klights

klights is a resource-efficient, event-driven Kubernetes-compatible cluster runtime. pronounced K-light-s.

Packaged with both ARM and x86 build on Ubuntu 24.04/26.04 and RHEL 9/10.

Start with QUICKSTART.md.

Build, run, configuration, and operations documentation doc/README.md.

Sonobuoy Conformance tests

Klights completed Sonobuoy Conformance tests with 3 raft controlplanes, one replicas and 2 workers nodes.

• Sonobuoy Version: v0.57.3
• Ran 424 of 7144 Specs in 2260.475 seconds
• SUCCESS! -- 424 Passed | 0 Failed | 0 Pending | 6720 Skipped

Baseline Memory usage

startup, with klights running and only CoreDNS deployed:

• klights process RSS: approximately 76.1 MiB
• klights process peak RSS: approximately 76.5 MiB
• klights.service cgroup: approximately 150.6 MiB
• klights.service peak: approximately 167.9 MiB

Processes inside klights.service at that point included:

• klights: approximately 76.1 MiB RSS
• containerd: approximately 54.4 MiB RSS
• containerd-shim: approximately 10.7 MiB RSS

The idle baseline was therefore roughly:

• klights binary only: approximately 76 MiB RSS
• klights + runtime cgroup: approximately 151 MiB

The cgroup number includes embedded runtime overhead, not just the klights process.

This beta currently offers:

• Near-zero CPU use during idle through async, event-driven runtime paths.
• A target minimum RAM requirement of 200 MB in the beta release.
• A single-node leader mode with embedded API server, scheduler, controllers, datastore, kubelet-facing runtime integration, and local networking.
• Kubernetes-compatible API access through the generated kubeconfig, so kubectl and Kubernetes clients can talk to klights as a cluster endpoint.
• Raft control-plane mode with exactly three control-plane voters.
• Single-leader mode, optionally paired with replica control-plane learners for manual recovery workflows.
• Worker-node and control-plane node joins with separate bootstrap tokens.
• Rootful container runtime integration through containerd and klights-managed CNI configuration.

Upcoming work includes but is not limited to:

• Rootless operation.
• Hybrid cluster with rootless and root nodes, using built-in CNI.
• CNI plugin support for standard Kubernetes CNI providers such as Calico and Flannel.
• Removing the containerd dependency.
• A pluggable datastore backend, with redb as the first target.
• Continued performance and stability improvements.
• An event-driven gRPC API as an alternative to stock Kubernetes polling API access patterns.
• GitOps deployment tooling built on the event-driven API, with near-zero CPU use during idle.

Current release support is limited to rootful local development mode. Rootless operation, hybrid rootless/root clusters, expanded CNI plugin support, and containerd-free runtime support are not available in this release.