MicroDACP: Microservice Fault Diagnosis Method Based on Dual Attention Contrastive Learning and Graph Attention Networks

With the wide application of microservice architecture, the fault diagnosis of microservice software system becomes difficult due to the complex dependencies between microservices. In order to diagnose the faults of microservices quickly and accurately, this paper proposes MicroDACP. First, a contrastive representation of the dual attention mechanism is used to learn to determine whether the microservice system is anomalous. Second, a graph attention network is used to learn the microservice invocation dependency graph, and the fault root cause scores are ranked using an improved PageRank algorithm. Then, a Sock-shop microservice system is built on a Kubernetes cluster to evaluate the performance of MicroDACP. We conduct extensive experiments on Sock-shop dataset, SMD dataset and AIOps 2020 dataset to compare and analyze our method with the baseline, and the results show that MicroDACP achieves improvements of up to about 0.13 in F1 score for anomaly detection and 0.32 in mean average precision for root cause localization.