I think this insight exists across a lot of fields. Basically, if you want to be a really excellent magician you also better be a decent blacksmith. More concretely in this case, if you’re unable to do the data “engineering” yourself then it will close a lot of doors for interesting and novel work on the “science” side. Beyond that, if the scientist’s job just involves gluing sklearn models together I think that job is more on the engineering side of things than the supposed scientist usually wants to admit.