Microservice Traces Anomaly Detection Based on Structural Clustering and Graph Attention Autoencoder

In a distributed microservice system, the trace can reflect the operating process of the system from a macro perspective, and it is of great significance for detecting abnormalities that occur during system operation. However, currently running systems have a large number of different microservices defined, resulting in complex trace structures and increasing the difficulty of anomaly detection. At the same time, both the call structure and response time of trace contain abnormal information, but existing methods focus more on detecting anomalies through response time, ignoring the call structure information, resulting in inaccurate anomaly detection. In this paper, we propose a trace anomaly detection method based on structural clustering and graph attention autoencoder. Firstly, clustering is used to classify the traces of calling similar microservices into the same category, in order to reduce the difficulty of anomaly detection. Secondly, an improved graph attention autoencoder is used to learn the structural information and response time distribution of each service. Finally, the best F1 method is used to obtain the threshold, and the presence of anomalies in the trace is determined by comparing the threshold with the obtained reconstruction loss. Experiments on public data sets show that our method has better performance than currently existing methods.