F-Q-LiTO: A Federated Q-Learning-Based Lightweight Intelligent Task Orchestrator for Multi-Tenant Container Clusters

Abstract

The demand for intelligent, scalable, and energy-conscious container orchestration has increased due to the growth of microservice-based designs and multi-tenant workloads. A novel federated reinforcement learning framework for adaptive task scheduling in heterogeneous container clusters, F-Q-LiTO (Federated Q-Learning-Based Lightweight Intelligent Task Orchestrator), is proposed in this research. In contrast to traditional orchestrators, F-Q-LiTO uses federated Q-learning to decentralise decision-making, guaranteeing convergence across dispersed nodes while maintaining data locality and minimising synchronisation overhead. The system has several lightweight components, including energy-conscious placement penalties, XOR filters for secure container fingerprinting, Count-Min Sketches (CMS) for constant-space resource estimation, and workload forecasting based on the Long Short-Term Model (LSTM) for proactive migration. In comparison to DeepPlace, F-Q-LiTO reduced task deadline misses by around 34% and achieved an average SLA satisfaction of 96.8% when tested on simulated multitenant workloads with over 1000 tasks. Ablation studies confirm that federated coordination and predictive migration materially improve performance. Global Q-values converged within six episodes, and SHAPbased explanations identify CPU forecast, SLA urgency, and node energy state as dominant decision factors. F-Q-LiTO demonstrates practical, interpretable, and low-latency orchestration suitable for dynamic edge– cloud deployments.OPEN ACCESS Received: 01/08/2025 Accepted: 09/10/2025 Published: 27/11/2025

Document