Self-Supervised Spatio-Temporal Representation Learning for Microservice Anomaly Detection

Microservice architecture is widely adopted in massive Internet applications due to its advantages such as agility and scalability. It is important to accurately detect anomalies to ensure the reliability of microservice systems. Unsupervised methods based on prediction or reconstruction have been proposed to monitor the time series of metrics for each virtual component (named pod). However, such methods have only used normal data for training representation of time series therefore the difference between normal and anomaly data is not fully represented. Moreover, the 2D spatial-temporal dependency among metric time series have been considered in literature but the interdependency among metric data implied by the physical co-deployment of components in the microservice system has not been considered. In this paper, we propose a self-supervised spatio-temporal representation learning model, named 3DCNN-VAE, for microservice anomaly detection. We customized VAE to learn comparative representations between normal and anomaly to construct discriminative features for multivariate time series representations. 3DCNN is introduced as an encoder in VAE to extract 3D spatio-temporal dependencies between multivariate metric data. Meanwhile, we use GNN to reduce the influence of metric indexing on feature extraction. The latent space representation in the VAE structure is used as the pre-trained time series representation, and the VAE decoder is transformed into MLP to complete downstream classification tasks for anomaly detection. Results on realistic dataset show that 3DCNN-VAE outperforms the best baseline by 9.8% on recall and 4.2% on F1-score.