The recurring problem I've found with node-based visual programming is that it's very challenging to work with multiple layers of abstraction and that refactoring can be incredibly challenging.
Those are definitely two of the major problems that need to be solved for visual programming to really work. Many people look at the problems and just throw up their hands, but I personally think they can be solved. I look at things like people building giant systems in Minecraft, something definitely not even designed for laying out logic or processing, and the way they talk about the physicality and geometry of the processing they lay out suggests to me that some breakthroughs might be right around the corner. Years ago when Supreme Commander was released (the game), they introduced an interface where you held down Ctrl and scrolled the mouse wheel in order to 'zoom' from high level overviews all the way down to single unit detail. It felt so intuitive and natural after using it for a bit that I found myself accidentally trying to do it in a file manager application after exiting the game. I think that would work for visual programming as well. And who knows the possibilities that VR brings along....