Slapping some scipy, sklearn, or pandas stuff into Django was a pretty workable solution even back when. Just want to sing the the praises. I've slayed a lot of normal size (as opposed to big) data problems with Django/Flask. Rub some JS charting on top of it and people thought you were a wizard.
Oh it was workable. Until it wasn't. I specifically remember extricating a high throughput web service from pandas. Having to scale made that stuff harder. If it was some app that a few thousand people were using, no problem.
Even then, scaling is an almost self-solving issue, as it usually comes with money attached, which buys quite a lot of engineering hours spent on scaling things. Overscaling in advance on the other hand with all of its associated complexity and slowdown killed many, many companies.
Inability to scale because the application was fundamentally unscalable has also killed many companies. And for those it didn't kill it significantly hurt their earning potential (and ultimately everyone's payout).
There's a balance and it's more delicate than people give it credit these days. Don't get me wrong though... over-engineering and/or premature optimization is just as bad. But there is an opposite extreme as well.
