Understanding Data Races in MySQL

Data races are notorious for their close relationship to many painful concurrency bugs that are very difficult to locate. On the other hand, it is both impossible and unnecessary to forbid every data race in large software systems, by locking every shared variable. It is impossible since we cannot afford for the performance deterioration by locking everything. It is unnecessary since most of the data races are actually benign, i.e., they do not compromise program’s correctness. In this paper, we study the behavior of data races inside MySQL, a modern database management system. We collect a large set of data races reported by Helgrind, after running a benchmark on MySQL that contains SQL query workloads coming from typical database applications. We then carefully examine these races and work out a taxonomy on programming paradigms that can cause data races. To learn the semantics implied in these paradigms, we further study the roles of the shared variables that are involved in the races, based on their read/write histories. Our findings in this paper will be helpful for programmers to get a deeper understanding on how to write more reliable multithreading programs.