A Deep Graph Neural Networks Approach for Service Failure Analytics

Microservices make it challenging for operators of cloud-based services to keep higher reliability. To strive for better reliability in service failure detection, some solutions use traditional approaches in their analytics that lead to missing or loss of important characteristics of telemetry data and thus cannot detect the impact of service failure on users. To enhance the detection of the impact of service failure, this paper contributes (i) a hidden service failure scenario based on a realistic use case, and (ii) a deep graph neural networks-based approach for improving service failure analytics. We built deep graph neural networks (DGNNs) to capture the diverse characteristics of telemetry data from observability practice. Our evaluation results, which are built on a digital payment use case, showed that DGNNs improved the detection rate of service failures by more than 30 percentage points compared to the traditional machine learning approaches.