Adaptive Scheduling Architecture for Multi-Cluster Kubernetes SystemsPetrović, Bogdan (Avtor) Jurič, Branko Matjaž (Mentor) kubernetesedge computingdockerThe default Kubernetes scheduler optimizes pod placement based on resource availability but does not account for real-time network conditions, limiting its effectiveness in latency-sensitive and geographically distributed environments. This thesis presents an adaptive, latency-aware scheduling architecture for multi-cluster Kubernetes systems. The proposed system introduces geo-distributed edge proxies that continuously measure round-trip latency to worker nodes and track incoming traffic intensity. These metrics feed into an adaptive scoring model that combines weighted-average latency with traffic-based edge-proxy importance, using exponential smoothing and sigmoid normalization. The scheduler uses these scores to place pods on optimal nodes across multiple clusters without relying on external federation frameworks. Experimental evaluation in a simulated multi-cluster environment with dynamically varying network conditions demonstrates that the proposed approach reduces the proportion of requests exceeding 500ms latency from 37.6% to 4.7% in simulated degradation scenarios. The system successfully adapts to changing network conditions and traffic patterns, which is an important factor for latency-sensitive applications.20262026-05-12 11:28:13Magistrsko delo/naloga182454VisID: 38415COBISS_ID: 278096899sl