Can't you translate GraphQL queries to Joinmonster/Dataloader/Haxl calls? It's up to the implementor of the GraphQL <-> backends layer to build these optimisations, and you could reuse those libraries for that.

Haxl seems particularly suited for this (I think it's the same thing as Dataloader, both from facebook?) as it can efficiently batch and cache queries to multiple sources at the same time. It seems like a good layer between GraphQL and your N data sources.