EDFI: Endogenous Database Fault Injection with a Fine-Grained and Controllable Method

EDFI: Endogenous Database Fault Injection with a Fine-Grained and Controllable Method

연구 분야: Databases

학회: BenchCouncil International Symposium on Intelligent Computers, Algorithms, and Applications

Database reliability is an essential issue in many applications that rely on databases, especially in modern artificial intelligence (AI) applications. One common method to uncover the reliability weakness of databases is fault injection, which can introduce faults into the running database to observe and evaluate its reaction on reliability and performance after the occurrence of faults. Moreover, the fault injection can generate a large and diverse amount of realistic data for training and evaluating anomaly detection algorithms, which can also enhance database reliability. However, existing fault injection tools for testing database reliability are either coarse-grained or imprecise to mimic real-world faults, which limits their applicability and effectiveness. In this paper, we present EDFI, a fine-grained and controllable fault injection framework for endogenous database fault. EDFI can inject endogenous database faults for specific SQL statements from specific user connections, and support extensible fault types and scenarios. We demonstrate the effectiveness of EDFI by generating large and diverse training data to validate commonly used anomaly detection algorithms. The results show that EDFI can effectively simulate realistic endogenous database faults and provide valuable insights to improve anomaly detection algorithms.