This abstraction will definitely leak (or performance will be horrible, pick your poison):

* A big benefit for SQL are transactions, I have a hard time seeing anything done efficiently with MongoDB.

* Additionally, joins cannot be consistent with the MongoDB model (shards, reading from secondaries, ...).

While this idea is charming, I don't see this reaching any serious production environment.

But it's cool, Citus implemented a somewhat working prototype which is useful for some use cases. Kudos ! :)