From Load Tests to Live Streams: Graph Embedding-Based Anomaly Detection in Microservice Architectures

A team of six Amazon researchers has developed a novel AI system that uses graph neural networks to detect anomalies in Prime Video's complex microservice architecture. The system, detailed in a paper accepted at FSE 2026, employs unsupervised node-level graph embeddings to learn structural representations from directed, weighted service graphs at minute-level resolution. By comparing embeddings from load tests with those from actual live events, it can identify services that behave differently under real-world traffic—something traditional load testing often misses.

Built on a GCN-GAE (Graph Convolutional Network - Graph Autoencoder) framework, the system flags anomalies based on cosine similarity between embeddings. During evaluation, it demonstrated 96% precision and an exceptionally low 0.08% false positive rate, though recall was limited to 58% under conservative propagation assumptions. The researchers also introduced a synthetic anomaly injection framework for controlled testing, showing the system's ability to detect incident-related services that were later documented in post-mortem analyses.

The technology has already proven practical within Prime Video, providing early detection capabilities during high-traffic events like Thursday Night Football streams and major video-on-demand premieres. While developed for Amazon's specific needs, the methodology provides a foundation that could be applied across other microservice ecosystems, offering a more sophisticated approach to reliability engineering than traditional monitoring tools.