An empirical analysis of database recovery costs

The time required for recovery from a failure is heavily inuenced by hardware setup and workload characteristics. In bad but still realistic cases, the recovery required during restart can take hours. For a database system based on write-ahead logging, we performed a qualitative study of how hardware and software congurations aect the behavior of the database and, consequently, how this behavior aects recovery time after a system crash. With the relevant parameters identied in the qualitative study, we performed an empirical quantitative analysis of recovery costs in multiple scenarios. We show that recovery costs tend to get worse as hardware and software improve in eciency, and we discuss possible approaches to make recovery time independent of system congurations and workload characteristics.