HiGraph: Learning Hierarchical Graph for Multivariate Time Series Anomaly Detection in Microservice Systems

Microservices enable flexible, scalable, and efficient application development in cloud environments. A service instance is a chain of microservices running in virtual pods. The service invocation trace and the pod performance metrics are recorded and analyzed to detect anomalies and ensure system stability. Spatiotemporal graph neural network models are popular for time-series representation as they can capture intra- and inter-correlation of time series. However, we find that the performance of anomaly detection depends on the dynamic correlation between containers in microservice system greatly, but has not been considered in existing work. To capture the correlation both on pod level and metric level, we propose a hierachcical graph neural network model, HiGraph, cascading of MetricLayer and PodLayer modules. MetricLayer is used to compute the correlation between metrics through mutual information. PodLayer is used to learn the dynamical correlation of the pods from trace data. It is important to note that trace data continuously records calls, while metric data has a sampling interval, making temporal alignment difficult. To address this, we innovatively use the edge weights of the graph to align them. We count the call frequency between different pods in the trace data over a specific period to update the weights of the spatial correlation graph. Extensive experiments on real data demonstrate that HiGraph improves RMSE by 4.9%, MAE by 4.8%, F1-score by 4.2% and recall by 9.2% compared to the best baseline.