Autonomous selection of the fault classification models for diagnosing microservice applications

Abstract

Microservices architecture is a new approach for deploying applications and services in the cloud, gaining popularity for constructing large-scale systems that are highly resilient, robust, and adaptable to dynamic customer needs. Given that the number of services in a microservice application can be in the hundreds or thousands, the complex dependencies between them can cause faults to propagate and affect multiple services, ultimately leading to degraded system performance. To run microservices reliably and with high uptime, it is essential to quickly troubleshoot when a fault occurs. Simultaneously, there is a need for unsupervised fault diagnosis methods, as existing supervised methods are usually time and labor-intensive. As a solution to these challenges, this paper proposes the Autonomous Selection of Fault Classification models (ASFC) for Diagnosing Microservice Applications. The method can automatically select optimal models from candidate unsupervised detection models for different faults and cascade them to perform fault diagnosis. Meanwhile, our method can determine the root cause of the fault by conducting a root cause localization of faulty services. Experiments on two widely used benchmark microservices demonstrate that ASFC outperforms the baseline methods. As shown in the experimental results, our method achieves an average macro-F1 of 82.9% and 88.6% on two benchmark microservices, respectively, with an average improvement of 34.4% and 24.7% over baseline methods. Furthermore, the method achieves high accuracy in the localization of the root causes, with an average Avg@5 of 0.857 and 0.525 on two benchmark microservices.