- Scott Wlaschin's Railway Oriented Programming (ROP): https://fsharpforfunandprofit.com/rop/ and https://fsharpforfunandprofit.com/posts/recipe-part2/

- Scott Wlaschin's The "Map and Bind and Apply, Oh my!" series at https://fsharpforfunandprofit.com/series/map-and-bind-and-ap... and its concept of an "elevated world".

- And for Elixir Programmers, Zohaib Rauf's "Railway Oriented Programming in Elixir" at https://zohaib.me/railway-programming-pattern-in-elixir/ based on Scott Wlaschin stuff.

- https://wiki.haskell.org/All_About_Monads. The introductory part (7 first pages of the pdf).

- Brian Beckman: Don't fear the Monad https://www.youtube.com/watch?v=ZhuHCtR3xq8

- Philip Wadler's "Monads for Functional Programming" at http://homepages.inf.ed.ac.uk/wadler/papers/marktoberdorf/ba...

- HN comment: https://news.ycombinator.com/item?id=16446139

ROP is a good way to introduce the Either monad without ever talking about monads, functors and co. Moreover it is highly useful.

(Personally I find the railway abstraction makes a lot of sense, but maybe that's just me.)

Let me set the record straight: The whole mysticism around the monad is the result of a misunderstanding. The monad is a very simple concept. It’s the diversity of applications of the monad that causes the confusion. As part of research for this post I looked up duct tape (a.k.a., duck tape) and its applications. Here’s a little sample of things that you can do with it:

- sealing ducts

- fixing CO2 scrubbers on board Apollo 13

- wart treatment

- fixing Apple’s iPhone 4 dropped call issue

- making a prom dress

- building a suspension bridge

Now imagine that you didn’t know what duct tape was and you were trying to figure it out based on this list. Good luck! So I’d like to add one more item to the collection of “the monad is like…” clichés: The monad is like duct tape. Its applications are widely diverse, but its principle is very simple: it glues things together. More precisely, it composes things.

===

Also would like to recommand: Functors and Monads For People Who Have Read Too Many "Tutorials" http://www.jerf.org/iri/post/2958

I got a feeling that Monad has no good metaphor to real life objects. Trying to make up an analogy just makes things worse.

They are a fantastic and incredibly powerful abstraction though. Their mathematical properties, termed the monad laws, make sure of that.

Most of the people complaining about monads are not in your situation. They are struggling to grasp even the most basic examples because they understand concepts only by analogy to other things they already know and monads don’t map to anything they know.

You mean like most programmers?

Part of the hardness comes from people wanting to use monads as an entry point to category theory.

The way I have found success is to have them write a program first where a monad would be useful, and then ask them to implement it without monads. They will identify the need for a monad themselves, and then you can show them the tool they want. Then they use it to solve the problem, and at that point you start giving them a more rigorous theory.

http://blog.sigfpe.com/2006/08/you-could-have-invented-monad...

Monads are hard to explain, they are _not_ hard to use, and use with great effect. Option#{map, flatMap} changed how I approached programming, and did not require a full understanding of why they existed or what they were a part of.

I don’t know any Haskell at all, and it seems to be the only language out there where people think monads are a thing you need to learn, so I definitely have no motivation to learn them outside of curiosity.

But I am curious about them, enough to read every tutorial I come across about them. But I still have no idea what they are.

You sound like a kind and curious person. If you ever want a one-on-one walkthrough please reach out to me (email in bio). Everyone deserves the chance to understand something if they want to.

Explaining things to people that already knows them is always easy.

I'd say that your list is full of things that have practical examples when you consider how real, physical computers work. Boolean logic can be shown by example by showing physical transistor circuits, and I'd argue that people can understand them quite easily. Pointers/variables can be shown by explaining how memory addressing works on real machines, and what CPU instructions correspond with the actual code you're writing.

Monads however, seem to be this purely abstract mathematical concept that comes from CS-as-a-math-discipline, which although useful, requires you understand ton more theoretical background before you can approach it. At least that's my experience -- I still don't understand them and every tutorial I've read either makes them look like magic (ie. doesn't actually explain them) or tries to take a mathematically rigorous approach and I get lost in the weeds.

I know intuitiveness is very subjective, but I'd wager that by numbers, many more people are able to understand pointers/mutable variables/macros/boolean logic/etc more easily than they can Monads.

Young man, in mathematics you don't understand things. You just get used to them.

—- John von Neumann

You could basically rephrase it as “you don’t need a lot of math, you just need a lot of math”.

Honestly I think a lot of effort is wasted starting with this stuff rather than more motivated examples... But the interest is the harder thing to come by, so at least they're learning something.

http://www.jerf.org/iri/post/2958

https://philipnilsson.github.io/Badness10k/escaping-hell-wit...

The first one I think cuts through all of the crap surrounding them. They're just a design pattern. And the second one helps motivate them, showing the many contexts where the design pattern many apply.

Well, if they were, all those people wouldn't be trying to understand them.

They are applied on Haskell (and Javascript, Rust, C#... but they are way more visible on Haskell). Write some Haskell and you will get them.

But, well, I don't have any reason for you to do that. You just asked how, and I gave you a way. If you don't have any interest in learning it, there's absolutely no problem, but posting "I never really tried, but that thing is bad because I haven't learned it already" FUD is pretty bad.

But so far, no monad tutorial has made a good case for why they’re actually useful, other than for learning Haskell. And most advice for how to learn them sounds exactly like what you said: “oh if you still don’t get it, you just need to write more Haskell.” You can see why that’s a bit of a disappointment, right?

Basically, you have people making this argument:

    - monads are important outside of Haskell
    - to understand why, you have to learn them
    - to learn them, you should write Haskell

If you want to use a language to program, knowing them will help you with approximately nothing, because most languages either have awful ergonomics for that kind of abstraction or bring them on much more limited ways that you are better learning how they specifically apply to the language than the general concept. If you learn the general concept, the specific cases will be easier, but unless you plan on learning several languages, it's not worth it.

Anyway, one large exception is if you plan on doing heavily abstract C# libraries. The C# support for monads is almost as general as Haskell.

I disagree -- these all seem pretty easy to explain to me.

Monad tutorials often avoid this problem by talking about monadic values, so perhaps in this example it could be called a functoric value?!

Yes, that is bogus. If you're tolerant of math jargon, I found this article demystified the topic more than anything else did:

https://en.wikibooks.org/wiki/Haskell/Category_theory

And this explains where the monad laws come from:

https://www.haskellforall.com/2012/08/the-category-design-pa...

The op is talking about something different, though. What is a term for List Int, a concrete type instead of a type constructor. We cannot call List Int a Monad, we cannot even call (List Int, pure_Int, bind_Int) a Monad. The definition really requires polymorphism. But, obviously, List Int is closely related to the List-Monad.

I've also heard "monadic value" used. I tend to just use "monad" even though I know it's an even greater abuse of language.

I get how Maybe types can be useful for parsing something or doing IO but once you get your Maybe value, wouldn't be the first action to unwrap it? If it is Nothing you can either provide a default value or just stop right there. I don't see the sense of dragging the Nothing further done with you and having to deal with potential nothingness further done the stack even if you have the tools to make it look nice.

Maybe there are other functors where this makes more practical sense?

I agree that this technique works well — as long as you need to use only one Maybe at a time. As soon as you want to combine multiple computations, each of which returns a Maybe, this quickly becomes horrible and cumbersome:

    case parseNumber str0 of
      Nothing -> Left "parse error"
      Just str1 ->
        case parseChar '=' str1 of
          Nothing -> Left "parse error"
          Just str2 ->
            case parseNumber str2 of
              Nothing -> Left "parse error"
              Just str3 -> …

And of course, then you get some data structures which can’t be unwrapped at all. The standard example is Promises as seen in e.g. JavaScript, but there are others.

More precisely, the monad interface allows abstracting this particular repeated structure in much the same way an iterator allows abstracting the repeated process of "get the next thing", but monad does not exist for Maybe just as iterators do not exist for arrays.

It's easy to make the mistake of focusing on the Maybe portion of that repeated code snippet, but the value of the monad interface in that code snippet is everything other than the Maybe portion, that pattern of descending into a computation over and over again with a closure, that fits into a number of data structures and an even larger number of data structures that can be coerced into fitting into that pattern without much effort.

Moreover, Maybe is a bad way of understanding the monad interface and pattern because it does not fully use the monad interface. It doesn't take advantage of the way the Maybe value is passed through to thread any additional values through. While a valid example of a monad interface usage, it is also a degenerate one that makes for a difficult lens to understand the interface.

Another benefit to using Maybes is that you avoid the rigamarole of null checks at every call site where you want to use the value. If you have a function that returns null or a value, whenever you call that function you'll always have to add an if guard to validate it's not null. If it is, that function itself may return null, and callers to it will again have to implement the same check.

I wrote a JS implementation of the Option object and the readme has lots of specific examples about these benefits: https://github.com/sbernheim4/excoptional and a few blog posts on these ideas[0][1]

[0] https://sambernheim.com/#/blog/the-engineers-schrodingers-ca... [1] https://sambernheim.com/#/blog/building-a-monad

    return myList?.deleteFirst?.takeFirst?.name

    if (myList != null){
        firstDeleted = myList.deleteFirst;
        if (firstDeleted != null){
            newFirst = myList.takeFirst;
            if( newFirst != null){
                return  name;
            }  
        } 
    }

Same reason why you iterate over an array instead of assigning each value to a temporary variable: you have less code.

For example for the list monad the transformation could be applied in parallel or not at all if the list is only partially consumed; for the IO monad you could perform every operation asynchronously in an event loop instead of blocking.

Here's a pretty odd one:

    data Needs r a = Needs { execute :: (a -> r) -> r }

    (execute x) length :: Int

    ex1 :: Needs (\view -> view "Hello, world")           -- contains one string
    ex2 :: Needs (\view -> view "Hello, " + view "world") -- contains two strings
    ex3 :: Needs (\view -> 0)                             -- contains no strings

    ex4 :: Needs (\view -> let n = view "Hello" in view (repeat n "world, ") \* n + 10)

    (execute ex4) (\a -> { push v a; () }) -- not really Haskell syntax

But Needs is still a monad.

    pure :: a -> Needs r a
    pure a = Needs (\view -> view a)

    (>>=) :: Needs r a -> (a -> Needs r b) -> Needs r b
    m >>= f = Needs (\view -> execute m (\a -> execute (f a) view))

Finally, I'll give up the ghost and note that Needs is known as the Continuation-Passing Monad. It helps you write code in CPS style without having to actually, deliberately pass continuations all around. This can be really useful for a few dumb-programmer tricks, but it's also a pretty useful framework for building embedded domain-specific languages.

I find this approach can make understanding the intuition behind monads more approachable.

[0] https://sambernheim.com/#/blog/building-a-monad

    (a -> b) -> T a -> T b
  T (a -> b) -> T a -> T b
  (a -> T b) -> T a -> T b

That said, I prefer a presentation using `times` and `join` rather than `ap` and `bind`:

   map   :: (a -> b) -> (f a -> f b)
   times :: (f a, f b) -> f (a, b)
   join  :: f (f a) -> f a

`times` lets you take a pair of sources and turn it into a source of pairs. This lets you build pipelines that flow together from multiple sources into a single result; with just `map` you can only assemble a unary pipeline with one source.

`join` lets you take a source of sources and meld the layered structures together into a single source. This lets you create pipelines whose contents dynamically extend the pipeline. Without `join`, you can only construct pipelines whose stages are fixed up-front.

(As a bonus, the presentation with `times` leads to the idea that an "applicative functor" is a monoidal functor. If you don't know what a monoid is, maybe it's a toss-up in terminology, but the term "monoid" is vastly more common than the term "applicative".)

Promise<_> is a functor because we can write a function `map : (Promise<A>, A -> B) -> Promise<B>` that takes a pipeline stage to run, and promises to run it on the A when it arrives. That stage produces a B, but we don't have it yet, either (since it needs that A). Just having a `Promise<B>` doesn't tell us how many pipeline stages it's built -- the fact that we called `map` isn't evidenced in the type itself -- but we know that there's a fixed number, and that all of the stages have already been given to it. After all, we call `map` in the current timestep; the stages are known long before they get to run.

Promise<_> is a monoidal functor because we can write a function `times : (Promise<A>, Promise<B>) -> Promise<(A, B)>` (but see note^). We don't necessarily have either the A or the B, and the interactions that produce them may not complete at the same time; but eventually we will have both of them, so we can wait until then and pair them together. Because we can wait for two pending interactions at once, we can wait for any number of them at once; think about how to write a function `[Promise<A>] -> Promise<[A]>` using `times`.

With `times`, we're just combining all of the stages in the existing promises together. Unlike `map`, which applies stages sequentially, `times` applies them concurrently -- like a branching series of pipes with Y-bends where `times` is used. With this, a `Promise<A>` still has a fixed number of stages known up-front -- all stages are still provided through `map,` which we still call before we ever have an actual `A` to run them with -- but we also now have a fixed number of forks/branches, where before it was a straight-line series of stages.

Promise<_> is a monad because we can write a function `join : Promise<Promise<A>> -> Promise<A>`. The nested promise means that we need to wait for one pending interaction, and then wait for another pending interaction, before we can get an A. This often happens because we interact sequentially with a backend service; we need to log in before we query the system, so if we try to put together pipeline stages like `(Username, Password) -> Promise<Session>` and `Session -> Promise<SensitiveInfo>`, we'll end up with a `Promise<Promise<SensitiveInfo>>`. This type only lets us make up to two interactions; if we can't get `SensitiveInfo` within two interactions, we simply can't write the program at all.

The `join` function lets us collapse two interactions within a single Promise; and just like `times`, if we can collapse two interactions, we can collapse any number of them. With `join`, a type like `Promise<SensitiveInfo>` may take an arbitrary number of interactions before we finally get the result. However, this is a lot more powerful than you might expect: those inner promises don't exist yet! They will exist, as each interaction completes; but now a single `Promise<A>` includes both the stages we've set now, and any stages that future interactions may generate. We can no longer tell how many stages it will take until the final `A` is computed, nor how many interactions it will take. We'll only know when it actually happens.

(^) Well, we also need a function `pure : A -> Promise<A>`, which lets us pretend we don't have an `A` yet. Or, put differently, it puts an A behind a "no-op" interaction that completes immediately and makes the A available to any downstream stages as soon as it's needed. Something like an echo server.

Possibly once you're immersed in the world and up to speed with Haskell type signatures are all you need. Unfortunately for people coming to the language we need more context to fit the pieces together, and the vast majority of documentation I found lacked that context.

To take another example, an iterator could be something like:

    next    :: T a -> a
    hasNext :: T a -> bool

As for the culture part, I think Rust is the sweet spot. Types are here, but you also have explanations. I personally find explanations without types worse than types without explanations, but ideally you should have both. Of course that's assuming meaningful types.

I think this is a fairly standard functional representation of iterable entities:

  next :: T a -> Maybe a
  rest :: T a -> Maybe (T a)

  step :: T a -> Maybe (a, T a)

The interesting thing is that the functor `Maybe (a, -)` can have many co-algebras, depending on which type and step function you pick. The type [a] -- that is, picking T to be [] -- gives a particularly special co-algebra, called the "terminal" coalgebra for `Maybe (a, -)`. In short, while any other choice of coalgebra may support additional operations, you can convert any one of them to a list without disturbing a process that only makes use of the nominated `step` function. They're observationally equivalent to lists.

That's true, but I was talking about iterators in general. My point was that functors, applicative and monads may seem weird and special, but we use the same thing in everyday life. Same thing about the signatures: iterator is that, just a signature, but nobody complains about it.

Your explanation about functional iterators is great, I think it shows really well how iterators are about treating any data structure as if it was the simplest data structure. In a imperative language, it's the array, in a functional language, it's the list.

Yes, and the parent I was replying to had only provided the type signatures.

Even more than that though, describing what something is is usually not sufficient. You also have to describe why it's useful, or rather what it's for.

This is analogous with comments in code - in well written code you don't need many comments telling you what the code is doing, it's often fairly clear. They're also for convenience not necessity because you can, with enough time, figure out what the code is doing because you have all the information. What are irreplaceable are "why" comments telling you why something is done. E.g. "this prints two newline characters at the end of the document" is inferrable, but "print two newline characters because X program is unreliable unless we end with a blank line" is invaluable.

The thing is, these things are very abstract. The signature is what they are and why they are useful. A way to think about them is that they were discovered. People wrote a lot of code, saw the same pattern come up over and over again, and named it so they can abstract over it and reason about it.

> This is analogous with comments in code - in well written code you don't need many comments telling you what the code is doing, it's often fairly clear. They're also for convenience not necessity because you can, with enough time, figure out what the code is doing because you have all the information. What are irreplaceable are "why" comments telling you why something is done. E.g. "this prints two newline characters at the end of the document" is inferrable, but "print two newline characters because X program is unreliable unless we end with a blank line" is invaluable.

I agree with you here. The problem is that it's very hard to describe why they are useful if you didn't encounter the problem they solve yourself.

But I think GP comment summarized it pretty well, it shows how they all relate to each other in a concise way. If its not enough of an explanation, ok, look for another.

Of course, there are lots of libraries that only have type signatures as documentation. I think most people think that's less than optimal, but the haskell community is much smaller than many others. So what you'll want to do is look for a blog post or something that explains it in more detail.

For example,

  (a -> b,T a) -> T b
  (a -> b,T<a>) -> T<b>

Note: I do understand why auto-currying is extremely useful for doing functional programming, and I have no objections to teaching it. I just don't think it's important for teaching functors, applicatives and monads.

In a separate nit-pick, these three function definitions don't mean much on their own, and there are many pure functions that satisfy them without forming a functor/applicative/monad.

Probably because Haskell (and related languages) use it.

> For anyone coming from a more traditional CS background, it's just extra cognitive overhead in understanding the meat of the argument

I find that curried function signatures are slightly lower cognitive overhead after a bit of exposure, and while I can see that its the kind of thing that is likely to vary from person to person a bit, I can't see that it's likely to be a big thing either way.

I think it's primarily because this is the form you use them in in a language like Haskell. Most explanations are exported from Haskell in one way or another, so they're deeply colored by Haskellisms.

It's the same reason why `ap` and `bind` are generally preferred over `times` and `join` -- in Haskell you can even use `f <$> a <*> b`, which is a deeply clever use of infix `fmap` and `ap` that drives a syntactic analogy with regular function application, `f a b`. Try doing that in another language and you'll barf. And do-notation desugars most cleanly into >>=, so that's where Haskell-oriented explanations start from.

If you implement these in Java (which you can, you just can't abstract over T), it looks like this:

    class List<A> {
      <B> List<A> map(Function<A, B> f) { ... }
      
      <B, C> List<C> map2(List<B> other, BiFunction<A, B, C> f) { ... }

      <B> List<B> flatMap(Function<A, List<B>> f) { ... }

      // bonus; note that these make more sense static
      static <A, B> List<Pair<A, B>> times(List<A> as, List<B> bs) { ... }

      // in fact, this one has to be static
      static <A> List<A> join(List<List<A>> aas) { ... }
    }

   trait Functor[F[_]] {
     def map[A, B](f: A => B)(x: F[A]): F[B]
   }

   trait Applicative[F[_]] extends Functor[F] {
     def pure[A](a: A): F[A]
     def apply[A, B](f: F[A => B])(x: F[A]): F[B]
     def product[A, B](xa: F[A], xb: F[B]): F[(A, B)]
   }

   trait Monad[F[_]] extends Applicative[F] {
     def bind[A, B](f: A => F[B])(x: F[A]): F[B]
     def join[A](x: F[F[A]]): F[A]
   }

Well, it indeed worked perfectly well. But the amount of noise added to satisfy the type checker was larger than reimplementing the log procedure at each place.

    (a -> b) -> (f a -> f b)

    =< :: (a -> m b) -> (m a -> m b)

Compared to other versions I have seen this Haskell-style is the most elegant and understandable.

What about this? Is this even possible?

(T a -> b) -> T a -> T b

I guess to complete the set you can also add this one which is not very interesting, it’s just “apply the function”

(T a -> T b) -> T a -> T b

> (T a -> b) -> T a -> T b

Yep; that's the signature for the "extend" method of a comonad.

https://hackage.haskell.org/package/comonad-5.0.8/docs/Contr...

Though I prefer the presentation with `duplicate :: w a -> w (w a)`, which is just the signature of Monad's `join` but flipped around.

That’s just a normal function application followed by return.

The actual definition of what? `return` is not part of the Comonad class, while `unwrap` (alternatively called `extract`) is not part of the Monad class. No matter what your preferred way of designing abstractions, it seems a strange complaint to say "feature X is useless with abstraction Y" if you don't have a Y in the first place.

> It’s not like any of this is relevant anyway. The usual abstractions liked by Haskellers are utterly pointless as this whole discussion nicely demonstrate.

Well, okay then.

You may not find value in this way of thinking, but other software developers -- many of whom, yes, ship actual software with actual business value -- find these ideas extremely valuable in organizing their thoughts about the problem domain.

I'm open to sharing that mindset with anyone of a curious mind, but most of us aren't going to shove it down your throat. You're welcome to do what works for you.

You can certainly do what you describe if you have a monad, but since you can't even try to unwrap a value if you don't know what the monad is concretely, you have no way to write a function `T a -> b` that's generic over T. Functions `T a -> b` only really make sense for comonads, which give you the `extract` and `extend` functions as primitive.

An alternate might be to have examples in multiple languages, colour-coded.

Where typeclass links to this other massive tutorial and then my eyes started to glaze over.

(Side note, this is probably the 50th or so Monad tutorial I've read and I still don't understand them. For context, I've been programming professionally for nearly 20 years now. My only conclusion at this point is either Monads aren't actually that important, or I actually do understand some useful subset of what Monads are and I just don't know it, and I just never had a need to use them "fully".)

Edited to add: I think my problem with Monad tutorials is that they all tend to explain the concepts using Haskell, and I can't read Haskell. It syntax is utterly impenetrable to me and I have no idea what any of the silly symbols mean. And any tutorial that tries to teach Monads using other languages ultimately feels like it's teaching me something I don't need to know... (how to do side effects in a pure functional language, when I don't code in pure functional languages and likely never will.)

As far as functors this could be rephrased more comprehensibly to those unfamiliar with Haskell as "What is a Functor, really? Functor is an interface here are some definitions:"

(typescript)

  interface ArrayFunctor<A, B> {
      map: (f: ((a: A) -> B), arr: Array<A>) -> Array<B>
  }

  interface PromiseFunctor<A, B> {
      map: (f: ((a: A) -> B), arr: Promise<A>) -> Promise<B>
  }

  interface ObservableFunctor<A, B> {
      map: (f: ((a: A) -> B), arr: Observable<A>) -> Observable<B>
  }

  interface Functor<F, A, B> {
      map: (f: ((a: A) -> B), arr: F<A>) -> F<B>
  }

> I actually do understand some useful subset of what Monads are and I just don't know it

This is certainly the case if you have ever used flatMap on an array or chained Promises (or other async structures like futures in other langs).

After half a year of trying to understand monads, I saw a video by MPJ/Funfunfunction [0] and it finally made sense.

[0] https://www.youtube.com/watch?v=9QveBbn7t_c

Almost anything that inspires this many people to be interested in learning, teaching, and talking about math cant be a bad thing.

  > Just 20 >>= half >>= half >>= half
  Nothing

I'd expect some kind of expection in that case.

This sounds like a dangerous example of monads (since you don't know if you had a Nothing value up in the chain or a valid value that can't be representated)!

Sorry, what exactly do you mean by ‘expectation’ here?

> This sounds like a dangerous example of monads (since you don't know if you had a Nothing value up in the chain or a valid value that can't be representated)!

Well, this particular case is of course oversimplified. There’s plenty of ways of dealing with different kinds of errors. The easiest is to define a sum type of errors:

    data DivisionError = ValueMissing | NonIntegralValue

    half x
        | even x    = Right (x `div` 2)
        | otherwise = Left NonIntegralValue