Skip to content

RKE2

Rancher’s security-focused Kubernetes distribution that runs the control plane as static pods with an embedded containerd and has a documented air-gap installation.

ID rke2
Layer orchestration
Version v1.37.2+rke2r1
Tiers T1, T2, T3
Requires nothing
License Apache-2.0
Evidence docs
Last verified 2026-11-22
  • A hardened, compliance-oriented Kubernetes distribution is wanted: the introduction page says RKE2 focuses on security and compliance for the U.S. Federal Government sector, provides defaults and options that let clusters pass the CIS Kubernetes Benchmark (v1.12 or v1.11) with minimal operator intervention, enables FIPS 140-2 compliance, scans components for CVEs with Trivy in its build pipeline and does not rely on Docker (control plane as static pods, embedded containerd).
  • A documented air-gap route with files is needed: the air-gap page gives four ways to load images (private registry, manual copy of the archive, Hauler, embedded registry mirror) and an install by binary or by script with INSTALL_RKE2_ARTIFACT_PATH.
  • A very small single-host or edge deployment is all that is needed and the compliance features do not matter: the K3s introduction page targets edge and small clusters with a single binary of under 100 MB; the RKE2 requirements page gives a minimum of 2 CPU and 4 GB RAM per node (recommended 4 CPU and 8 GB).
  • GPUs: the GPU Operators page installs the NVIDIA GPU Operator on a running RKE2 cluster with a HelmChart resource that points to the chart repository https://helm.ngc.nvidia.com/nvidia (chart gpu-operator, version v26.7.2 on the default tab; v26.3.x is marked deprecated and v25.10.x end of support), with the toolkit’s CONTAINERD_SOCKET set to /run/k3s/containerd/containerd.sock. It requires the NVIDIA kernel drivers and the libnvidia-ml library on the host, installed by the operator or, for supported operating systems or a vendor-supplied driver image, by the GPU Operator; the page says installing the NVIDIA components on the host is out of its scope. v26.7.x requires containerd 2.0 or later, the DRA mode (Kubernetes 1.34 or later, NVIDIA driver 580 or later) is marked experimental, and the page warns that the operator restarts containerd, which restarts RKE2. The drivers, the GPU Operator, its chart and the images it deploys (device plugin, container toolkit and others) are outside the catalog; the page describes no offline procedure for them and shows the chart repository on the internet.
  • The packaged components (the packaged components page names rke2-coredns and rke2-metrics-server as examples) are deployed from the images in the archives; the image lists published with the release (rke2-images-*.txt files), not this entry, say which images a given CNI choice needs.
  • No paid edition or license check of RKE2 itself was found in the RKE2 documentation read; SUSE support and the Rancher Prime distribution are separate offerings that this entry does not describe.

No container images.

Packages:

  • rke2-images.linux-amd64.tar.zst
  • rke2.linux-amd64.tar.gz
  • sha256sum-amd64.txt
  • install.sh

What the software tries to fetch while it runs, and how to stage it instead.

  • Install script downloads: without INSTALL_RKE2_ARTIFACT_PATH, install.sh finds the release through https://update.rke2.io/v1-release/channels (INSTALL_RKE2_CHANNEL_URL, channel stable by default) and downloads the release files from https://github.com/rancher/rke2/releases/download (INSTALL_RKE2_ARTIFACT_URL), and on hosts where yum exists it uses the rpm method by default (source at the release tag); the documentation’s offline mode runs it with INSTALL_RKE2_ARTIFACT_PATH pointing to a directory that holds rke2-images.linux-amd64.tar.zst, rke2.linux-amd64.tar.gz and sha256sum-amd64.txt, and says not to run the curl commands on the offline hosts.
  • Container images outside the image archive: with the private-registry route, the images of the archive are loaded into Docker with docker image load, retagged and pushed to a private registry, registries.yaml is written on each node and the configuration file sets system-default-registry to the registry’s host and optional port. Images that are not in the archive are pulled from their registries unless a mirror and the settings of the private registry page cover them.
  • rke2-selinux RPM: the air-gap page says it must be installed manually on air-gapped nodes with SELinux enabled, with container-selinux, selinux-policy-base, selinux-policy, policycoreutils and selinux-tools (SUSE) or libselinux-utils (RHEL) available offline, and iptables-nft and libnftnl when RKE2 itself is installed from RPM. The exact RPM file names are in airgap.notes. The operating system’s packages are outside the catalog.
  • Automated upgrades (optional, not part of the first installation): the air-gap page says the images rancher/rke2-upgrade (tag of the RKE2 release with + replaced by -), rancher/system-upgrade-controller and rancher/kubectl must be in the private registry, with the last two read from the system-upgrade-controller manifest; the page’s version numbers are examples and this entry names no tags for them.
  • Security responder (enabled by default since v1.37, see telemetry): a CronJob that sends one request every eight hours; the documentation says it operates in disconnected and air-gapped environments without requiring a successful request, retries a failed request up to three times and does not queue data. Its image, docker.io/rancher/rke2-security-responder:v0.1.9, is in the release’s image lists (rke2-images.linux-amd64.txt and rke2-images-core.linux-amd64.txt), so it comes with the image archive.

No license server is contacted.

What How to turn it off
RKE2 security responder (bundled with RKE2 v1.37 and newer, enabled by default): a CronJob in kube-system, every eight hours, sends one request with non-identifying cluster metadata to a service managed by SUSE and logs the available newer versions and CVEs the answer lists. In the default mode (recommended) the documented fields include the Kubernetes version, a cluster UUID (the UID of the kube-system namespace, for deduplication), server and agent node counts, allocatable CPU and memory, the CNI plugin and ingress controller with versions, operating system, kernel and architecture, SELinux status, GPU node count, vendor and GPU operator version, whether Rancher manages the cluster with its version and installation UUID, the system default registry and the IP stack. The documentation says it collects no workload data, logs, secrets, hostnames, cluster names or IP addresses as payload fields. The endpoint, from the README of the responder project and not from the documentation page, is https://security-responder.version.rke2.io/v1/check (chart value check.endpoint). Add disable: [rke2-security-responder] to /etc/rancher/rke2/config.yaml on each server node and restart RKE2 (the documentation says no further reports are collected). To keep the update notices but send less, create a HelmChartConfig named rke2-security-responder in kube-system with valuesContent mode: minimal (default is recommended).
Traefik version check in the default ingress controller. The networking services page says Traefik is deployed by default from RKE2 v1.36, and the ingress migration page says that from v1.37 only traefik or dual ingress are supported, so v1.37.2 deploys Traefik (the chart rke2-traefik 42.1.301, from charts/chart_versions.yaml at the release tag). That chart’s values (read from the chart file at https://rke2-charts.rancher.io) set global.checkNewVersion: true and global.sendAnonymousUsage: false, and no RKE2 file read overrides them (searched the repository at the tag and the RKE2 documentation for checkNewVersion and sendAnonymousUsage; the RKE2 documentation does not mention either). Traefik’s documentation (the data-collection page at v3.7.16, the Traefik version RKE2 v1.37.2+rke2r1 ships: the chart’s appVersion is v3.7.16 and scripts/build-images at the tag names rancher/hardened-traefik:v3.7.16-build20261028; the page is identical at v3.7.15) says checkNewVersion is enabled by default and that Traefik then periodically contacts update.traefik.io, sharing the running Traefik version and the public IP address of the instance, and states that the collected IP addresses are also used for marketing purposes; sendAnonymousUsage (a daily report of version and an anonymized static configuration) is opt-in and off here. Create a HelmChartConfig named rke2-traefik in the kube-system namespace, for example at /var/lib/rancher/rke2/server/manifests/rke2-traefik-config.yaml, with valuesContent global: checkNewVersion: false (the networking services page shows a HelmChartConfig for rke2-traefik and the Helm add-ons page describes the mechanism; the key global.checkNewVersion is from the chart’s values and Traefik’s data-collection page). Alternatively ingress-controller: none in the RKE2 configuration deploys no ingress controller (server configuration reference), which this design may or may not accept.

Offline behavior is documented but not verified: the entry author did not run RKE2. The files in airgap.packages are for linux/amd64 (x86_64) hosts and follow the documented script install: rke2-images.linux-amd64.tar.zst is the image archive (it includes the images for the Canal CNI plugin; load it with the private-registry route or copy it to /var/lib/rancher/rke2/agent/images/ on every node), rke2.linux-amd64.tar.gz is the binary tarball, sha256sum-amd64.txt the checksum file that install.sh uses to verify the artifacts, and install.sh the script from https://get.rke2.io, all kept with the same relative names in one directory. The images inside the archive are not listed separately. Binary route (documented as the alternative to the script, not a carry item of this entry): download rke2.linux-amd64 instead of the tarball, install.sh and checksums, place it in /usr/local/bin, make it executable and run it with the configuration file. CNI and extra archives named on the air-gap page: instead of rke2-images.linux-amd64.tar.zst the page allows rke2-images-core.linux-amd64.tar.zst plus the archive of the CNI plugin (for example rke2-images-cilium.linux-amd64.tar.zst), and rke2-images-vsphere.linux-amd64.tar.zst when the vSphere CPI/CSI charts are enabled (–cloud-provider-name=rancher-vsphere); the release also publishes archives named rke2-images-NAME.linux-amd64.tar.zst for calico, canal, flannel, harvester, ingress-nginx and multus, and each archive as .tar.gz as well as .tar.zst, which the page does not describe. Files for other architectures, named by the release v1.37.2+rke2r1 (the requirements page says x86_64 and arm64/aarch64): for arm64 the release has rke2-images.linux-arm64.tar.zst, rke2.linux-arm64.tar.gz, rke2.linux-arm64 (binary) and sha256sum-arm64.txt, and eight per-component image archives, each as .tar.zst (and as .tar.gz): rke2-images-core.linux-arm64.tar.zst, rke2-images-calico.linux-arm64.tar.zst, rke2-images-canal.linux-arm64.tar.zst, rke2-images-cilium.linux-arm64.tar.zst, rke2-images-flannel.linux-arm64.tar.zst, rke2-images-harvester.linux-arm64.tar.zst, rke2-images-ingress-nginx.linux-arm64.tar.zst and rke2-images-multus.linux-arm64.tar.zst. There is no vsphere archive for arm64 in this release (rke2-images-vsphere exists only for linux-amd64). The release publishes no Linux file for other CPU architectures. Windows: the release also has Windows files and the documentation has a separate Windows air-gap page, which this entry did not read. SELinux (air-gap page and SELinux page): on air-gapped nodes with SELinux enabled the rke2-selinux policy RPM must be installed by hand, before RKE2; its dependencies are listed in runtime_downloads. The policy RPM is distribution-specific and is not in packages. The air-gap page points to the releases of rancher/rke2-selinux (releases/latest, which on 2026-11-22 is the tag v0.24.latest.1; the tag v0.24.stable.1, published later, has the same file names): the el8 file is rke2-selinux-0.24-1.el8.noarch.rpm, the el9 file is rke2-selinux-0.24-1.el9.noarch.rpm, and the release also has rke2-selinux-0.24-1.el10.noarch.rpm, rke2-selinux-0.24-1.sle.noarch.rpm and rke2-selinux-0.24-1.slemicro.noarch.rpm (each with a .src.rpm); the SELinux page’s own example names the older rke2-selinux-0.23-1.slemicro.noarch.rpm from the tag v0.23.stable.1. The SELinux page says that on an Enforcing node two things must hold on every node, the policy module installed and RKE2 started with selinux: true in the configuration file (or RKE2_SELINUX=true), that the tarball path of install.sh (the one used with INSTALL_RKE2_ARTIFACT_PATH) never installs the policy and only applies the right file contexts if the policy is already installed, and that without it the files carry the wrong contexts (unconfined_service_t) and RKE2 does not work as intended on an Enforcing node. Hosts without a default route need a dummy default route (air-gap page). Documented alternatives to the archive route: the Hauler tool (hauler store add file …, on the connected side, and hauler store load on the air-gapped host) and the embedded distributed registry mirror (registry_mirror page); neither is a carry item of this entry. Version selection: the default channel of install.sh is stable (the script at the tag and the installation configuration page both say so). On 2026-11-22 a GET of https://update.rke2.io/v1-release/channels (HTTP 200, JSON) lists stable = v1.36.6+rke2r1 and latest = v1.37.2+rke2r1, and a GET of https://update.rke2.io/v1-release/channels/stable answers 302 to https://github.com/rancher/rke2/releases/tag/v1.36.6+rke2r1 (the script reads the version from the last segment of that redirect); so install.sh without carried files and without a version pin (INSTALL_RKE2_VERSION) would install an older release than the v1.37.2+rke2r1 this entry pins, and the channel moves over time. With INSTALL_RKE2_ARTIFACT_PATH set, the tarball path of install.sh (do_install_tar in the script at the tag) stages the local checksum file, image archive and tarball and does not call the function that looks up a channel, so no channel is consulted and the installed version is the one of the files carried; the files carried therefore have to be those of the intended release. Outbound calls: the documented security responder and the Traefik version check, both described in telemetry, enabled by default in this release. Two official pages differ in details: the documentation says in minimal mode the Rancher version and installation UUID are sent as an empty string, while the responder README at v0.1.9 says the Rancher version is sent as the string redacted; and the documentation gives the schedule 0 */8 * * * while the README says the minute and the hour offset derive from the cluster UUID. This entry does not resolve either. No other telemetry, analytics or update-check documentation was found on 2026-11-22 (searched the pages of rancher/rke2-docs for telemetry, analytics, usage statistics, phone home, update check and call home; a search is not proof of absence).

For running the tool itself, these skills go further than the catalog:

All values were read on 2026-11-21 from the pages in sources. The version v1.37.2+rke2r1 is the latest GitHub release (published 2026-11-05, not a prerelease); the file names under airgap.packages are the release assets listed by the GitHub API for that release. The rancher/rke2-docs repository has no tags, so its pages are cited at commit 3931bc85192e5d74abdfc43a499b0c6fae2274d2 (the head of its default branch on 2026-11-21); its commands use the example versions v1.33.1+rke2r1 and v1.34.5+rke2r1, while this entry names the files of v1.37.2+rke2r1. The security responder README is cited at v0.1.9, the tag of the image that the release v1.37.2+rke2r1 lists. RKE2 was not run by the entry author: the offline behavior is documented but not verified in a disconnected network.

Sources:

Source of this page: skills/airgap-architect/catalog/components/rke2.toml.