[clin_]

Here's a handful of things I really appreciate about Haskell, although some of this apply to other ML family languages as well:

1.) Algebraic Data Types let you enforce a huge amount of correctness in your programs simply by how you define your types.

For example, say you have a data type that represents a player's state in a game. If the player is alive, he has a "HP" and "inventory" record. If the player is dead, he should just be dead—HP and inventory may or may not be set to 0 and None, respectively. In C or OOP languages, you just get a player object/struct and you have to enforce this state logic manually, and you can easily write a function that puts the player in an inconsistent state (player.alive = False, but player.HP = 15). In Haskell, you can define the player such that code that produces this kind of inconsistent state will not compile.

Of course, you should still write tests in Haskell, but it's amazing to me that the type system can catch the vast majority of programming errors before your program even compiles. It takes longer to get something up and running, but you make up for this up front cost with long term gains in maintainability.

2.) Functional purity enforces two very useful properties:

a.) Data is never mutated by accident. This ensures you never have to look beyond the scope of the function you're working on (or closing over, in some cases) to know what's going on. You can then compose these functions to create more complex programs.

b.) When you do need mutability and side effects to happen, Haskell forces you to be very explicit about this and enforces a strict separation between "pure" and "effectual" code. This seems arbitrary until you realize that almost all of your errors not caught by the compiler happen because of IO. Whereas you have complete control over the pure part of your program, IO is nondeterministic and full of exceptions and edge cases. Separating this part of your code makes it far easier to reason about what's happening, adding to maintainability.

As a side note on the last point: people make the argument that this separation makes Haskell impossible to do printf debugging. Not true—the standard library provides ways of circumventing the type system specifically for this reason.

3.) "Monads" and other fancy sounding abstractions can be hard, but with practice they become as intuitive as inheritance in OOP. The real barrier to entry is the awful, awful documentation everywhere. I make no apologies for that: it's just really shitty and the community needs to do a hundred times better.

[SamReidHughes]

> In C or OOP languages, you just get a player object/struct and you have to enforce this state logic manually, and you can easily write a function that puts the player in an inconsistent state (player.alive = False, but player.HP = 15).

False. In OOP you can do this using the visitor pattern. The difference is that Haskell makes this real convenient, with its algebraic types feature baked into the language, while OOP makes it very cumbersome.

> As a side note on the last point: people make the argument that this separation makes Haskell impossible to do printf debugging. Not true—the standard library provides ways of circumventing the type system specifically for this reason.

It's still far more annoying when the language is lazily evaluated and has Haskell syntax.

[clin_]

> False. In OOP you can do this using the visitor pattern. The difference is that Haskell makes this real convenient, with its algebraic types feature baked into the language, while OOP makes it very cumbersome.

I don't think we're on the same page here. I meant to say that Haskell protects against inconsistent states: you literally cannot create a dead player with HP and Inventory records. I don't have a strong grasp on the visitor pattern, but I'm not sure how it can be used to accomplish compile time guarantees that the properties of an object will always be correct.

> It's still far more annoying when the language is lazily evaluated and has Haskell syntax.

Python:

  def f(x,y):
      print x, y
      return x + y

Haskell:

  f x y = traceShow (x, y) (x + y)

Laziness is a separate issue altogether. Debugging lazy behavior is difficult in any language, and you would have similar problems in Python if you were testing deeply nested generators.

That said, laziness should not be an issue if you are simply testing that functions are correct, since most functions are pure and you will almost never use lazy IO.

If you need to debug evaluation order or a memory leak, that's much harder, and admitted one of Haskell's biggest weaknesses. My answer here is that you should almost always prefer strict data structures and functions unless you really need laziness. That's not a cure all, but it is a good policy when programming in Haskell.

[SamReidHughes]

> don't think we're on the same page here. I meant to say that Haskell protects against inconsistent states: you literally cannot create a dead player with HP and Inventory records. I don't have a strong grasp on the visitor pattern, but I'm not sure how it can be used to accomplish compile time guarantees that the properties of an object will always be correct.

An example of the visitor pattern:

    class LivePlayer;
    class DeadPlayer;

    class PlayerVisitor {
    public:
      virtual void VisitLive(LivePlayer *p) = 0;
      virtual void VisitDead(DeadPlayer *p) = 0;
    };

    class Player {
    public:
      virtual void Visit(PlayerVisitor *v) = 0;
      virtual ~Player() { }
    };

    class LivePlayer : public Player {
    public:
      void Visit(PlayerVisitor *v) { v->VisitLive(this); }
      LivePlayer(int hp, std::vector<InventoryItem> inventory)
        : hp_(hp), inventory_(inventory) { }
      int hp() const { return hp_; }
      const std::vector<InventoryItem> &inventory() const { return inventory_; }
    private:
      int hp_;
      std::vector<InventoryItem> inventory_;
    };

    class DeadPlayer : public Player {
    public:
      void Visit(PlayerVisitor *v) { v->VisitDead(this); }
      DeadPlayer() { }
    };

This is equivalent to the Haskell code

    data Player = LivePlayer Int (Vector InventoryItem) | DeadPlayer

including the fact that in this example the objects are immutable.

Instead of using a case expression to pattern match over the player, you'd have to construct a PlayerVisitor and implement the visit methods on that type.

Things can be made a bit less cumbersome than that, for example in a language with lambdas, you can just have the Visit method take a lambda for each subclass, and each subclass's implementation calls one. That makes it arguably equally convenient to use the types (with some extra parentheses), but it's still much more annoying to define the types that way in the first place.

[platz]

I think even the visitor pattern isn't required here, all you need is subclassing. Although this clutters things up more by having abstract methods in the base class, technically that's all you need.

[SamReidHughes]

Sure.

[quchen]

To be fair, if the place you're trying to debug is not evaluated (be it due to laziness or because it's generally dead code) that should be a pretty strong indicator something is wrong.

[dllthomas]

And generally a good clue as to what is wrong (or at least where to put the next couple checks...).