I was surprised too, when I learned this technique, that people weren't already doing it. I think it's just an age thing. B-trees are old so they have all the kinks worked out, at least in mature implementations. Fractal trees are new enough that there are still a bunch of tradeoffs to play with and evaluate.

When you actually go and implement the system, you find all these behaviors that really aren't expressed in the theory. There are lots of things we're still experimenting with. Flushing isn't too hard, if a flush makes another node way too big, you can just flush that one too. We don't allow our cascading flushes to fan out, to prevent latency-style problems, but they can flush all the way down to a leaf. There are some other fun problems around query strategies and concurrency control too. It's definitely a fun structure to play with.