The processes that use locks are often short-lived. They live in short-lived containers with no state, or maybe they're just lambdas executing under a strict resource limit. Either way, there's nobody to clean up after them or restart them once they're killed. When they begin a database transaction and then disappear for any reason, the best practice is to roll back and pretend they never did anything.
In this brave new world of YOLO lock holders, antirez's position makes a lot of sense. There's definitely still room for old-fashioned durable locks, but these are different use cases.