Hacker News new | ask | show | jobs
by tikhonj 3924 days ago
The reason that the Monad structure is interesting, in my view, is that it's simple and neatly captures the notion of a "computation" that we can compose in different ways.

The core useful operator for monads in Haskell is >>=. It has the following type:

    m a -> (a -> m b) -> m b
If you squint, this is like function application with a few extra m's thrown in. Here's normal function application for comparison:

    a -> (a -> b) -> b
So what is this extra m useful for? It's like a hole where we get to plug in some custom logic. In a sense, it lets us change what it means to "apply" a "function". This turns out to be useful for a whole bunch of things, not just IO. (Honestly, from a pedagogical standpoint, I think IO is a bit of a distraction!)

For example, take the Maybe type. It's Haskell's nullable: a Maybe a means you either have an a or Nothing.

    data Maybe a = Just a | Nothing
Remembering the signature of >>= above, what's the most natural way to implement "application"? If we break the function into cases, it becomes pretty straightforward. Here's the specialized function signature we want to implement:

    (>>=) :: Maybe a -> (a -> Maybe b) -> Maybe b
    x >>= f = ...
If x is Nothing then we don't have anything to pass into f so the whole result has to be Nothing. If x has a value, we can just get that value out and pass it into f normally, returning the final result.

    Nothing >>= f = Nothing
    Just x  >>= f = f x
(In case you're not familiar with Haskell syntax, the above is actually a valid definition of (>>=) for Maybe!)

For Maybe, being a monad gives us a standard way of working with values while automatically dealing with Nothing. It abstracts over repetitive null checking and lets us easily build up Maybe values based on other Maybe values.

Other examples of monads are the same in spirit. The list monad, for example, lets us handle any number of inputs in a way that's similar to Maybe. The State monad similarly lets us combine values while carrying along an implicit state internally.

The rest of the Monad structure (namely the return function and the laws) are just a formal way of codifying behavior behavior that's already intuitive.

So how does this all apply to doing input and output?

Well, the problem in Haskell is that it's a language of evaluating expressions at heart: executing effects makes no sense any more than it would in arithmetic. To work with effects we instead have a special, opaque type IO; normal expressions get evaluated to IO actions that can be run to produce the desired effect—namely the IO type.

Critically, the IO type does not have to be a monad. It could be completely self-contained and have custom functions for doing one action after the other. We could imagine something like:

    after :: IO a -> IO b -> IO b
which would let you run an IO statement then run a second one and only return the value of the last one. It's like an imperative block of code!

However, we would also like some way of using the results of an IO statement, perhaps assigning them to a name. We can't do this normally because the IO statements are run separately from expression evaluation. We'd have to have some sort of function that could take an IO value, unwrap it and do something with it. And how would we express an interface like this? With a normal function!

    doSomething :: IO a -> (a -> IO b) -> IO b
Hey, doesn't that look familiar? It's exactly (>>=)!

I'm hand-waving a bit again, but the rest of the monad structure comes up when you try to make sure after behaves consistently and intuitively.

So IO being a monad emerges naturally from the desire to be able to compose actions and depend on their results in a way that's separate from normal variable bindings and expression evaluation.

The causation here is important: it's not that IO is a monad, but rather the IO type (which could exist on its own) happens to naturally and usefully form a monad. But it does a lot of other things too, including some specific capabilities (like spawning threads) that are hard to generalize.

So my point, I suppose, is twofold: monads are useful for combining some notion of computation and IO happens to be an interesting example, but the fact that we wrap statements with external effects in a custom type called IO does not inextricably depend on the idea of a monad.

Did that explanation help? I wrote a blog post on a similar topic that might be interesting too: http://jelv.is/blog/Haskell-Monads-and-Purity

4 comments

> Well, the problem in Haskell is that it's a language of evaluating expressions at heart: executing effects makes no sense any more than it would in arithmetic.

You know what else doesn't make sense in arithmetic? Computation. I think it is important to wonder how come mathematical definitions of computation didn't come about until the 20th century[1] even though math has had most of its big breakthroughs (to date) by then (or, at least, there was no revolution in mathematics in the 20th century on the same scale as there was in physics). And yet, as advanced as math was, computation was only defined very late in the game. The reason is that, perhaps ironically, the concept of computation is absent from most of math, which is "equational".

But computation is the very core of computer science. So, for example, in arithmetic, 4 and 2+2 are the same thing because 4 = 2 + 2. So in arithmetic, equality means "sameness". But in computer science they are most certainly not the same, because the process of moving from one of the representation to the other is precisely what a computation does. Not only that, the process isn't even necessarily symmetrical, as moving in one direction may entail a much higher computational complexity than going in the other direction.

Any distinction between calculations and effect (other than actual IO) is, therefore, arbitrary. Whether or not the particular distinctions Haskell makes (where waiting for one second may or may not be an effect depending on how it is implemented) in the form Haskell makes it -- using pure functions and equational reasoning and describing effects with monads -- is actually useful and beneficial is a question for psychologists. Given Haskell's design, monads are certainly useful in Haskell, but that utility can't be generalized to languages with other designs.

[1]: https://en.wikipedia.org/wiki/Computability_theory

Computation in mathematics is well represented in the constructive and intuitionistic camps. It was thoroughly examined in the 1930s via the invention of things like lambda calculus and turing machines. It was also subsequently rejected by most mathematicians not because it fails to be equational but just because they decided that they don't really care.

Including notions of computation makes equational theories hairier. That might be very much what you want, and there are plenty of ways in which mathematics talks about (2 + 2) being unequal to (4). Typically you instead replace equality with weaker notions going all the way down to the weakest (most intricate) examples of actual algorithmic computation.

So why avoid this kind of stuff? Because you want to pick a "cognitive abstraction", in your terms, which is most useful. It turns out that for a lot of purposes inclusion of computation is too hairy and obscures what people care about. It also turns out that in many cases what they care about is somewhat unaffected by the computational substrate involved.

I really can't say I agree that your beef with equational reasoning makes much sense to me, but I think it's just because you're shooting too broadly. You can easily have problems with the particular choice of granularity of equating taken by the Haskell community and language... but that's just a much smaller thing.

> It was thoroughly examined in the 1930s

That's what I said. Only in the 20th century for something that might seem to an outsider to be fundamental to mathematics (after all, mathematicians sometimes compute, right?)

> So why avoid this kind of stuff? Because you want to pick a "cognitive abstraction", in your terms, which is most useful.

Oh, absolutely, but there are many ways and many forms of reconciling "equational" mathematics and computational mathematics, and PFP is just one of them. As I've shown in another comment, Prolog takes this much further than Haskell because it's truly equational (or, rather, relational), while Haskell gladly allows a direct (one-way) computation if it defines it as pure, i.e. in Haskell 2+2=>4 is still as different from 4=>2+2 as it is in Java. Other ways are the plain procedural style (which simply lets you name and parameterize a computation), and my favorite (although I don't know how practical), temporal logic used in, e.g. Esterel (and TLA+).

My beef is not with equational reasoning as a concept, and not even with Haskell's flavor of it. Certainly "equationality" in programming languages is a spectrum (as computation is inherently non-equational), and finding useful points on it is a matter for lots of empirical study. My "beef" (if that's what you want to call it) is with the claim that Haskell's particular design is somehow more fundamental or mathematical than other designs, and, to bring us back to the topic of the discussion, that (cognitive) monads are somehow essential. The reality is that there are many (good) justifications for Haskell's design, but they've been made by people to match certain human aesthetics and certain human goals[1] and they don't have any transcendent mathematical justification. They are just a point in the design space. Similarly, monads are only (pretty much) essential in Haskell given its design. Some people try to present Haskell (and monads) as the language (and abstraction) God, nay, better -- Math -- has given us, and that's just wrong. It's just a language like many others, with its own design choices. Besides, everybody knows God uses Lisp. My other "beef" is with the unjustified belief that a certain way of expressing an algorithm is automatically cognitively optimal just by virtue of it being equational (be it either in the Haskell or the Prolog sense).

[1]: One of them is a particular balance between correctness and specification effort, which, again, is also arbitrary and defined by the limitations of HM type inference.

P.S.

To clarify something I wrote in another comment, Haskell's definition of purity, while fundamentally arbitrary (from a computational theory point-of-view), makes a lot of sense if you've chosen that particular design.

Ahhh, well if you're taking aim at Haskell somehow being "right" then I'm not going to say a thing. :)

I think as far as monads go I quite like your distinction drawn elsewhere between abstractions for mathematical sake and abstractions for human cognitive sake. I think ultimately it has a lot to do with someone's goals as a programmer and person as to which abstractions they should hold and realize.

I think, personally, monads have a very high power-to-weight ratio here, even outside of Haskell, but I've never been one to argue that one cannot live without them. Or even shouldn't.

I will add that in general, there is an inherent mismatch between "equational" math and computation, and there have been various approaches to bring the two closer, e.g. Hoare Logic, various process calculi and temporal logic (my personal favorite). Haskell's approach is basically using lambda calculus[1] to unify some computations such as 2+2=4 with 2+2 -computes-> 4 [2](Prolog goes one step further and unifies 2+2=4 with both 2+2 -computes-> 4 and 4 -computes-> 2+2, as equality is a kind of relation and Prolog is relational) and other computations as monads wrapping various "unsafe" operations[3]. It works, and it's quite elegant if LC as the one true expression of computation is your cup of tea. I think it's clear why this approach can be a fine design choice, but it's also clear why it can be not so fine. Personally I think that design doesn't sit too well with most programmers as a cognitive abstraction.

So the answer to "why are monads useful?" is that if for whatever reason (e.g. it makes programming easier for you) you decide to model computation as Haskell does, as LC and "equational reasoning", monads are invaluable as an elegant description of various computations that wouldn't otherwise be convenient to express. Once you use other approaches, monads (as a cognitive abstraction) quickly lose their usefulness and appeal.

[1]: Which is one common definition of computation and one of the "original three" definitions, but not a very good or powerful description of computation as it lacks good measures of complexity (it also lacks introspection which makes it less powerful than the Universal Turing Machine, but that is no a very common capability in programming languages other than Lisp).

[2]: In fact, "computational math" doesn't care in the least that 2+2 equals 4. All it cares about is the computation 2+2 -computes-> 4 (or 4 -computes-> 2+2, the latter is completely separate from and unrelated to the former). It is those various approaches of verifying computation that care about the relationship between 2+2=4 and 2+2 -computes-> 4, but those can (and should?) be completely separate from the programming language.

[3]: That split between some computations and others is exactly why a `sleep` operation in Haskell can be implemented to either be "effectful" (using monads) or "pure" (using only "pure" computation). That artificial distinction doesn't really have a rigorous definition in computation theory; rather, the definition of "pure" is a language-level design choice. For example, in Koka a divergent computation (i.e. infinite sleep) is an effect, but not a finite sleep; in Haskell, neither is an effect. In both Haskell and Koka, a stack overflow is not an effect and neither is memory allocation. OTOH, in Haskell spawning a new thread is considered an effect (which seems really weird to me). I hope that at this point it is clear that the distinction between "pure" and effect is arbitrary and not a fundamental concept of computation (aside, perhaps, from IO).

I appreciate your explanation but again it fails to actually show it being useful outside of the context of working around Haskell's strict type system. What the OP and myself would like to see is concrete examples why this is a better approach than the way something would be done without explicitly caring about monads (I say explicitly because I know there is a tendency to say that some given structure is a monad and we didn't realise it; but that still doesn't really prove that it is all that useful in the general sense)

One problem I suppose is that not everyone speaks the abstract language of monads and so it does not serve much usefulness, yet, and maybe there is a little chicken and the egg with that.

You know how ES6 has this big hoorah coming because they're finally getting something to make callback hell a little less hellish, async/await?

Well. It's a monad. It's about 30s of work to implement it if you know what you're doing.

Of course, you could implement it as a monad in Javascript as well. It'd be a neat trick, but you'd never love it. The reason is simple: you can't benefit from monads unless you use them so pervasively that everything will integrate together.

And if you do, that integrate together bit works fantastically. People often complain about monads not composing---but I think they're often just misinterpreting a rather technical result. Actually, imo, monads compose incredibly well and it's astounding when you get used to it.

When you use them pervasively, monads mean that you get to "pick your own semantics" on the fly, whenever you want. You can mix and match semantics as is interesting and work with your custom mixes as if they were built into the language.

So people, e.g., talk about how it was easy to write STM because of monads.

It wasn't "because of monads". Of course STM is a monad and anything which looks even halfway like it in any language will also be a monad no matter how hard you try to avoid it. They're "just a pattern".

But when you've got a language which allows you to cut into the "STM monad" exactly and whenever you want, when you've got a hold of the root of the semantics of your language, when you've got programmable semicolons, then there's something really special.

And without that you've got a couple of years of bickering between standards committees to fix what end up being trivial looking problem.s

> I say explicitly because I know there is a tendency to say that some given structure is a monad and we didn't realise it; but that still doesn't really prove that it is all that useful in the general sense

Exactly. That is what I call mathematical- vs. cognitive-abstraction. They are not the same thing, and in the case of programming languages, only the latter matters. A mathematical abstraction is simply the naming of a pattern; a cognitive abstraction is a human solving a problem by explicitly thinking of said pattern. The former is either true or false; the latter is either useful or not (and the utility is completely orthogonal to any mathematical considerations).

> I appreciate your explanation but again it fails to actually show it being useful outside of the context of working around Haskell's strict type system

I think it isn't useful as a cognitive abstraction (and the problem isn't Haskell's type system, but its purity from effects). I believe that monads as a cognitive abstraction are foreign and harmful to imperative languages, which have an equally-powerful cognitive abstraction -- the thread -- that fits much better with the rest of the language's abstractions and I think is much more useful for most people.

I'll take a shot at it.

Quoting tikhoni:

> For Maybe, being a monad gives us a standard way of working with values while automatically dealing with Nothing. It abstracts over repetitive null checking and lets us easily build up Maybe values based on other Maybe values.

To expand on that a bit, in a way that might work for you (or might not): I've got a function f a that exists. It works. It does exactly what I want. Unfortunately, the data I've got is a Maybe a, not just an a. So I can't use my nifty function f, because it takes an a. You can feel my annoyance/frustration - so near, and yet so far.

But, in fact, I can use my function f, because Maybe is a monad. That lets me apply f to my Maybe a, without having to change either f or a whatsoever. I don't even have to write the magic code to do this - Maybe already has it.

Isn't Maybe being an instance of Applicative enough in this case?
Maybe...

But seriously, I guess I think it's not a Monad. The Monad signature (in this case) is

  (>>=) :: Maybe a -> (a -> Maybe b) -> Maybe b
and I think what we really want (and what I said we wanted to do) is

  Maybe a -> (a -> b) -> Maybe b
I don't know whether that's Applicative or Functor, but I don't think it's Monad.
That operation is called fmap, and it's characteristic of a functor. In Haskell, it's called liftM for monads, but they're really the same thing (and I think recent proposals are going to iron out this redundancy). fmap is one of the properties of a functor. All monads are functors, so all monads provide an operation with that signature.
I'll try and keep it simple. It helps make code more readable.

When I see a bunch of Maybe monads, for example, I can instantly know that the programmer's intent was maybe to not just have a type safe null and avoid null pointer exceptions, but also to possibly make use of Maybe's toxicity.

What I mean by toxicity is you know how in some languages you have "not a number" aka NaN? A NaN times / divided by / added to / subtracted from anything is a NaN, that means that if you have a complex formula, if you throw a NaN in there anyway you get back a NaN. Maybe works the exact same way but with all functions not just arithmetic operators. That's what I mean by its toxicity.

Without Maybe you'd have to explicitly apply that logic everywhere you'd otherwise apply a monadic "join." The reader would then have to study your code much more carefully to understand that intent, and probably look at it with some skepticism that the toxicity works perfectly every time or is in every place you might expect. And of course not everyone values brevity and expressiveness, but when done right it makes code more readable not less, and Monads done right facilitate that.

Maybe's toxicity is just an example of pretty much any "magic" you can build into a Monad. You could for example have a Future Monad that will handle asynchronous control flow for you.

As a side matter, it's not just working around Haskell's strict type system that make Monads a good fit for IO, but also Haskell's non-strict execution.. monads help things happen in the right order which is crucial for IO. That's kind of neat when you think about - the bridge between the mathematical rigor of pure functional code and real world devices that make computing useful. Functional purity makes code easier to reason about and less error-prone.

I'd say STM (software transactional memory) is a compelling example. Your computations that mess with shared variables can only mess with shared variables, the type system enforces this, and the only way you can actually run your shared-memory-changing transaction is through a function `atomically :: STM a -> IO a` (this gives a nice effect: `atomically $ do ...`)
IO, Maybe, ST, STM all being monads gives you the same sort of benefits that you get from Int, Float, Double, Integer, Complex all being numbers.
Ha, you're quick on the draw! I think the exploration of `after` and `doSomething` highlight good insights that I missed.

The idea that the `IO` type forms a monad (as opposed to `IO` being a monad just because the wizards said so) is also really important and missing from my explanation. Great work.

Why doesn't the English language have a Super Thank You in it, some way of saying "Holy Shit Was That Uniquely And Exquisitely Useful" ?!

My man tikhonj... thanks.