Deep Q-Learning Based Circuit Breaking Method for Micro-services in Cloud Native Systems

More and more modern applications are being deployed on cloud-native systems. High concurrent requests that exceed the processing capacity lead to service failures, and the failure of a single microservice result in the entire application unavailable. Currently, most existing methods rely on manual adjustment of microservice’s circuit breaking strategies based on empirical experience. It is very complex to select an appropriate circuit breaking strategy, considering the processing capacity, ect. In this paper, a reinforcement learning-based method is proposed which tries to select optimal circuit breaking strategy by interacting with the system continuously. The proposed method is evaluated on a real Kubernetes and Istio based cluster. Experimental results demonstrate that our approach achieves a successful service rate of 92.6% for requests while 7.4% of requests experience failures. In contrast, the static circuit breaker achieves an average of 48.4% successful service and 51.6% failure rate.