Speaker
Description
Bare-metal Kubernetes deployments eliminate cloud abstraction layers, exposing critical hardware-level friction in latency, routing, and storage. Project Noctilucent presents an architectural framework designed to systematically remove latency and operational overhead across all cluster layers.
To prevent etcd quorum instability during latency spikes, Noctilucent replaces traditional consensus clustering with a primary-secondary architecture that mirrors state to two read-only secondary nodes. Network traffic is streamlined using OVN-Kubernetes in no-overlay mode, leveraging control-plane iBGP Route Reflectors and MetalLB with frr-k8s for direct eBGP service advertisements along the shortest path. DNS lookup latency and Linux conntrack table exhaustion are eliminated using NodeLocal DNSCache agents.
Persistent block storage utilizes OpenEBS Mayastor for low-latency NVMe-over-TCP with strict workload locality, complemented by OpenZFS snapshotters for ARC single-copy caching, native compression, and dataset isolation. Node lifecycle management enforces an explicit state machine covering PXE provisioning (AlmaLinux 10 with ZFS), Kured automated reboot serialization, and Flux-based GitOps workflows, alongside a Slurm bridge for HPC workload offloading. Together, these core components provide a highly optimized, resilient foundation for bare-metal Kubernetes.