Unsupervised Anomaly Detection in Microservice Telemetry: Identifying Silent Failures via LSTM Autoencoder Reconstruction Errors

Modern cloud-native applications built using microservice architectures generate high-dimensional telemetry data, making the detection of silent failures and performance anomalies increasingly challenging. Traditional rule-based and supervised monitoring techniques often fail to identify previously unseen or subtle anomalies. This paper presents an unsupervised anomaly detection approach for microservice telemetry using an LSTM-based autoencoder. The model is trained exclusively on normal operational data and identifies anomalies by analyzing reconstruction errors across multivariate time-series metrics. By learning temporal dependencies and service interaction patterns, the proposed method effectively detects silent failures that do not trigger explicit alerts. Experimental evaluation on microservice telemetry data demonstrates that reconstruction error distributions provide a reliable signal for anomaly detection, enabling early identification of latent faults and performance degradation. The approach is scalable, does not require labeled anomaly data, and is well-suited for dynamic microservice environments. The implementation is developed in Python and leverages deep learning frameworks for time-series modeling and evaluation.