Issues of correctness in database concurrency control by locking

Our aim in this paper is to show that there is a mathematically inherent reason why existing systems enforce D-serializability (rather than just because of its simplicity): it is because they are based on locking. Our main result is a characterization of the power of locking which states that if a locking policy is safe then it must allow only D-serializable schedules. Furthermore any such schedule can be produced by some safe locking policy. The rest of the paper is organized as follows. In Section 2 we formalize our concepts and describe the model. In Section 3 we characterize D-serializability in semantic terms. In Section 4 we examine when a set of transactions can be let to run safely by themselves without locking or any intervention from the scheduler. Section 5 is concerned with locking policies and in Section 6 we discuss some implications of our results.