Take the approach of the Singularity project and apply it to OLTP databases.
How would yo architect such a DBMS? Would you use the same abstractions (such as sealed processes), or what others would you come up with, and why? How would DBMS profit from this? How will they be better than classical OLTP DBMS?