In enterprise container platforms, heterogeneous clusters where overcommitted virtualised nodes co-exist alongside dedicated baremetal nodes demand scheduling systems that reason about physical resource reality, not merely declared requests. Kubernetes treats every virtual CPU core as equivalent in capacity, causing performance-critical workloads to experience throughput degradation on overcommitted nodes while cost-efficiency workloads unnecessarily occupy baremetal processors that could be reserved for latency-sensitive applications. This thesis proposes, implements, and evaluates a node-aware Kubernetes scheduler plugin that integrates live Prometheus metrics and a hypervisor-level monitoring mechanism into the scheduling decision pipeline. The plugin implements two scoring modes: performance mode, which steers workloads toward baremetal nodes by structurally reducing the effective capacity of virtualised nodes according to their overcommitment ratio; and cost-efficiency mode, which maximises pod density on overcommitted guest nodes within safe utilisation thresholds. A custom Prometheus exporter surfaces L0 hypervisor utilisation metrics, enabling detection of shared hypervisor host stress entirely invisible to standard Kubernetes metrics. The prototype assumes pod and node labels are present and uses CRD-defined overcommit classes cached locally at runtime. Validated on a live OpenShift cluster at IBM Deutschland Research and Development comprising two baremetal IBM Z Logical Partitions (LPARs) and two overcommitted z/VM guest nodes, the scheduler achieves 100% placement accuracy across all workload classes, reduces physical CPU consumption by 35.5% for cost-efficiency workloads contributing 28.6% beyond the overcommit operator alone and correctly avoids a loaded z/VM node scoring 36 versus 100 for idle baremetal, preventing a measured 22.8% throughput penalty at the node level. All results are achieved at a scheduling overhead of only +2.6 milliseconds relative to the default Kubernetes scheduler. These results demonstrate that overcommitment-aware, workload-class-sensitive scheduling is both technically feasible and operationally viable, establishing a practical foundation for extending Kubernetes scheduling capability to heterogeneous enterprise infrastructure.