Streaming Telemetry Analytics for AI-Driven Service Performance Prediction in Large-Scale Microservices Architectures

Abstract

Distributed systems today, such as cloud infrastructures, microservice architectures, and hybrid cyber-physical networks, are increasingly difficult to monitor with respect to fault detection and diagnostics because of their large size, complexity, and dynamics. Telemetry data, a record of detailed operational metrics and event logs, is important for understanding how a system behaves and for proactively maintaining it. The paper describes an AI-based telemetry system capable of predicting service performance in a microservices system. The architecture leverages streaming telemetry, such as system metrics, logs, and traces, from cloud-native setups. The data are cleaned and normalized, and their features are extracted using Principal Component Analysis (PCA) to enhance data quality and reduce dimensionality. To learn temporal patterns of dependencies and identify anomalies in service behavior, an LSTM-based deep learning model is employed. A privacy protection layer is a safety measure that ensures the secure handling of sensitive information. The presented model achieves 94% accuracy with excellent predictive performance, low detection time, and few false positives. The comparative analysis reveals that the framework is superior to existing solutions and can be used to reliably monitor microservice performance in real time.