The problem of colour is rather that even though two functions should have the same structure, regardless of their effects, the language still forces you to write two functions each with its own syntactical structure. In type theory terms, this is a problem stemming from a lack of effect polymorphism.
Ideally you'd write one function, polymorphic in its effects, and then just be able to use it in both sync and async contexts. In the sync context, the function's type would specialize in a way that it lacks the `Async` effect, so it wouldn't be able to do async stuff there. In the async context it would specialize in a way that it does have the `Async` effect. Therein you would be unable to apply it in contexts requiring the lack of an `Async` effect.
I recently tried Go to speed up some Cloud Functions that use Firestore, it's very difficult to know which of the chained method calls of the Google-provided SDK is the one to actually "blocking" the green thread, perhaps they are all are unnecessarily? Without studying the source code, it's hard to know. I prefer explicitness over absolute simplicity. It also makes it easier to optimise code in my opinion as it's easier to identify sequential async calls that could actually happen concurrently, either by merging SQL queries or by using Promise.all() or the equivilent.
JavaScript does a good job with the await keyword in my opinion, you know exactly when something could take a significant delay, when it's doing something more behind the scenes. Rust does one better by making ".await" a suffix, allowing for nifty call chains. It communicates that this is an explicit yield point and could take a while.
Well it also sucks if unpure code has to add a bunch of boilerplate annotations to every single function and call.
> I recently tried Go to speed up some Cloud Functions that use Firestore, it's very difficult to know which of the chained method calls of the Google-provided SDK is the one to actually "blocking" the green thread, perhaps they are all are unnecessarily?
If you have to mark enough of your functions as "possibly blocking" then you can get just as lost about which one is actually blocking in a particular situation.
Since this could be read in a very abstract, purist way, I'd like to add that this is also very much an architecture concern for many pieces of code.
At work, I regularly encounter issues with a piece of code that is written in one of the worst imaginable spaghetti styles. And, since it changes things in a database (not even in proper transactions, btw), and is written to pretty much prevent being tested, no one wants to touch it.
Yet, what the code does would really match an "FP" approach:
- Read in all inputs. You might even read during computation if you're simulating FP through transaction isolation
- Compute the outputs and the new desired state
- Pass that to someone who writes it back to the database and commits the transaction
Outputs in our case are messages that get stored in the database, too (in order to allow recovering after transmission issues), so "output" in effect also just means "state". You don't always have to read "FP" in the strictest sense -- there's a spectrum to this.
Now, if that code was written in such a way it would be easily testable and thus refactorable. In fact, most ideas in the clean/hexagonal architectures or "functional core, imperative shell" are to achieve exactly this. But since the author of said code never had any pushback, they just kept making an already grim situation worse with every change. For years.
I see the same issue with the "ice nine" argument for async. If async infects all your business logic, chances are your abstractions suck.
Edit: Also note that goroutines do effectively the same thing as synchronous code in a larger asynchronous function call: You group the code you can sensibly do in one go into a single goroutine. Then you pass the result to someone else to perform the next step.
So, it's very often just another way to express the same process. CSP will invariably be a better abstraction for some problems, and async/await for others.
Then you keep pushing stuff to other threads "just to be sure", which may hurt performance but the worst part becomes reading that code.
Now you will want to blow your brains out because you will have to refactor all of your function calls down the chain.
It could get even worse if at some point, you are passing a function to a library/framework and for whatever reason it refuses to tolerate a function of your color (assuming you can't lambda-ize the color). Now you have to rewrite the library. Or use a library that does tolerate it. Which might lose other features or ergonomics.
For example, one's expectations of a pure function would mean that you could do some arbitrary numeric computation on inputs and return a value. Unfortunately, the compiler is not so naive. If you should do some transcendental math and call a standard library function to do so, you have manipulated hardware state. You are now impure, and so are all your callees.
And so you have to decide at that point: is the code you are writing wrong? Or is it just using too coarse an assumption for its invariants?
A huge amount of what determines our code architectures are the way in which we've organized our own thoughts. In this respect we always go from looser to tighter specification over time as the code matures, until tightness has produced brittleness and cracks. I believe coloration forces some tightening decisions to be scheduled earlier. In that respect it should be treated with some suspicion, as it may produce surprising couplings.
You see the same thing happen in async circles. All of a sudden every single function has to be labeled async, whether it needs to or not.
Haskell’s use of monads is perhaps the ultimate in colouring (for values not functions, but that doesn’t make that much of a difference); you routinely colour everything with which pieces of state it needs to read and/or write to do its thing. The result is you either have everything painted in “global state”, which just feels like unnecessary boilerplate, or you spend most of your time defining and converting between elaborately carved pieces of said state.
People could and did invent various intricate constructions which would do the conversions and possibly even the carving for you, but overall I think most agreed it was a gigantic pain.
It’s possible the situation is fixable with a correct combination of features (records, polymorphic variants, implicit parameters, possibly even algebraic effects), as a good portion of the pain comes from not having a standard solution every library would use (they all have drawbacks). But then the checked exceptions story in Java, IIRC, is from pre-Java 5 times, and not being able to be polymorphic in the exception signature of a callback you take is really stupid however you look at it, as well as also a case of inadequate language features.
And later on, checked exceptions were incompatible with new stuff like lambdas and streams
Zig actually does pretty well on this front. you can write a library that accepts functions that return errors, and your functions will be specialized to return those errors in addition to their own if you don't handle them. Callers who want to handle errors exhaustively with a switch will get compile errors if the error set changes and they fail to handle new errors.
In a similar vein, as another commenter above said, really the only problem with function coloring is that we're not treating it polymorphically.
public class SomeClass<E extends Exception> {
public someMethod() throws E {}
}
The other problem with this approach is that it requires the caller to explicitly specify all the exceptions thrown by the function/object being wrapped. So every time you do map() or fold(), you have to spell out the full exception spec for the closure you're passing in. And god help you if there's nested maps...
Check out Koka lang for a very interesting implementation: https://koka-lang.github.io/koka/doc/book.html#why-effects
Sadly I don't see that happening anytime soon.
If we really need to be explicit, we begin to need different libraries for each combination of effects/capabilities/whatever, sort of like how Rust users need at least 4 separate sync/async/nostd-sync/nostd-async libraries just to speak websockets. This seems like a waste to me.
Zig's error system in which errors must be handled and functions can opt to have their error sets inferred by the compiler (again, this includes specialization to different error sets for instances of the function that are parameterized by functions with different error sets) sort of gives us the same thing for errors that we have for async-ness, but I can sympathize with a perspective that feels that special casing these two things and not allowing any others is inelegant or whatever.
> Async is probably one of the less interesting constraints, but I find it so much more confusing when languages have colorless functions... I can't see suspension points, I can't tell wtf is going on. It's extremely frustrating.
Zig's view on this, which I endorse, is that suspending to your event loop is not substantially different from suspending to the kernel. Zig sort of assumes that we don't want to manually mark each function with which syscalls it and its callees are allowed to make, but maybe it's worthwhile to make a language that does do that.
The problem is that the library author must then code for the worst case and the compiler won’t help them; injecting preemption into synchronous code can expose things about it that used to be unobservable, but now can make the difference between success and deadlock. (Similarly to how injecting mutation into pure code can expose—and bar the optimizer from changing—things that were previously unobservable, like “in which order does map() visit elements?”.)
There still will be async-compatible and async-incompatible libraries, except now you can’t tell which one is which except by looking at the documentation—look at the lengths R5RS goes to in order to define things like returning multiple times from inside a letrec while R6RS just flat out prohibits that.
Yes, this seems to be a pro-colouring comment while elsewhere in this discussion I make what seem to be anti-colouring comments. I really don’t know what the answer is.
Based on what I've seen in Rust's async, you don't actually want this in a low level language. Those calling scenarios are fundamentally different. If you magically make something async, other code can now suddenly mutate state that your function is referencing on its own stack and really mess things up.
In Rust this isn't an issue because you're not allowed to pass references into an async function that outlive the future. If you take a synchronous function that holds references and add the async keyword, you probably will get compile errors about this. I don't think Zig has anything like that.
>Zig's view on this, which I endorse, is that suspending to your event loop is not substantially different from suspending to the kernel.
No, that doesn't make sense, not even in the context of Zig. Suspending to the event loop allows other coroutines to run. Blocking in a syscall stops the whole program and all the coroutines.
But to address your comment directly, no it's not, not when you're actually building a different state machine with different constraints each time. I actually wish we could put this protocol code in one C library in a way that only handles the protocol logic and not the I/O part, and then we'd be done with it forever. But you know that wouldn't be good enough for some, those libraries exist and people still don't use them and instead they go and write their own in the new fancy language with their cool new concept for a state machine.
You can have one Python library at least[1], h11 and h2 in particular make a pretty convincing case as the ultimate in Python HTTP. (You will always have wrappers, though, because not all levels of detail are appropriate for all uses—similar to how imagemagick/libgd/etc. wrap libtiff which wraps libjpeg and wrap libpng which wraps zlib and so on.)
The problem, as far as I can see, is that to avoid calling out to I/O you need inversion of control, and either you convert your code into CPS and state machines manually, which almost impossible, almost unreadable, and makes layering libraries a distant dream (not that people don’t do it anyway), or you use some sort of generic inversion-of-control mechanism like async/await or call/cc, which really wants dynamic memory allocation.
It seems like being able to accurately track stack usage (per module, not per program) could solve this, but that means (we must abandon arbitrary recursion and) the compiler must be able to track arithmetic equations, which is liable to make Gödel and Turing rise from their graves just to shout NOPE! at you. There’s also ABI: what do you do when a new version of a protocol library requires a dozen more bytes on “the” stack, and you were putting these on your stack or in static storage?
(As far as I can tell, Rust’s no_std answer to ABI of this kind is “no.”)
To that end, Zig does have another kind of function coloring that's useful there: functions which take an allocator parameter. I like that about Zig, I wish C++ and Rust were better about doing it. Sadly you can still hide this by referencing a static allocator in Zig so it's not totally there yet on the language level.
Yes, this is a great approach. Unfortunately people in e.g. the Rust world tend to write crates that assume they can call malloc, recv, and send, rather than one crate just implementing the protocol logic and another that does all that other stuff.
Zig's way of doing things causes the library people naively write and use to call malloc, recv, and send to also be a library that just implements the protocol logic, reads from and writes to static buffers, and doesn't call malloc.
Edit to reply: Yes, Rust has no_std, but most libraries that implement a protocol are not divided into a no_std library that implements just the protocol logic and a separate std-using library that wraps it, nor are most libraries that implement a protocol asynchronously wrappers around a no_std library that is not async and only implements the protocol and makes 0 syscalls. A consequence of the fundamental non-composability of naively written Rust libraries is that users who want the wrong combination of asyncness and malloc-calling very frequently need to reimplement existing Rust libraries. A consequence of the fundamental composability of naively written Zig libraries is that they never need to do that.
Another thing that you could do is in the command pattern for undo/redo, to make sure changes to your data model only happen from command objects. In Python, I had a decorator that said "when this property is set, assert that Action.do is on the call stack". It would be really powerful to be able to do that at compile time, too.
You denote that the color is optional in the function's type signature, and inside the body of the function, you have a construct that lets you execute different code depending on the state of the function's colors. The call site could either automatically infer the colors used or the colors used could be manually specified.
Basically a type parameter with compile-time if statements.
It makes sense to avoid this in a javascript runtime because you actually almost always want (3) there. It's too late for node to change it unfortunately. But I'm also pretty confused as to why zig, a (relatively) low level language, considers it important to remove this distinction.
Imagine you have 2 singleton objects in your 'void main()', redkey and bluekey of types Redkey and Bluekey.
Somewhere else, you declare your function: 'int foo(Redkey redkey, int x, int y)' that needs a Redkey object of which only one exists: in your main.
This by itself forces every call on the path between 'main()' and a call to 'foo()' to also include a parameter Redkey.
In the extreme cases where a pattern like this is useful (tracing IO calls, you named it) it can really help cut through a codebase after an hour of refactoring. But it can be limiting. Async and Checked Exceptions are probably the most colored functions and they both need escape hatches because of that.
The two scenarios are the same, yet in a colored functions language f' will need to be annotated if r does, for example io.
The way I see it, a language should either allow 'closing' over the color (or effect or capability, or whatever) hiding it, so that f' doesn't beed to care about the details of r; or it should provide enough of polymorphic behaviour so that f' the signature of f' can be inferred from that of r.
Most languages with async do not provide this capability.
> Most languages with async do not provide this capability.
Those where async is syntax sugar for something else (e.g., JavaScript, Python, I forget if C# qualifies this way) allow hiding; you don't need annotation to call an async function, you only need it if you choose to use async syntax in the calling function.
> you don't need annotation to call an async function, you only need it if you choose to use async syntax in the calling function.
The proposed thing is "I want to get a function, and call it, and use the result, and if it is effectful in a way then I am effectful in that way, and if it is not then I am not" where "the result" is defined (among other things) as the value of the expression that appears in a return statement in the body of the function. Redefining "the result" to be some object representing a continuation or whatever is missing the point.
The lack of async-hiding in JavaScript is why the original "What Color is Your Function?" blog post exists.
But continuing on your idea, one colouring I would really like to see everywhere is pureness. C++'s `const` comes close, and perhaps Haskell does it right by having everything be (mostly) pure by default and using the type system to encode other properties. And yeah - different monads (like IO) are likely a good way to handle this whole topic in a user-definable way, but not many people like them (at least when they know they are using them, because unknowingly many programmer do use it, eg. most async syntactic sugar is basically a do notation for a single hard-coded Monad)
the async/non async functions in javascript are good and well designed. the whole colored functions thing is nonsense and it's weird that it gained so much traction on here
Those two groups are minorities in the set of all people that are even aware of monads.
There’s also Haskell’s system, which IIRC uses an IO monad.
https://www.academia.edu/71720554/A_Modular_Verifiable_Excep...
So here you go, http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p070...
As for being useless, well that is the usual argument from no exception, no RTTI folks, that would be better off using another language instead of messing up with C++'s design.
The advantage of something like JavaScript's async/await over Haskell's "do" is the fact because async/await is more limited than `do` (or yield and coroutines in JS) it's a lot easier to create tooling for.
I really don’t get how is Haskell “behind”.
What kind of tooling are you referring to?
Despite not liking the idea of function colors, I do think I like your idea of having functions express what capabilities you have to have permission for to call them. A long time ago, I saw someone talk about fine-grained permissions within a programming language, and I think they are the same idea.
I would love if that could be done without function colors, though.
In the "real code" I linked to in the post, I have a concept I call "context stacks". The idea comes from Jai, and I use to do implement things such as error handling and allocators in such a way that functions don't need to take an allocator argument. To make callees use your allocator, you push your allocator onto the allocator context stack and then call the functions you want to use that allocator.
I made it useful for many more things, and users could add their own.
In particular, I realized that it could be extended to where functions could get permissions (such as popping up an Admin dialog box on Windows or asking for sudo on Linux), push a capability [1] onto the capability context stack, and then call the functions that actually do the work. These functions could then grab that capability and do what they need to do.
One good thing about this is that permissions don't leak up to callers that should not have them. In other words, this system keeps permissions from infecting code that really shouldn't have them, but might need them because they might call functions that do.
I don't know if all of this made sense, sorry if not. I wrote it really fast.
[1]: https://en.wikipedia.org/wiki/Capability-based_security
Let’s take the example of JS. Does JS have function colors?
In one sense, yes, there are async functions and sync functions and you can only call an async function from an async function. So under a certain definition, JS functions are colorful.
But wait... a JS async function is just a normal function that returns a Promise object. So under another definition, JS functions are colorless. Perhaps someone might not like how Promises work in JS, but this has nothing to do with what color the functions are.
There are some languages that do break this and implement async functions as a completely different entity to normal functions. I believe this kind of function colors is lazy and bad language design and these async implementations tend to feel like they don’t belong with the rest of the language and were duct taped on.
Not only does this mean that I can express arbitrary constraints on callers, it also means that all constraints and stateful expectations are explicit, verifiable, and clear to the reader.
The post is interesting because it highlights that, colorless or not, functions have constraints and you can try to hide them and handle things automatically, via inference as is demonstrated here, but they're always there. Async is probably one of the less interesting constraints, but I find it so much more confusing when languages have colorless functions... I can't see suspension points, I can't tell wtf is going on. It's extremely frustrating.
> However, before I continue, let me say a little bit of praise for Zig: this is innovation. This is a piece of good work that, most of the time, async functions can be called the same as normal functions.
It's definitely cool and I like the definition of async they have. I think suspension is probably one of the less interesting things that a function can do, vs, say, IO, which is similarly "colorless" in languages. I just wish there were colors for all of these things.