Right if you want to use a sharded and replicated system with a document data model, might as well use the real thing rather than introduce the risk and impedance mismatch of an emulation layer. The whole point here is for folks to not have to have the cognitive and performance downsides of a tabular data model