It is not only Cabal, but also the constant churn in a lot of packages. For example, conduit was a great package when I used Haskell. But usually I had to update quite a bit of code when I would bump conduit up to the latest version after not touching a program for a year.
Maybe I have been unlucky enough to use Haskell during some transitions (pre-arrows -> arrows rule, let's use them everywhere -> arrows suck, basic IO monad -> iteratees, old fashioned field access -> various lens proposals -> lens). But I never used a language ecosystem before or after where things would change or break so often.
Luckily, Haskell's static typing coupled with the extremely helpful error messages of the compiler makes refactoring a cheerful task: If it worked before and all type errors have been fixed, then we can be quite confident that it works again.
Many areas have stabilized now. For instance, Edward Kmett's lens package is the de facto standard for all things related to records, getters, setters, and all those things on steroids. (Personally, I love how this turned out. Before lens, we quite frankly had the problem that field accesses were much more painful than they are in mainstream object-oriented languages like Python or Java. Nowadays field accesses are as simple as they are in other languages, and in addition to this we have the flexibility of composable first-class lenses.)
... by someone trying to actually address the issues. There are almost always wrong ways you can fix a type error. Fortunately, that's usually the situation at hand...
I don't understand why it keeps receiving so much heat. What does stack do that cabal can't?
Today I use Racket and I won't go back but I am glad Stack is taking over Cabal.