Current architectures for main-memory online transaction processing (OLTP) database management systems (DBMS) are based on one of two design choices. In the partition choice, the data is assumed to be well partitioned. Transactions run with little or no concurrency control inside a partition. In the non-partition choice, the data is not required to be partitioned and the system carefully controls communication to achieve high performance. The partition choice has the advantage of performance if the data partitions well. The non-partition choice offers flexibility, with a loss in performance in some cases.

In this talk, I will present a third, middle way, between these two design extremes: partitioning is not required, but statistically based transaction scheduling is used to achieve a performance (almost) as good as the partition choice. To understand the design trade-offs of our new approach, we construct an extension to a main-memory OLTP system that implements a variety design decisions.

Bio:
Tieying (Michael) Zhang is a Postdoc in the Computer Science Department (Database Group) at Carnegie Mellon University, where he is working with Andy Pavlo, on large-scale database systems. Prior to coming to CMU, he was an assistant professor at Chinese Academy of Sciences. He received his Ph.D. degree in 2011 from Chinese Academy of Sciences.