Cloud-native systems resilience assessments based on kubernetes architecture graph

Chaos engineering is a popular approach in cloud-native system resilience assessments. It uses fault injection to test if the system is still reliable. For Kubernetes platform, which is widely used for deploying deep learning models in AI-powered system, chaos engineering mainly uses "random selection and injection" with "single-point faults", which may ignore "deep" system failures involving combinations of different instances and fault types, hence failing to assess the resilience effectively. To address this problem, we first analyze the injection-points combinations and fault-types combinations in the Kubernetes platform. Then we propose a novel method called RA-KAG (Resilience Assessments Based on Kubernetes Architecture Graph), which assesses the resilience of systems on Kubernetes. The Kubernetes architecture graph contains both architecture-level information and real-time monitoring information of the cloud platform and the system under test. By applying graph computation algorithms and constraint filtering, combinations of injection-points with high relevance and high network traffic are selected to create failure scenarios with higher quality. RA-KAG is tested on the microservice test suite Sock-Shop and the experimental results show that it is more efficient and effective than the baseline. RA-KAG improves the rate of valid test cases in all test cases by 206% on average with 2-points fault injection, and 105% on average with 3-points fault injection. It can also cover a variety of valuable injection-points combinations modes and fault-types combinations modes.