Joint evaluation of recovery and performance of a COTS DBMS in the presence of operator faults

Abstract

A major cause of failures in large database management systems (DBMS) is operator/administrator faults. Although most of the complex DBMS available today have comprehensive recovery mechanisms, the effectiveness of these mechanisms is difficult to characterize. On the other hand, the tuning of a large database is very complex and database administrators tend to concentrate on performance tuning and disregard the recovery mechanisms. Above all, database administrators seldom have feedback on how good a given configuration is concerning recovery. This paper proposes an experimental approach to characterize both the performance and the recoverability of DBMS. Our approach is presented through a concrete example of benchmarking the performance and recovery of an Oracle DBMS running the standard TPC-C benchmark, extended to include two new elements: a faultload based on operator faults and measures related to recoverability. A classification of operator/administrator faults in DBMS is proposed. A set of tools have been designed and built to reproduce operator faults in an Oracle 8 i DBMS, using exactly the same means used in the field by the real database administrator. This experimental approach is generic (i.e., can be applied to any DBMS) and is fully automatic. The paper ends with the discussion of the results and the proposal of guidelines to help database administrators in finding the balance between performance and recovery tuning.