Think email. When you send an email and CC five other people as well then seven people now have the same copy of the email stored on their email servers. That is, there’s no central database that contains a single email that is referenced by others.
This is basically how sharding with relational DBs works as well.
This sort of data denormalization is almost a requirement as applications scale and especially for many-to-many applications that have a high write to read ratio.
Low write to read and you can get away with a single master to many slave relational DB architecture for quite astonishing numbers of requests and data!
*(actually they just onboarded a second production PDS yesterday.. progress!)
Haven't done any research to determine if there are plans for direct messages.