Detection And Localization Of Performance Anomaly For Microservices

Microservices is a popular architectural idea to construct applications from a set of small services with independent functions and low coupling.The advantages of microservices such as high cohesion,low coupling,high availability,and scalability make multi-service application architecture gradually become the benchmark for IT application delivery.However,the numerous and complex interactions between components make microservice performance anomaly diagnosis challenging.How to design a performance anomaly detection and root cause location method for microservice architectures is a hot topic of current research in service computing.Many methods have been proposed in academia for performance anomaly detection and root cause location for large microservice clusters.However,these methods suffer from loss of call chain structure information,the coarse granularity of root cause location and are not suitable for low cross-rate scenarios of call chain track,which are insufficient to solve the technical problems faced by microservice performance anomaly diagnosis in production environments.In response to the shortcomings of existing work,we propose a novel method for microservice performance anomaly detection and root cause location.Our method consists of three parts: call chain anomaly detection model,container-level root cause localization algorithm,and root cause promotion algorithm.First and foremost,we propose an unsupervised anomaly detection model based on graph attention networks and variational autoencoders for anomaly detection in call chains.The model uses event representation to construct call events and their relationships into a call chain dendrogram as a feature vector.This method incorporates the microservice features and incorporates the call structure relationship of the call chain,which makes the anomaly detection model have a better fit and thus locate the anomaly to the call chain level.Then,we design a multi-target localization algorithm based on the control variable.Our algorithm uses the control variable method to detect whether the response time of each microservice container in the abnormal call chain is abnormal to locate the abnormality from the call chain level to the container level.After that,we propose a root cause promotion algorithm to construct inter-component dependency relationships through component dependency graphs and use a voting mechanism to promote container-level anomalies precisely to the component level from the bottom up.We evaluate the effectiveness of our proposal in an open-source benchmark microservice environment and an actual production environment.Experimental results show that our solution outperforms existing methods in anomaly detection and root cause localization in terms of accuracy and recall.