Typically, the best you can hope for out of your data is that it is a directed graph. Relational databases exist in a sweet spot of efficiency, widespread availability, and fairly decent ability to represent directed graphs.

But you need to know the shape of your graph first. The document-store model tends towards giving up the graph entirely and just going with a tree. Us programmers love tree structures. They make things nice and tidy and can be traversed in finite time.

But I've never worked on a project that could easily be modeled as a tree. Unfortunately, this was often realized after I had been working on the project for a year. You know how phenomenally fucked you are having a tree-like data structure when what you really need is a directed graph? You're basically all the fuckeds.

Directed graphs generalize trees, so trying to shoehorn a tree where a directed graph is needed is basically bringing a Fischer Price tool set with you to Habitat for Humanity.

Thus, from experience, I never start with a tree now. I always start with a directed graph. If I learn that the data is a tree after a year and not actually a directed graph, oh well, look at all this flexibility we didn't need.