Fighting Back: Using Observability Tools to Improve the DBMS (Not Just Diagnose It)

Observability tools increase the transparency of the system and help practitioners identify a wide variety of bugs and bottlenecks. In this paper we claim that such tools (DTrace, in our case) can play a much more active role: the same features which allow effective, but largely passive, measurement also have the potential to provide a powerful tool for actively improving the system. This capability proves particularly useful with database engines, which have complexity approaching that of operating systems but are nevertheless hosted by -- and often impeded by -- an underlying OS. We illustrate our point with several non-traditional use cases for DTrace where the DBMS and OS communicate, leading to elegant solutions for OS-DBMS interaction problems that are otherwise difficult to solve.