I find it very difficult to believe that a large source of errors in ML programs are caused by programmers forgetting to annotate type information on top level functions.
Scala does that because ad hoc polymorphism doesnt resolve well with hindley milner type inference. F# also has this limitation and they get some of the way around this with a strict vertical projects file layout. The scope of its type inference is a middle ground between Scala and Ocaml which I think is a nice optimum, you leave most things untyped except the overloads which can't be determined so that the eager generalization you encounter in ocaml is less of a problem there too. The price in project layout is quite steep though but like in all things you train your self around it.
Plus with an IDE in Scala (or F#) you can just hover for the types.
Good post, one small issue though, "type inference, which you can do in any language", is not really true. Once you get dynamic enough type inference becomes nearly impossible (Cannon, 2005: http://www.google.com/url?sa=t&source=web&cd=4&v...)
Yes, but for many languages which are that dynamic, we now have JIT VMs which gather tracing data and can empirically tell us what the types actually are. This is especially true for server apps, which can provide us with large datasets. Also, as noted in a recent article posted to HN, after a certain span of time, most apps in dynamic langs behave as if statically typed. Why don't we use this information? It happens to be move available when it is most valuable. (When a codebase has "matured" and things are moving more towards a "maintenance" mode.)
Eh? I'm not disputing that at all, I'm a contributor to PyPy, probably the leading python JIT. None of that has anything to do with whether it's semantically possible to do a full type inference on a language like Python.
12 comments
[ 4.5 ms ] story [ 38.8 ms ] threadPlus with an IDE in Scala (or F#) you can just hover for the types.
Take a look at "Declarations as Assertions" section in SBCL manual as well.
The article is talking about partial type inference, and most implementations of type inference I've seen do partial type inference.