There's probably a sampling bias, because of the attitude that one should learn Haskell "to be a better programmer". The FP universe is far broader than the pure FPs. Need I remind: Twitter went from RoR to Scala, ostensibly for practical (performance) purposes, though with hindsight they could have figured out how to scale RoR. So it's hard to argue that nobody has gone the other way strictly for "practical" reasons.

It's also hard to argue that Whatsapp hasn't been wildly successful at scaling to a billion users on a miniscule engineering team in spite of "impracticality of FP".