VAE-based Fault Diagnosis for Microservice System

Microservice architecture is widely used on large-scale Internet systems, and the data generated during system operation such as Traces and Metrics are useful in diagnosing system faults, but the heterogeneity of different data makes it difficult to model the normal behavior of the system and detect their faults. In this paper, we propose an unsupervised anomaly detection model based on a variational autoencoder (VAE) for anomaly detection in microservice domain. We find that anomalous data exhibit temporal and spatial anomalies, and represent various types of faults uniformly, to address these, we propose a multi-source heterogeneous data fusion method and then use a two-layer VAE model to reconstruct the spatial structure and temporal information of the fused data, respectively. In addition, to address the problem that the VAE model cannot make use of the global information of the graph data, we also propose to learn the hypersphere of graph data representation. On two microservice datasets, our proposed method obtains a better AUC.