = obj#foo 1 2;;
val f : < foo : int -> int -> 'a; .. > -> 'a = <fun>
I think the more dynamic nature is really the crucial difference. In OCaml, if the function above had an if/else, and called #foo in one branch, and #bar in the other, the type of the object would be inferred as having both methods, and would enforce that at compile time. With duck typing, you could pass something that only has #foo, so long as the right branch is taken.
def foo(o):
if (b := o.a()) is not None:
b()
If you do a tiny modification to "def foo(o, a): ..." then this doesn't apply any more, and you're outside the realm of structural types and into the realm of dependent types, which means you must drink deeply of the static typing kool-aid and want to write type signatures that run the risk of being more complex than the code they describe.
(But I like to quote Conor McBride: "if you're not going to write types your type system has to be stupid enough a computer can do it", from https://www.youtube.com/watch?v=3U3lV5VPmOU&feature=youtu.be...).
interface I { void m(); }
class A { void hello() { ... } }
I i = new A()::hello;I love how you can just return a custom object literal from a function and TypeScript figures out the type, without ever giving it a name. This lets you explore with pretty much the same flexibility and speed as with dynamically typed languages, and only after you shaped your code sufficiently well, you can decide whether you want to name some of the types and make everything a bit more "solid".
The example they have is not illustrative or explanatory at all.
class Duck:
def fly(self):
print("Duck flying")
class Sparrow:
def fly(self):
print("Sparrow flying")
class Whale:
def swim(self):
print("Whale swimming")
for animal in Duck(), Sparrow(), Whale():
animal.fly()
Duck flying
Sparrow flying
AttributeError: 'Whale' object has no attribute 'fly'
Go isn't duck typed, except for interfaces which are. Pretty simple.
I think the difference is that in duck typing, there is no real type. A thing is a duck if I can use it as one, and that same thing can be a chair if I can also use it as one.
So on duck typing, we say, does it have the properties I need, if so I can happily use it, don't care what it is. And that can only be done at runtime, since you can't fully predict the runtime properties the thing will have.
In structural typing, we say that the type of a thing is defined by its structure. This thing is a Duck because it is of the expected structure of a duck, which might be to define a quack string -> string method. At this point, you say the thing is a Duck. You can do this at compile time, because you have defined statically a set of named things and their structure, and then you can track the things passed to you and match their structure to the list of structures you have to find the one that fits and consider it of that type.
quack() in duck typing, has to have no arguments and return a string (let's say). We can have a NotarizedDuck class where quack(true) instead returns the number of times the NotarizedDuck has quacked, so long as quack() itself returns a string, it will pass the unit test.
In structural typing, it's illegal to call that quack, because it's the wrong function signature.
How frequent are problems where classes unintentionally use an interface due to identical signatures? Seems likely to happen in theory if so many interfaces have few functions.
func InsertFoo(*sqlx.DB, Foo) (int64, error)
Would those be mixed up?
There's some "self-limiting" - the smaller the interface, the more likely you get accidental implementation, but also the less likely you are to confuse them in a way that typechecks but actually does something unexpected.
A practical example might be io.Closer - I am certain I implemented this in a dozen types I never intend to pass to any io functions. But no io functions actually take just a Closer, and if I did write some generic cleanup routine that only requires a Closer, it's probably fine and maybe even good if it works on byte IO and gRPC connections and Kafka streams. (Especially without generics, it might be the best I can do.)
You might care if you expect the Id string to look a particular way. But then you probably wanted to use a type other than string for the return type.
That's the point. In order for (String userId) to not be confused with (String password), I wrap them with nominal types UserId and Password.
If the language looks at both UserId and Password and decides they're interchangeable because they both satisfy the contract 'contain a String', then that completely negates the point.
I used to worry about this sort of thing, but now I think that any language designer making a new language should just consider it a non-issue. If you happen to make it impossible for some other reason, sure, cool, I guess, but don't put even a little bit of effort into worrying about this case.
(Note I'm not saying it can't happen, or even be bad if it did. I'm sure it's happened to some people, and of that set, I'm sure there must be at least one person with a story of how it went really badly wrong somehow. I'm just saying that compared to the sorts of issues that affect every programmer in your language every day, worrying about this edge case is not productive. If this is your biggest problem in your language, or even fits into the top 100, please let me know about your surprisingly perfect language as I am very interested in using it.)
It's interesting to me that people are very concerned about Go's structural subtyping, but they generally have no qualms about passing functions around. Given that the contract for a function is the function signature, and the contract for an interface is [all function signatures and their associated names], surely the latter is less error prone?
In that case, the fault lies with the thing that put unrelated functions together and passed them to a thing expecting them to be related. I'm not claiming that structural typing will somehow prevent programmers from deliberately writing wrong things, because I mean, what type system can make that promise? My point is that it is in my experience very rare for something expecting "something that can write bytes" to be accidentally passed "something that can write novels" or something equally unrelated, and then the world blows up in a bad way (that is, not just an exception thrown, because that's just a risk of dynamic languages, but actual bad things happening).
Dynamic languages can at least still get what I said wrong; in Go it's even harder because as others have pointed out, Write([]byte) and Write(NovelInput) Novel still can't cross. But even in dynamic languages, this isn't a problem I had.
func receiver(next func() bool, getItem func() Foo) {}
type FooIter interface { Next() bool; Item() Foo }
func receiver(iter FooIter) {}
Any interface can be implemented by something that's accidentally (or intentionally) stateful - there's nothing that prevents that in Go. And a interface containing more than one function almost certainly has an implicit relationship between those functions that an implementation could accidentally violate.
> and accidentally implementing an interface is itself vanishingly rare (I don't think I've ever heard of this happening in practice).
Accidentally implementing an interface must happen all the time - not in the sense of unrelated functions having a name collision, but in the sense of implementing a type that looks like another type and not explicitly looking through all the interfaces that other type implements and what expectations they carry.
> It doesn't make sense to me to take grievance with Go's interfaces (which offer strictly stronger protections than callbacks), but be okay with naked callbacks on the basis that "people tend not to use them in a stateful manner" (never mind the stateless scenario in which bare callbacks also provide weaker guarantees than Go's interfaces).
Naked callbacks make it clear what the expectation is and where the responsibility lies. It's the same as the idea that bad encryption is worse than no encryption at all, or a datastructure that is "mostly threadsafe" is worse than one that immediately errors when used from multiple threads.
The problem with structural typing isn't that completely unrelated things end up sharing a type. It's that you have no way to express the subtle distinctions between similar but different things, which are actually when having distinct types is most important.
He gives an excellent example leading to this dictum: single property interface "Drawable". Let Shape implement drawable with widely expected results. But: let Cowboy implement Drawable -- now with disastrous consequences. The solution to this dilemma is as stated above, not to allow concepts/interfaces/{wave hands} to refer to a single object -- for example, a number is a concept which refers to multiple operations (i.e. methods).
To my mind, this feels right, albeit initially surprising. It makes me think that e.g. the "Neg" interface in Rust is wrong.
The problem with "Drawable" is that it is property/data, and not messaging-based, thus breaking encapsulation, even if it was behaviour-based.
That's definitely a good reason to prefer small interfaces; I hadn't thought of that but it certainly makes sense.
That said, I think we may be referring to different things when we use the word "property". Stroustroup is talking about properties in the mathematical sense, i.e. he's talking about the properties of the operation "draw()" in the concept Drawable.
Am I misunderstanding you? To my mind that's an operation and closer to "message" (emphasis on the verb aspect) rather than data or property.
So I suspect if we seek more meaningful objects and models, they become more active, have better names, improved design, etc.
If you ask a janitor about Drawable, will he say it's all about draw()? So I think it's kind of a dreary default that we need to avoid and focus on who the actors are, what messages they will pass, and the hidden treasures they encapsulate (don't show, but tell!).
Anyway a serious answer to your comment would be to create a field or ring concept. This would unify some or all of the arithmetic operations.
A neat bonus would be the ability to overload the one and zero literals for the set. So in field of nxn matrices on real numbers 0 is the zero matrix and 1 is the identity.
Ha!
> Anyway a serious answer to your comment would be to create a field or ring concept.
Yeah, this is true. We really, really struggled to create a set of traits for numerics, even regardless of these more advanced ideas... so it's unclear this would work. But it's certainly closer than I had thought :)
To use a popular quote "if it acts like a duck-type it is a duck-type".
Personally, don't think Go needs to change anymore and be "duck-typed" via more powerful generics but we'll how it goes :p