[im_down_w_otp]

We kind of ran into a similar problem. Our core data (multi-context execution/event traces) is fundamentally graph shaped, but it's also among the pathologically poor graph structures for most general purpose graph databases and their associated query DSLs and execution planners/optimizers, so we had to build a solution that was tailored more for our domain (verifying system properties of complex interacting components).

Which I suppose kind of typifies the problem. Graph databases are fantastic because they let you flexibly and coherently model practically anything. But, perhaps principally because of this, they can become an impediment once you better understand the nuances and idiosyncrasies of your domain, and thus need something that has more optimal (or perhaps predictable) performance for the kinds of questions you know you need to ask over a representation of your domain/data that you know is sufficient?