I've been implementing row polymorphism in my fork of Elm in order to support proper sum and substraction operations on unions, and it's far from trivial.

Example usecase: an Effect may fail with 2 types, but actually you have handled one/catched one so you want to remove it.

Elm-like HM systems handle fine, as you say it, row polymorphism mostly over records.

I'm not an expert in all of this, started studying this recently, so take my words with a grain of salt.