That paper is pretty good but it's comparing bw-tree with much simpler in memory data structures. I think bw-tress might work specially well for fast-disk storage.
I understand the benefits of a single numeric type, but a type that doesn't support full 64 bits is just painful. It makes documentation smaller and benchmarks look good but often becomes a problem in real usages.
I did in my previous job. It was a good experience except around the time we started seeing corrupt LDTs (Large Data Type). I think a proper solution was never found and they eventually deprecated the feature. I'd recommend it IF you really think you can model your business on top of the KV interface.