It is true that data can’t be searched for now (though not because the data is distributed), but search is a crucial feature, so it won’t be like this forever; see, for example, my idea, which is especially geared towards “matchmaking” (semantic hashes.) This, or similar ideas, will be implemented in the future.
As for now, maybe if you can find a way to partition your graph (e.g. as clusters around the more highly connected nodes), and then store the partitions in Structured Data blocks (which are mutable), then you could build a more-or-less searchable graph database even on top of the current architecture without having to resort to a split design with a centralized database. You’ll still need to have to have an application server 
because only the owner can update SD blocks.