A Graph Attention-Based Autoencoder for Critical Path Anomaly Detection in Microservices

In complex microservice architectures, detecting performance anomalies is a critical challenge for ensuring system stability and efficiency. This study introduces CPAnoGAT (Critical Path Anomaly detection with Graph Attention Network and Autoencoder), a novel real-time anomaly detection model leveraging causal graphs and critical path analysis. The model utilizes Graph Attention Networks (GAT) and a critical-path-based edge weighting strategy to focus on crucial system relationships. Features such as operation name, duration, and status codes are embedded as graph nodes, with interrelations modeled for optimal analysis. Experiments on the TrainTicket dataset, a benchmark for microservice architectures, demonstrate that CPAnoGAT outperforms state-of-the-art models such as TraceAnomaly and MultimodalTrace, achieving superior metrics including 99.98% precision, 73.38% recall, and an F1 score of 0.8464. By reducing false positives and enhancing accuracy, CPAnoGAT provides a robust tool for monitoring distributed systems. Future directions include evaluating the model on diverse datasets and integrating real-time anomaly detection in streaming environments.