Authors: Parav Sharma, Rajesh Chauhan, Akshay Bhardwaj
Abstract: In today's complex cloud-native microservice architectures, the traditional rule-based monitoring approaches are insufficient for system reliability. The solution to this is the integration of Artificial Intelligence (AI) into DevOps, often known as AIOps. This paper introduces an AI-powered observability framework that combines the LSTM (Long Short-Term Memory) autoencoder with the Prometheus-Grafana observability stack for anomaly detection in Kubernetes-based microservice environments. The system collects real-time CPU utilization metrics from different microservices, trains an LSTM model on metric records, and detects anomalies using reconstruction error thresholding. Under controlled CPU stress injection, the characteristic scale of reconstruction error, measured as the mean plus two standard deviations of each condition's own reconstruction errors, is approximately 27 times higher than under baseline conditions, indicating that the model's reconstruction error responds strongly to abnormal workload patterns. Because a single training run may not be representative, the model is retrained ten times on a 6,407-timestep dataset; the reconstruction-error threshold is 0.000676 +/- 0.000038 and the anomaly rate 4.38% +/- 0.26% (mean +/- SD), a coefficient of variation of 5.6%, indicating that the result is reproducible rather than an artifact of one random initialization. The trained model is then deployed as a continuous detector that scores live metrics every 60 seconds; in a controlled test it flagged all 17 readings taken while an injected workload was active and returned to normal after its removal. The LSTM-based model detects anomalies without relying on fixed alert boundaries, offering a more adaptive alternative to static threshold approaches and enabling proactive detection of system abnormalities in cloud-native DevOps environments.
