Michael
Abebe,
PhD
candidate
David
R.
Cheriton
School
of
Computer
Science
Workload access patterns can limit the performance achievable with static data placement in distributed database systems. Dynamic physical database designs in which data item master locations, partitioning schemes and replica locations can vary with the workload help improve system performance.
I will present DRP, a system that dynamically and iteratively alters the physical design of a distributed database as transactions execute. DRP supports changing the data partitioning, replication, and master placement schemes as first-class operations. DRP alters the physical design using a learned cost model that can quantify the trade-offs for each physical design change. Using these techniques, DRP provides low latency transactions, adapting the distributed physical design based on workload observations.