The model in the article is roughly based on the version in Date and Darwen's work which is almost the same as what you mention: project, rename, extend, join, union and not matching. The summarize by is kind of syntax sugar in this system. I left out some of the operators which weren't used anywhere in the code in the article. Industrial grade conversion from SQL to relational algebra is a real challenge!
SQL to RA is definitely quite a task, but it was pretty fun (up until the end, when it just started getting annoying...). Then again, I think compilers and programming languages are a blast anyway. Haskell makes it even more so :)
When I wrote this code, I wasn't really comfortable using monads, so all of the threading is explicit. If I were to revisit it, I would probably use a state monad, which might also make it easier to make it tail-recursive (as it is, my desugarer is not written in a tail-recursive manner).
How would you handle correlated subqueries when converting from SQL to RA?
I was curious that you used a different version of relational algebra than the one I had been taught. I've usually seen RA described in terms of six fundamental operators: Project (pi), Select (sigma), Rename (rho), Cross Product, Union and Difference. What gave rise to the model you chose?
Anyway, it's great to see that more people are using Haskell. I would love to be able to work in it some day.