Joe Armstrong: Why OO Sucks

OO vs. FP is just a matter of whether you focus the nouns or the verbs. The counterargument to the OP is that surely a function that manipulates "data" is less powerful and abstract than one that manipulates objects.

For example, take an "interface" or abstract data type like Array, consisting of a length() and a get(i) method. (This is really called List in Java and Seq in Scala.) There may even be an associated type, A, such that all items are of type A. This is very powerful because functions written against the Array interface don't depend on the implementation; we can store the data different ways, calculated it on demand, etc.

The "binding together" Joe is complaining about is binding the implementation of length() and get(i) to the implementation of the data structure, which is surely understandable. The alternative, seen in Lisps and other "verb-oriented" languages, is that there is a global function called "length" which takes an object... er, a value... and desperately tries to figure out how to measure its length properly, perhaps with a giant conditional.

The original OO (SmallTalk) was about message passing rather than abstract data types; just the idea that an object was responsible for responding to certain messages, and that these communication patterns completely characterized the object. This is how we think about modern cloud services, too; it's kind of inevitable. Who would complain that S3's "functions" and "data" are too coupled? Who would ask for a description of S3 in terms of what sequence of calls to car and cdr it makes internally? OO concepts allow a functional description of a system that starts at the top and can stop at any point.

The "everything is an object" philosophy gets a bad rap. It's a big pain in Java, especially, because of how the type system works. Ideally I'd be able to define a type of ints between 1 and 3, an obvious subclass of ints in general, whereas in Java I find myself declaring "class [or enum] IntBetweenOneAndThree" or some nonsense.

"If NASA's Remote Agent[1] had been implemented in Haskell, OCaml or ML, do you think debugging DS1 would have been as simple as connecting to a REPL and changing some definitions in a living core?"

Erlang in particular is known for rock-solid support for this, and Armstrong never advocated for pure functional languages.

"Sure, you can use a simple array to build a b-tree, but don't you want to explicitly state that variable x is a b-tree if that's the case?"

You can do this in Erlang with atoms.

OFFTOPIC - Man I can't believe how often I am coming across "sexist in IT" stuff now. I read your second link and I thought to myself it all makes sense. Demanding that compsci have more women is like demanding [insert women dominated field here] have more men, even though most men just aren't interested in the stuff. Why bother? Why not just let people work on what they want to work on?

Is this making me a sexist misogynistic jerk? I am not trolling, I am honestly curious. Explain to me why I am wrong. I will reflect and make up my mind and be intense and stuff.

I expect this to be a busy thread.

I think the point about state has some traction. For example, Chaper 15 of Effective Java (awesome book on Java, btw) is entitled "Minimize Mutability," so I think this idea is one that has caught on even in fairly traditional OO languages.

As for the other points... I do think sometimes that using OO to model real world objects may not always be wise, esp. if the result is a deep hierarchy, as in the OO 101 example of, say, Boston Terrier < Dog < Mammal < Animal < Thing... And then someone changes Dog and gives your Boston a tail... I dunno. I have only vague intuitions here, but perhaps the "objects as models of reality" might be perfectly suited for reality simulators of some sort that require stateful elements a la Sim City but not generally. Or, perhaps a better example: You could model a chess game as classes of Pieces on a Board, with methods like King.isInCheck(), or Queen.canMoveTo(Square) but this to me seems clumsier than simply having an 8x8 array of enums with the logic living in functions and not inside individual pieces.

Yes, a Lisp-like language could use a giant conditional for a generic "length" function. But that's not how the source code usually looks, and there are Lisp compilers which optimize it so it's not done as a conditional.

In CLOS, it will look as a separate DEFMETHOD for each type you want to define the function on, and these can be in separate modules. DEFMETHOD will just extend the generic function. And the function itself can be optimized to use the normal tricks -- things kind of like vtables -- to optimize performance.

The S3 example is a bit facile since (1) S3 objects are like mud, they're all the same and (2) they map cleanly to the OO paradigm since they encapsulate discrete chunks of external state. At the other end we can throw around examples like BLAS and LAPACK, whose functions are much closer to the "functional ideal" (ignoring mutability, here). For example, if I want to solve a linear equation, how do I express that as a single dispatch method? Does it go on the matrix, or on the vector? Or do I have to create a new object from the two called LinearSystem, just so I can invoke a single method?

We think about things in terms of coupled data, as pure functions, and as dirty generic functions (like the Lisp example). The only real lesson is that people get pissed off at languages that force them to shoehorn everything into one category (like Java).