Carmack talks practical use of Haskell, Lisp in game prog at Quakecon keynote

Some of us who argue in favor of dynamic typing have a much more nuanced and informed view...

I, for one, am of the mind that the there are precisely zero production-caliber statically typed environments that possess a sufficiently powerful type system for the kinds of problems I tackle on a regular basis. Haskell doesn't count, since you need to turn on about a dozen GHC language extensions in order to incorporate the last 20 years of research. There's also quite a bit of design warts that newer academic languages are starting to iron out. In particular, I don't think monad transformer stacks are a reasonable solution to computational effects.

That's not to say you can't write any program in an environment where the type system is constraining you. You can. You simply implement a "tagged interpreter", which is something that's so simple to do, people do it all the time without realizing. Either you have a run-time map or you pattern match on an sum type data constructor, then loop over some sequence of those things with a state value threaded through. Poof! You've got a little interpreter.

I find that this happens a lot. And, I also find that a lot of problems are easier to reason about if you create a lazy sequence of operations and then thread a state through a reduction over that sequence. Now, in Haskell, I've got a type correct interpreter for an untyped (unlikely turing-complete) language! Sadly, I can't re-use any of the reflective facilities of the host language because my host language tries to eliminate those reflective facilities at compile time :-(

I'm in favor of optional, pluggable, and modular type systems. I think that a modern dynamic language should come with an out-of-the-box default type system that supports full inference. If, for some reason, I build a mini interpreter-like thing. I should be able to reuse components to construct a new type system that lets me prove static properties about that little dynamic system. This level of proof should enable optimizations of both my general purpose host language and my special purpose embedded "language".

Additionally, I require that type checking support external type annotations, such that I can separate my types from my source code. In this way, type checking becomes like super cheap & highly effective unit tests: The `test` subcommand on your build tool becomes an alias for both `type-check` and `test-units`. You just stick a "types/" directory right next to your "tests/" directory in your project root. Just as a stale unit test won't prevent my program from executing, neither will an over-constrained type signature.

Have you ever had a bug due to something being null/nil when you didn't expect it? How about a string or list being unexpectedly empty? Perhaps you've discovered an XSS or SQL-injection vulnerability? What about an exception that you didn't anticipate, or really know what to do with?

In a more robust type system, these could all be type errors caught at compile time rather than run time. A concrete example of the null/nil thing; in Scala, null is not idiomatic (although you can technically still use it due to Java interop, which is understandable but kind of sucks). To indicate that a computation may fail, you use the Option type. This means that the caller of a flaky method HAS to deal with it, enforced by the compiler.

My "come to Jesus" moment with the Option type was when writing Java and using a Spring API that was supposed to return a HashMap of data. I had a simple for-loop iterating over the result of that method call, and everything seemed fine. Upon running it, however, I got a null-pointer exception; if there was no sensible mapped data to return, the method returned null rather than an empty map (which is hideously stupid, but that's another conversation). This information was unavailable from the type signature, and it wasn't mentioned in the documentation. The only way I had of knowing this would happen was either inspecting the source code, or running it; for a supposedly "statically-typed" language, that is pretty poor compile-time safety.

This particular example of a stronger null type would be doable in the "weaker" languages, but it isn't done for several reasons - culture and convenience are the two most prominent in my opinion. In this sense, "convenience" means having an interface that does not introduce significant boilerplate; any monadic type essentially requires lean lambdas to be at all palatable. "Culture" refers to users of the language tolerating the overhead of a more invasive type system, which admittedly does introduce more mental overhead.

- The Maybe/Option type: you explicitly declare that a value may be missing, so you cannot call methods/functions on it willy-nilly; the compiler will force you to handle both cases. Say bye-bye to NoneType object has no attribute 'foo' errors.

- Different types for objects that have the same representation: in a language like Python, text, SQL queries, HTTP parameters, etc. are all represented as strings. Using a statically-typed language, you can give them each their own representation and prevent them from being mixed with one another. See [1] for a nice description of such a system. See also how to separate different units of measurements instead of using doubles for everything.

- Prevent unexpected data injections. With Python's json module, anything that is a valid JSON representation can be decoded. This is pretty convenient, but it means you must be very careful about what you decode. With Haskell's Aeson, you parse a JSON string into a predefined type, and if there are missing/extra fields, you get an error.

- When I was doing my machine learning class homeworks, I very often struggled with matrix multiplications errors. An important part of that was that the treatment of vectors vs Nx1 matrices was different. I feel that if I could encode the matrix size in the types, I'd have had an easier time and less errors.

These are simple examples, but whenever I code in Python, I inevitably make mistakes that I know would've been caught by the compiler if I had been coding in OCaml or Haskell.

[1] http://blog.moertel.com/posts/2006-10-18-a-type-based-soluti...

A bit of insight into the use of Racket here (scroll down): http://comments.gmane.org/gmane.comp.lang.racket.devel/6915