Microservice architectures are inherently vulnerable to partial failures, and they rely on resilience patterns to tolerate them. However, it is challenging to design and implement the resilience logic. Unforeseen interactions with faulty services can lead to errors in dependent services, resulting in incorrect system behavior. Fault injection testing techniques aim to uncover these errors by examining the system’s behavior in response to different fault combinations. However, existing automated techniques primarily focus on service-level fault injection, which limits their broader applicability.

We present an automated fault-injection testing method that operates at the network level, enabling broad applicability. The method models the system’s resilience behaviors dynamically through the observed test executions and uses this information to reduce the set of fault combinations to explore. We implemented our method in a prototype tool, called Reynard. Our evaluation demonstrates that Reynard efficiently explores system executions by significantly reducing the number of fault combinations to test. It incurs minimal overhead and can be easily integrated into existing benchmarks. Furthermore, we applied Reynard to test an industrial system, showcasing its applicability and effectiveness in a real-world context. Moreover, we uncovered a previously unknown resilience bug.