There are 2 reasons for that (personal opinion):
* performance: MongoMK has an increased latency compared to a local SegmentStore/TarMK.
* availability: you introduce a single point of failure (and you need to operate Mongo)
The only benefit you might have is the reduced amount of storage you need to the NodeStore (the binaries are offloaded into an objectstore, e.g. S3). But even that might be questionable, because a typical Mongo-Cluster consists of 3 nodes, so with up to 3 publishs even that argument does not count.