I was completely baffled by "algebraic effects" for years. They looked far too confusing for me to want to spend my time on them, and took the "Don’t feel like you have to [get curious about them]" approach.
But then at some point it struck me: underlying all these effect systems is just passing stuff in. So I developed my own effect system for Haskell, Bluefin[1], based on capabilities, which means the "capability to perform some effect" is represented by just passing stuff in (that is, a function can do some effect as long as it has been passed the capability to do it).
From this point of view it's hard to understand the excitement over "resume with" and "the part you can’t do with try / catch. It lets us jump back to where we performed the effect, and pass something back to it from the handler". Programming languages have had that feature since forever: a "resumable exception" is a "function call". A dynamically chosen "resumable exception" is the call of a dynamically chosen function, i.e. the argument to a higher order function.
So I don't know why people love the complexity around "algebraic effects". Maybe the mystique has a certain allure. But if you want the most straightforward possible approach I can recommend you try out Bluefin. I'm happy to answer questions on the issue tracker[2].
(Caveat: Bluefin is able to simplify things dramatically by dropping support for "multi-shot" continuations. But mostly you don't want multi-shot continuations.)
A real effect system allows you to do things like NOT continue execution after using the effect (like the error effect does - if you "implement" this by using Exceptions, you're not using effects at all, just using Exceptions with extra steps) or only continuing it after some asynchronous work happens (the Future effect), or even "continue" execution several times. That just cannot be done with "just passing stuff in". You still don't seem to have understood effects.
Thanks for your response. Perhaps I'm missing some fundamental things. Could you help?
> A real effect system allows you to do things like NOT continue execution after using the effect
Right, Bluefin's Request allows you to do that too. For example here is an example of handling the request by continuing or not, depending on what the value yielded to the Request is.
example :: Either String ()
example = runPureEff $ try $ \ex -> do
forEach
( \r -> do
request r True
request r True
request r False
request r True
request r True
)
( \case
True -> pure ()
False -> throw ex "Stopped"
)
> if you "implement" this by using Exceptions, you're not using effects at all, just using Exceptions with extra steps
Not sure I follow that. Above you can see I used an exception (Bluefin's Throw capability), but I couldn't have used only an exception because that would have aborted unconditionally. What am I missing here, that makes "using Exceptions" "not using effects at all"?
> only continuing it after some asynchronous work happens (the Future effect)
I'm not really sure what "a Future effect" is, but I don't see how it's not something that can be run as a function call, at least in Haskell.
> or even "continue" execution several times
Right, these are the multishot continuations which Bluefin doesn't support. I haven't discovered many particularly compelling use cases for multishot continuations but would be very interested in finding some. The developer of the Kyo effect system for Scala, Flavio Brasil, suggested parsing, with multiple parse results, which makes sense.
I'm also not entirely sure Bluefin couldn't simulate common use cases of multishot continuations with threads, but I haven't thought about it very hard.
> You still don't seem to have understood effects.
Possibly true, and part of my puzzlement! I'm always happy to try to improve my understanding. Can you help me see what I've missed?
Equating "algebraic effects" with "continuations" is like saying "if" is just "goto" (which isn't even true, e.g., an if can turn into a cmov or whatever).
The only mystique around algebraic effects is the same mystique there is around monads. I don't know if people have started equating algebraic effects to burritos yet but that's a pretty good way to take something simple and turn it into something confusing.
> Equating "algebraic effects" with "continuations" is like saying "if" is just "goto"
Fair enough. But are you responding to something I said? I didn't make that equation.
> The only mystique around algebraic effects is the same mystique there is around monads. I don't know if people have started equating algebraic effects to burritos yet but that's a pretty good way to take something simple and turn it into something confusing.
Ah, are you saying that fundamentally there isn't really much to algebraic effects and they're much simpler than they're made out to be? If so then it perhaps we agree?
I don't think I said they're just continuations. In fact I'm trying to make the point that they're mostly just function calls (and I think in my career I've come across one case where I wanted something beyond function calls (for a constraint solver)). There are "multi-shot" continuations (whether you consider that interface or implementation I don't really mind), which have behaviour than function calls can't express, but I don't know of any algebraic effects beyond that.
What do you think algebraic effects are, if they're not continuations?
EDIT: Ah, based on your comment at https://news.ycombinator.com/item?id=48334737 you might say they're a feature of an intermediate language? So you might take a surface language and "compile to an intermediate language of lambda calculus + algebraic effects", without specifying how that intermediate language is implemented (because it may not even be implemented, per se).
They're really just a protocol. You can implement them in various ways. They will always be some sort of delimited continuations but a "function call" or continuation passing style or anything of the sort does not have to be involved at all.
For example, let us say I don't allow "multi-shot" continuations like in your library, and I'm implementing Algebraic Effects in my own interpreted language.
One way I can implement effects and handlers is to have a handlers get registered in a stack, then when an effect is triggered, save the IP and current stackframe, search for the right handler and jump to it. "resume" then just resets the stackframe, pushes a value into the stack, and sets the saved IP.
(Only saw your edit after posting, sorry, but yes)
Thanks, that clarifies where you're coming from. Is it possible to specify this protocol somehow, by defining an interface for it? Or by extending lambda calculus with the bits it needs?
(Maybe that's what the Koka folks do in their papers, and if so feel free to say, "yeah read their papers").
I'm thinking a less formally than that. Protocol in a very layman-y "perform is supposed to do this, resume is supposed to do this".
For example, Koka compiles handlers differently depending if they do multi-shot continuations or not. It can do this because all that matters is "perform is supposed to do this, resume is supposed to do this", not what they turn into (same as "if" turning into "cmov" in certain cases). I think it uses a continuation-passing style sort of implementation, but I can't quite remember.
Daan's libhandler implements effects for C in an entirely different manner. It captures the stack much like my example or a stackful coroutine library would.
I'm sure there a formal definitions in both the koka papers and the libhandler paper, but I just skim that stuff ;)
I submitted this because I've been getting really interested in effect systems, especially now that OCaml 5 has a working production quality example they'd been iterating on for years prior. I wanted to see what it'd look like in Rust too so maybe one day we can get rid of async function coloring, and with OxCaml by Jane Street maybe we could see how that would look in practice.
Another reason for submitting this is that React actually has a quite robust effect system, that people don't necessarily realize they're using one every day if they use hooks.
It’s worth clarifying that for the most part, this article just discusses plain effects, ie. "reified side effects" or "resumable exceptions". Algebraic effects are about the composition (ie. algebra) of effects, exactly like algebraic data types are about the composition of types. This part is generally not meaningful in an untyped language like Javascript because effects are all YOLO and you never know what’s going to happen, what effects a function might throw, or whether there’s any handler up-stack to catch your effect.
So this looks like dynamically scoped callbacks. Instead of passing callbacks along as parameters they are declared as “handlers”, and any function down the call stack can invoke them. Is this a correct understanding?
That's what I was thinking. You could get almost all of this pretty directly in Javascript by putting a callback function in an AsyncLocalStorage instance or, in other languages, in a thread local variable.
That's very interesting, thanks! It gave me a brainwave and I wondered I could implement that in Bluefin. I'm pretty sure Bluefin's Request[1] is a second class stackful coroutine, and sure enough it turns out to be possible, so I'm pleased about that.
-- ghci> example
-- Hello
-- World
-- Timed out
example = runEff $ \io -> awaitYield (receiver io) sender
receiver ::
(e1 <: es, e2 <: es) =>
IOE e1 ->
Await String e2 ->
Eff es ()
receiver io a = do
r1 <- await a
effIO io (putStrLn r1)
r2 <- await a
effIO io (putStrLn r2)
mr3 <- timeout io 0 (await a)
effIO io $ case mr3 of
Nothing -> putStrLn "Timed out"
Just r3 -> putStrLn r3
sender ::
e1 <: es =>
Yield String e1 ->
Eff es ()
sender y = do
yield y "Hello"
yield y "World"
yield y "More"
timeout ::
e1 <: es =>
IOE e1 ->
Int ->
Eff es r ->
Eff es (Maybe r)
timeout io t m = withEffToIO
(\effToIO -> System.Timeout.timeout t (effToIO (\_ -> useImpl m)))
io
As the article notes, effects (at least as described in the article) are already possible in JavaScript using generators, as long as every calling function is a generator, which I assume the author finds problematic because generators are cumbersome.
I was wondering how well TypeScript can type generator-based effects. My hypothesis is that TypeScript can let you compose functions with effects, but it is not possible for it to narrow down that the result from an effect corresponds to the effect (i.e. the result of an effect will be any possible effect result used by the function).
My incomplete, untested attempt in TypeScript[0] tries to implement `enumerateFiles` and `withMyLoggingLibrary`. The type errors demonstrate TypeScript's limitation that it can't associate an effect call with its result.
One of the things I find most exciting about effect systems is that they solve some of the issues with typeclasses and traits. The issue with these is you can provide exactly one implementation for a type, and this implementation then applies to your entire program. (The underlying theoretical issue, I guess, is coherence.) This means that code cannot request injection of behaviour _except_ through the type of its data.
In practice, what that means is you get a strong temptation to hang behaviour on data even when it doesn't fit perfectly. Because of the natural desire to reduce the number of entity definitions, you end up defining a typeclass on a data type that doesn't fit exactly just to get the behaviour to the right place without having to introduce a new policy or something.
Effects change this by essentially letting you provide multiple names implementations for the same data type, and you don't need to pass around a policy type because the polymorphism lets you tie handlers to scope.
So, if the fancy type-safe library-based control flow doesn't really do much for you, I think that their potential for code design is a good reason to still be excited!
CL conditions do what you actually need if you program. CL gives you deterministic state, safe resource management etc.
Nondeterminism is not a feature you want. Algebraic effects treat the execution stack (continuation) as data, you have total freedom over what you do with it. This flexibility is exactly where you get nondeterminism. This is how logic solvers or probabilistic algorithms work, but you don't want it as a programming language feature in general purpose programming language.
I can get behind the sentiment, but you absolutely need nondeterminism. You can separate the d from the non-d, but only Haskellish languages even attempt it. It's a coarse separation to make (IO vs non-IO), which is where effect systems come in - I guess you can categorise code into more descriptive buckets. The 'algebraic' part is currently beyond my knowledge.
It’s pretty annoying to comment “this thing was already present in CL 30 years ago”, then to have someone correct you (pointing out that this is a common misconception, and algebraic effects are not equivalent to conditions at all), and then to respond not to admit you’re wrong but instead to say “algebraic are bad anyway, you shouldn’t want nondeterminism”.
By the way, nondeterminism is not the only difference between the two.
Yes, this article is doing a bad job at explaining why you would want effects, and one of the main advantages is exactly that it becomes part of the type system, essentially coloring every single function with a set of effects it needs to be called. As the article used JavaScript it shows what untyped effects would look like, which in my opinion is awful.
If you want to use algebraic effects today, I highly recommend Unison. If you’re on the JVM, Flix is doing major advances with effects!
When you need to use an effect, you need it in the type. If you directly call a function using some other effect, it propagates into your function. So far, so colourful.
But you can have generic effects. Your arguments and return type can specify "any effect", indicating your function can use a type with any effect safely, or can be used in any effect context safely.
Passing an async value to a function doesn't mean that function must now also be an async function. It can be a "for all effects, do the thing" function. The code duplication problem is gone.
No, they are function colouring. That's the point.
Someone writes a post lamenting red and blue functions, and everyone eats it up.
Substitute colour for something meaningful and the idea becomes idiotic.
"Top level function declares that it is non-blocking, but when I try to call a small blocking function from it, I have to change the declaration to blocking???"
How do you handle logging then? If f() calls g(), how can I add logging to g() without having to change or recompile f() (and everything in the call stack above it)? ‘You can’t’ is not an acceptable answer.
Not sure why people are saying "you can't" when it seems to me the whole point of algebraic effects that you can. You can define g so that it has no ability to do "general IO", all it can do is yield log messages. Then f can call g in a way that turns the log messages into writes to stdout. For example, here's how you would do it in Bluefin:
type Log = Yield String
-- workWithLogging cannot do arbitrary IO!
-- All it can do is yield log messages, which
-- must be processed elsewhere.
workWithLogging ::
(e1 :> es) =>
Log e1 ->
Int ->
Int ->
Eff es Int
workWithLogging l x y = do
yield l ("x was " <> show x)
yield l ("y was " <> show y)
let result = x + y
yield l ("result was " <> show result)
pure result
-- ghci> example
-- x was 5
-- y was 7
-- result was 12
-- 12
example :: IO Int
example = runEff $ \io -> do
-- forEach determines how each log message
-- should be handled.
forEach
(\l-> workWithLogging l 5 7)
(\logMsg -> effIO io (putStrLn logMsg))
I’ve used effects in scala3 with cats-effects and haven’t been impressed. All in all, the way it was used was just to reimplement a very similar interface to exceptions.
Cats Effect is monadic effects. What is discussed here is sometimes called "direct-style" effects, and is an alternative representation.
I think you've missed the point regarding effect systems. Concurrency and resource handling, implemented in a way that is composable and reasonably easy to reason about, are two of the big ticket features.
Everything he lists is solved by effect-ts [1] bar, obviously, the language support (effect has its own fiber-based runtime like ZIO's scala).
I've been using it for 5+ years and my 4 men team can scale to supporting 6 different products (each running millions $ in business, sometimes daily), as we reuse the same patterns and architecture. This would not be possible without Effect, even though I'm lucky to have terrific engineers as colleagues, we just wouldn't be able to without the endless goodies from Effect.
The amount of features is basically endless, as effects and runtimes weren't enough, from SQL to AI, from effectful schemas (encoders/decoders), first-class OTEL support, CLI, debuggers, editor extensions, and many others. There's still countless modules I have yet to see or use.
Runtimes are available for each platform, including cloudflare workers.
There's absolutely nothing in TypeScript land to have such a wide scope.
v4 will also bring durable workflows (I'm already using v4 beta and that feature in prod) and many other goodies. That's quite important for us needing to have procedures that need to survive redeploys, crashes, etc.
I would never go back to writing standard TypeScript.
There is a learning curve, but you can adopt it incrementally. Nobody adopting it has ever gone back.
That being said, it would be great if there was a proper effect-based language (I've seen few projects like Effekt, but there's way too many things missing) as TypeScript is verbose, and effect adds its own verbosity.
When I started a typescript server project, I spent quite some time
- setting up dependency injection framework to pass context (tsyringe)
- finding a good schema validation library (zod)
- use neverthrow for typed error.
- wrote a few withSpan, recordDuration OTEL helper functions.
Then I heard about Effect-ts, checked the doc, and realize that Effect-ts already has all these things, in a single package.
IMO Effect-ts is currently the most practical effect system right now. While it does not support resuming like in other algebraic effect languages, it is powerful enough to express common patters, but not too powerful so that the code becomes hard to understand.
I hope effect-ts gets more traction. The biggest obstacle to sell it right now is that the API doc is not great. I had to trace the source code several times just to see how a type is defined. I hope Effect-ts team is aware that more people will use it if it has a proper documentation.
Effect is pretty nice, I'm not sure how worth it it is for the frontend, but I've heard good things on the backend, but sadly I don't use TypeScript for backend work, mainly Rust, and would love to see something like that there. I'm not sure how much Rust's type system would make it possible though however.
I know parts of Effect like its schema are incrementally adoptable but if you use it substantially with many of its features, isn't it viral in a sense? In that you need to do things the Effect way and wrap libraries into Effect functions?
It does tend to naturally bubble upwards as you point out, but you can decide where to stop.
E.g. you could describe a complex effect that has retry, scheduling, etc and run it only once with `Effect.runFork(yourEffect)` in a random place of your existing code.
That's in general how teams adopt it, in general there's a champion in the team that sells using one feature, and as people get accustomed and the champion does a good work mentoring it slowly takes over whole projects.
The learning curve is quite steep but once you get it you become effect pilled. You can completely separate the WHAT the application does from HOW it will do it.
I swear I'm not trying to be inflamatory, but this is the _worst_ programming language feature I could ever imagine. I'm not trying to be hyperbolic, if I try to reason about it there is nothing I can come up with that I would dislike more in the realm of recent features that have been pitched in the PL community
I was already in the camp that try/catch is "considered harmful", I dislike the concept of having a second, hidden, control flow that might get sprung up upon function callers, because it has side effects buried in the implementation of a callee that are not defined in the parameters or the returns, and I am not 100% sold on the benefits of "Things in the middle don’t need to concern themselves with error handling.", which I guess informs this opinion.
Now since I hate that, I really, really would hate that on top of this, another programmer could write a hidden control flow upstairs that could, potentially, not just crash my code, but also do a lot of other things, such as coming up with default values for unexpected NULLs or whatever, which could THEN take something that would have crashed immediately, and turn it into something that crashes later down the line, away from the problem, with a varialble set to an inexplicable value that I have never put there myself
I agree but I also think it's important to point out that Algebraic Effects typically refers both to a runtime feature (the try/handle delineated continuation stuff) and also a type system feature.
The latter is very important because in your example it would not really be hidden. If your function does not have the "exn" effect, it cannot call functions that throw exceptions, full stop. Same with any other effect including IO if you want.
Basically function coloring taken to the extreme. In a statically typed language with statically typed effects you actually cannot get surprised, which was your major complaint.
Type systems that support algebraic effects also typically support row polymorphic effects (fancy generics) so you can make a function generic over "color", avoiding the "function coloring" problem.
Now, having said that, why did I say I agree with you? Well because algebraic effects are a lousy user-facing feature. You almost never want to implement your own handlers, at best you'll plug in a custom handler for the IO effect and that's about it.
Where they shine is for the language implementer. They provide a framework on which exceptions, generators, async/await and even prolog-like backtracking can be implemented, while (very importantly) defining how they compose. That's really the bit that makes them so interesting from a research point of view and why they might make it into the mainstream languages, even if the language doesn't actually ever expose them for you to use.
> I dislike the concept of having a second, hidden, control flow that might get sprung up upon function callers, because it has side effects buried in the implementation of a callee that are not defined in the parameters or the returns
You might like my capability-based effect system for Haskell, Bluefin[1], then. If a Bluefin effectful function throws you can see it in the type system. If you want to have the capability to throw, you need to pass in an argument of type Throw. For example here "workWithThrow" can only throw an exception because it is passed the Throw capability.
workWithThrow ::
(e1 :> es) =>
Throw String e1 ->
Int ->
Int ->
Eff es Int
workWithThrow t x y = do
let result = x + y
when (result > 10) $ do
throw t "Too big"
pure result
-- ghci> example
-- Left "Too big"
example :: Either String Int
example = runPureEff $ try $ \t -> do
workWithThrow t 5 7
Because you must statically declare dependency on an effect, it's opt-in.
To be clear, you still pay indirection cost: when you do opt in you have to hope the upstairs implementation is compliant to the contract. But that does also apply to interfaces/typeclasses.
I was completely baffled by "algebraic effects" for years. They looked far too confusing for me to want to spend my time on them, and took the "Don’t feel like you have to [get curious about them]" approach.
But then at some point it struck me: underlying all these effect systems is just passing stuff in. So I developed my own effect system for Haskell, Bluefin[1], based on capabilities, which means the "capability to perform some effect" is represented by just passing stuff in (that is, a function can do some effect as long as it has been passed the capability to do it).
From this point of view it's hard to understand the excitement over "resume with" and "the part you can’t do with try / catch. It lets us jump back to where we performed the effect, and pass something back to it from the handler". Programming languages have had that feature since forever: a "resumable exception" is a "function call". A dynamically chosen "resumable exception" is the call of a dynamically chosen function, i.e. the argument to a higher order function.
So I don't know why people love the complexity around "algebraic effects". Maybe the mystique has a certain allure. But if you want the most straightforward possible approach I can recommend you try out Bluefin. I'm happy to answer questions on the issue tracker[2].
(Caveat: Bluefin is able to simplify things dramatically by dropping support for "multi-shot" continuations. But mostly you don't want multi-shot continuations.)
EDIT: I was too pessimistic, bazoom42 has noticed this :) https://news.ycombinator.com/item?id=48334067
[1] https://hackage.haskell.org/package/bluefin
[2] https://github.com/tomjaguarpaw/bluefin/issues/new
A real effect system allows you to do things like NOT continue execution after using the effect (like the error effect does - if you "implement" this by using Exceptions, you're not using effects at all, just using Exceptions with extra steps) or only continuing it after some asynchronous work happens (the Future effect), or even "continue" execution several times. That just cannot be done with "just passing stuff in". You still don't seem to have understood effects.
Thanks for your response. Perhaps I'm missing some fundamental things. Could you help?
> A real effect system allows you to do things like NOT continue execution after using the effect
Right, Bluefin's Request allows you to do that too. For example here is an example of handling the request by continuing or not, depending on what the value yielded to the Request is.
> if you "implement" this by using Exceptions, you're not using effects at all, just using Exceptions with extra stepsNot sure I follow that. Above you can see I used an exception (Bluefin's Throw capability), but I couldn't have used only an exception because that would have aborted unconditionally. What am I missing here, that makes "using Exceptions" "not using effects at all"?
> only continuing it after some asynchronous work happens (the Future effect)
I'm not really sure what "a Future effect" is, but I don't see how it's not something that can be run as a function call, at least in Haskell.
> or even "continue" execution several times
Right, these are the multishot continuations which Bluefin doesn't support. I haven't discovered many particularly compelling use cases for multishot continuations but would be very interested in finding some. The developer of the Kyo effect system for Scala, Flavio Brasil, suggested parsing, with multiple parse results, which makes sense.
I'm also not entirely sure Bluefin couldn't simulate common use cases of multishot continuations with threads, but I haven't thought about it very hard.
> You still don't seem to have understood effects.
Possibly true, and part of my puzzlement! I'm always happy to try to improve my understanding. Can you help me see what I've missed?
Equating "algebraic effects" with "continuations" is like saying "if" is just "goto" (which isn't even true, e.g., an if can turn into a cmov or whatever).
The only mystique around algebraic effects is the same mystique there is around monads. I don't know if people have started equating algebraic effects to burritos yet but that's a pretty good way to take something simple and turn it into something confusing.
> Equating "algebraic effects" with "continuations" is like saying "if" is just "goto"
Fair enough. But are you responding to something I said? I didn't make that equation.
> The only mystique around algebraic effects is the same mystique there is around monads. I don't know if people have started equating algebraic effects to burritos yet but that's a pretty good way to take something simple and turn it into something confusing.
Ah, are you saying that fundamentally there isn't really much to algebraic effects and they're much simpler than they're made out to be? If so then it perhaps we agree?
We disagree that they're just continuations (only one of many possible implementation strategies) but agree they're nothing special ;)
I don't think I said they're just continuations. In fact I'm trying to make the point that they're mostly just function calls (and I think in my career I've come across one case where I wanted something beyond function calls (for a constraint solver)). There are "multi-shot" continuations (whether you consider that interface or implementation I don't really mind), which have behaviour than function calls can't express, but I don't know of any algebraic effects beyond that.
What do you think algebraic effects are, if they're not continuations?
EDIT: Ah, based on your comment at https://news.ycombinator.com/item?id=48334737 you might say they're a feature of an intermediate language? So you might take a surface language and "compile to an intermediate language of lambda calculus + algebraic effects", without specifying how that intermediate language is implemented (because it may not even be implemented, per se).
They're really just a protocol. You can implement them in various ways. They will always be some sort of delimited continuations but a "function call" or continuation passing style or anything of the sort does not have to be involved at all.
For example, let us say I don't allow "multi-shot" continuations like in your library, and I'm implementing Algebraic Effects in my own interpreted language.
One way I can implement effects and handlers is to have a handlers get registered in a stack, then when an effect is triggered, save the IP and current stackframe, search for the right handler and jump to it. "resume" then just resets the stackframe, pushes a value into the stack, and sets the saved IP.
(Only saw your edit after posting, sorry, but yes)
> They're really just a protocol.
Thanks, that clarifies where you're coming from. Is it possible to specify this protocol somehow, by defining an interface for it? Or by extending lambda calculus with the bits it needs?
(Maybe that's what the Koka folks do in their papers, and if so feel free to say, "yeah read their papers").
I'm thinking a less formally than that. Protocol in a very layman-y "perform is supposed to do this, resume is supposed to do this".
For example, Koka compiles handlers differently depending if they do multi-shot continuations or not. It can do this because all that matters is "perform is supposed to do this, resume is supposed to do this", not what they turn into (same as "if" turning into "cmov" in certain cases). I think it uses a continuation-passing style sort of implementation, but I can't quite remember.
Daan's libhandler implements effects for C in an entirely different manner. It captures the stack much like my example or a stackful coroutine library would.
I'm sure there a formal definitions in both the koka papers and the libhandler paper, but I just skim that stuff ;)
> Protocol in a very layman-y "perform is supposed to do this, resume is supposed to do this".
OK, but at the very least it has two primitives "perform" and "resume"? And they're supposed to interact in some particular way?
Continuations (“function calls”) reduce the amount of optimisations you can do around memory - both space wise and computationally
Not sure I follow what you're trying to say. Can you elaborate?
Oh neat, I'm in the second chance pool.
I submitted this because I've been getting really interested in effect systems, especially now that OCaml 5 has a working production quality example they'd been iterating on for years prior. I wanted to see what it'd look like in Rust too so maybe one day we can get rid of async function coloring, and with OxCaml by Jane Street maybe we could see how that would look in practice.
Another reason for submitting this is that React actually has a quite robust effect system, that people don't necessarily realize they're using one every day if they use hooks.
(2019)
It’s worth clarifying that for the most part, this article just discusses plain effects, ie. "reified side effects" or "resumable exceptions". Algebraic effects are about the composition (ie. algebra) of effects, exactly like algebraic data types are about the composition of types. This part is generally not meaningful in an untyped language like Javascript because effects are all YOLO and you never know what’s going to happen, what effects a function might throw, or whether there’s any handler up-stack to catch your effect.
[dead]
So this looks like dynamically scoped callbacks. Instead of passing callbacks along as parameters they are declared as “handlers”, and any function down the call stack can invoke them. Is this a correct understanding?
That's what I was thinking. You could get almost all of this pretty directly in Javascript by putting a callback function in an AsyncLocalStorage instance or, in other languages, in a thread local variable.
You need more than that for the example with setTimeout(). It requires to be able to freeze the stack and then go back later.
You need stackful coroutine (like goroutine) for that.
That's very interesting, thanks! It gave me a brainwave and I wondered I could implement that in Bluefin. I'm pretty sure Bluefin's Request[1] is a second class stackful coroutine, and sure enough it turns out to be possible, so I'm pleased about that.
Yes. dynamically scoped, and statically typed.
As the article notes, effects (at least as described in the article) are already possible in JavaScript using generators, as long as every calling function is a generator, which I assume the author finds problematic because generators are cumbersome.
I was wondering how well TypeScript can type generator-based effects. My hypothesis is that TypeScript can let you compose functions with effects, but it is not possible for it to narrow down that the result from an effect corresponds to the effect (i.e. the result of an effect will be any possible effect result used by the function).
My incomplete, untested attempt in TypeScript[0] tries to implement `enumerateFiles` and `withMyLoggingLibrary`. The type errors demonstrate TypeScript's limitation that it can't associate an effect call with its result.
[0]: https://www.typescriptlang.org/play/?#code/C4TwDgpgBAMg9gcwK...
One of the things I find most exciting about effect systems is that they solve some of the issues with typeclasses and traits. The issue with these is you can provide exactly one implementation for a type, and this implementation then applies to your entire program. (The underlying theoretical issue, I guess, is coherence.) This means that code cannot request injection of behaviour _except_ through the type of its data.
In practice, what that means is you get a strong temptation to hang behaviour on data even when it doesn't fit perfectly. Because of the natural desire to reduce the number of entity definitions, you end up defining a typeclass on a data type that doesn't fit exactly just to get the behaviour to the right place without having to introduce a new policy or something.
Effects change this by essentially letting you provide multiple names implementations for the same data type, and you don't need to pass around a policy type because the polymorphism lets you tie handlers to scope.
So, if the fancy type-safe library-based control flow doesn't really do much for you, I think that their potential for code design is a good reason to still be excited!
For some funsie here's my fully working delimited continuation in C with effect handler example: https://godbolt.org/z/3ehehvo6E
No ASM involved so technically portable (although it depends on built-in).
Flix equivalent (copy paste to https://play.flix.dev/):
Pardon my ignorance, but isn’t this more or less a fancy goto?
Standard feature of Common Lisp condition system (over 30 years old).
https://lisp-docs.github.io/cl-language-reference/chap-9/j-b...
The CL condition system always gets brought up when people unfamiliar with effects see effects for the first time (example: https://news.ycombinator.com/item?id=38813484, another example: https://lobste.rs/s/12m2f0/algebraic_effects_another_mistake).
But while the condition system can do many things you can also do with effects, they cannot do everything.
Here's another discussion on this: https://news.ycombinator.com/item?id=44078743
CL conditions do what you actually need if you program. CL gives you deterministic state, safe resource management etc.
Nondeterminism is not a feature you want. Algebraic effects treat the execution stack (continuation) as data, you have total freedom over what you do with it. This flexibility is exactly where you get nondeterminism. This is how logic solvers or probabilistic algorithms work, but you don't want it as a programming language feature in general purpose programming language.
> Nondeterminism is not a feature you want.
I can get behind the sentiment, but you absolutely need nondeterminism. You can separate the d from the non-d, but only Haskellish languages even attempt it. It's a coarse separation to make (IO vs non-IO), which is where effect systems come in - I guess you can categorise code into more descriptive buckets. The 'algebraic' part is currently beyond my knowledge.
It’s pretty annoying to comment “this thing was already present in CL 30 years ago”, then to have someone correct you (pointing out that this is a common misconception, and algebraic effects are not equivalent to conditions at all), and then to respond not to admit you’re wrong but instead to say “algebraic are bad anyway, you shouldn’t want nondeterminism”.
By the way, nondeterminism is not the only difference between the two.
Do effect systems actually avoid colored functions? Don’t most typed effect systems require the used effects in the signature?
Yes, this article is doing a bad job at explaining why you would want effects, and one of the main advantages is exactly that it becomes part of the type system, essentially coloring every single function with a set of effects it needs to be called. As the article used JavaScript it shows what untyped effects would look like, which in my opinion is awful. If you want to use algebraic effects today, I highly recommend Unison. If you’re on the JVM, Flix is doing major advances with effects!
https://www.unison-lang.org/
https://flix.dev/
When you need to use an effect, you need it in the type. If you directly call a function using some other effect, it propagates into your function. So far, so colourful.
But you can have generic effects. Your arguments and return type can specify "any effect", indicating your function can use a type with any effect safely, or can be used in any effect context safely.
Passing an async value to a function doesn't mean that function must now also be an async function. It can be a "for all effects, do the thing" function. The code duplication problem is gone.
No, they are function colouring. That's the point.
Someone writes a post lamenting red and blue functions, and everyone eats it up.
Substitute colour for something meaningful and the idea becomes idiotic.
"Top level function declares that it is non-blocking, but when I try to call a small blocking function from it, I have to change the declaration to blocking???"
Yes, yes you do.
Total functions can't call non-total functions.
Deterministic functions can't call nondeterministic functions.
Non-IO functions can't call IO functions.
‘Non-IO functions can't call IO functions.’
How do you handle logging then? If f() calls g(), how can I add logging to g() without having to change or recompile f() (and everything in the call stack above it)? ‘You can’t’ is not an acceptable answer.
Not sure why people are saying "you can't" when it seems to me the whole point of algebraic effects that you can. You can define g so that it has no ability to do "general IO", all it can do is yield log messages. Then f can call g in a way that turns the log messages into writes to stdout. For example, here's how you would do it in Bluefin:
"You can't" is simpler, because the inevitable reply is "but how do I do actual logging inside g"
"Actual logging" as in direct access to IO?
Yes
Don't declare it as non-logging.
You can’t is an acceptable answer. The entitle point of such a feature is to prevent people from doing that.
Previously discussed (with pretty decent comments) https://news.ycombinator.com/item?id=20496043
I’ve used effects in scala3 with cats-effects and haven’t been impressed. All in all, the way it was used was just to reimplement a very similar interface to exceptions.
Cats Effect is monadic effects. What is discussed here is sometimes called "direct-style" effects, and is an alternative representation.
I think you've missed the point regarding effect systems. Concurrency and resource handling, implemented in a way that is composable and reasonably easy to reason about, are two of the big ticket features.
Everything he lists is solved by effect-ts [1] bar, obviously, the language support (effect has its own fiber-based runtime like ZIO's scala).
I've been using it for 5+ years and my 4 men team can scale to supporting 6 different products (each running millions $ in business, sometimes daily), as we reuse the same patterns and architecture. This would not be possible without Effect, even though I'm lucky to have terrific engineers as colleagues, we just wouldn't be able to without the endless goodies from Effect.
The amount of features is basically endless, as effects and runtimes weren't enough, from SQL to AI, from effectful schemas (encoders/decoders), first-class OTEL support, CLI, debuggers, editor extensions, and many others. There's still countless modules I have yet to see or use.
Runtimes are available for each platform, including cloudflare workers.
There's absolutely nothing in TypeScript land to have such a wide scope.
v4 will also bring durable workflows (I'm already using v4 beta and that feature in prod) and many other goodies. That's quite important for us needing to have procedures that need to survive redeploys, crashes, etc.
I would never go back to writing standard TypeScript.
There is a learning curve, but you can adopt it incrementally. Nobody adopting it has ever gone back.
That being said, it would be great if there was a proper effect-based language (I've seen few projects like Effekt, but there's way too many things missing) as TypeScript is verbose, and effect adds its own verbosity.
[1] https://effect.website/
When I started a typescript server project, I spent quite some time - setting up dependency injection framework to pass context (tsyringe) - finding a good schema validation library (zod) - use neverthrow for typed error. - wrote a few withSpan, recordDuration OTEL helper functions.
Then I heard about Effect-ts, checked the doc, and realize that Effect-ts already has all these things, in a single package.
IMO Effect-ts is currently the most practical effect system right now. While it does not support resuming like in other algebraic effect languages, it is powerful enough to express common patters, but not too powerful so that the code becomes hard to understand.
I hope effect-ts gets more traction. The biggest obstacle to sell it right now is that the API doc is not great. I had to trace the source code several times just to see how a type is defined. I hope Effect-ts team is aware that more people will use it if it has a proper documentation.
Effect is pretty nice, I'm not sure how worth it it is for the frontend, but I've heard good things on the backend, but sadly I don't use TypeScript for backend work, mainly Rust, and would love to see something like that there. I'm not sure how much Rust's type system would make it possible though however.
I know parts of Effect like its schema are incrementally adoptable but if you use it substantially with many of its features, isn't it viral in a sense? In that you need to do things the Effect way and wrap libraries into Effect functions?
It does tend to naturally bubble upwards as you point out, but you can decide where to stop.
E.g. you could describe a complex effect that has retry, scheduling, etc and run it only once with `Effect.runFork(yourEffect)` in a random place of your existing code.
That's in general how teams adopt it, in general there's a champion in the team that sells using one feature, and as people get accustomed and the champion does a good work mentoring it slowly takes over whole projects.
The learning curve is quite steep but once you get it you become effect pilled. You can completely separate the WHAT the application does from HOW it will do it.
[dead]
I swear I'm not trying to be inflamatory, but this is the _worst_ programming language feature I could ever imagine. I'm not trying to be hyperbolic, if I try to reason about it there is nothing I can come up with that I would dislike more in the realm of recent features that have been pitched in the PL community
I was already in the camp that try/catch is "considered harmful", I dislike the concept of having a second, hidden, control flow that might get sprung up upon function callers, because it has side effects buried in the implementation of a callee that are not defined in the parameters or the returns, and I am not 100% sold on the benefits of "Things in the middle don’t need to concern themselves with error handling.", which I guess informs this opinion.
Now since I hate that, I really, really would hate that on top of this, another programmer could write a hidden control flow upstairs that could, potentially, not just crash my code, but also do a lot of other things, such as coming up with default values for unexpected NULLs or whatever, which could THEN take something that would have crashed immediately, and turn it into something that crashes later down the line, away from the problem, with a varialble set to an inexplicable value that I have never put there myself
What a nightmare to debug! I mean, come on
I agree but I also think it's important to point out that Algebraic Effects typically refers both to a runtime feature (the try/handle delineated continuation stuff) and also a type system feature.
The latter is very important because in your example it would not really be hidden. If your function does not have the "exn" effect, it cannot call functions that throw exceptions, full stop. Same with any other effect including IO if you want.
Basically function coloring taken to the extreme. In a statically typed language with statically typed effects you actually cannot get surprised, which was your major complaint.
Type systems that support algebraic effects also typically support row polymorphic effects (fancy generics) so you can make a function generic over "color", avoiding the "function coloring" problem.
Now, having said that, why did I say I agree with you? Well because algebraic effects are a lousy user-facing feature. You almost never want to implement your own handlers, at best you'll plug in a custom handler for the IO effect and that's about it.
Where they shine is for the language implementer. They provide a framework on which exceptions, generators, async/await and even prolog-like backtracking can be implemented, while (very importantly) defining how they compose. That's really the bit that makes them so interesting from a research point of view and why they might make it into the mainstream languages, even if the language doesn't actually ever expose them for you to use.
> I dislike the concept of having a second, hidden, control flow that might get sprung up upon function callers, because it has side effects buried in the implementation of a callee that are not defined in the parameters or the returns
You might like my capability-based effect system for Haskell, Bluefin[1], then. If a Bluefin effectful function throws you can see it in the type system. If you want to have the capability to throw, you need to pass in an argument of type Throw. For example here "workWithThrow" can only throw an exception because it is passed the Throw capability.
[1] https://hackage.haskell.org/package/bluefinBecause you must statically declare dependency on an effect, it's opt-in.
To be clear, you still pay indirection cost: when you do opt in you have to hope the upstairs implementation is compliant to the contract. But that does also apply to interfaces/typeclasses.
[dead]
The article discusses algebraic effects but React is mentioned. Don’t make the mistake of saying thinking React is functional. See https://mckoder.medium.com/why-react-is-not-functional-b1ed1...