Skip to content

External runtime dependencies#

k0s is packaged as a single binary, which includes all the needed components. All the binaries are statically linked which means that in typical use cases there's an absolute minimum of external runtime dependencies.

However, depending on the node role and cluster configuration, some of the underlying components may have specific dependencies, like OS level tools, packages and libraries. This page aims to provide a comprehensive overview.

The following command checks for known requirements on a host (currently only available on Linux):

k0s sysinfo

Linux specific#

Linux kernel configuration#

Needless to say, as k0s operates Kubernetes worker nodes, a certain number of Linux kernel modules and configurations are required on the host. This essentially stems from the requirement to run containers and set up networking for them.

The needed kernel configuration items are listed below. All of them are available in Kernel versions 4.3 and above. If running on older kernels, check if the distro in use has backported some features; nevertheless, it might meet the requirements. k0s will check the Linux kernel release as part of its pre-flight checks and issue a warning if it's below 3.10.

The list covers ONLY the k0s/Kubernetes components' needs on worker nodes. Your own workloads may require more.

Note: As part of its pre-flight checks, k0s will try to inspect and validate the kernel configuration. In order for that to succeed, the configuration needs to be accessible at runtime. There are some typical places that k0s will check. A bullet-proof way to ensure the accessibility is to enable CONFIG_IKCONFIG_PROC, and, if enabled as a module, to load the configs module: modprobe configs.

Control Groups (cgroups)#

cgroup v2 is required by default, k0s will fail its pre-flight checks if it's unavailable on the node.

As of Kubernetes 1.31, cgroup v1 is in maintenance mode (KEP-4569). As of Kubernetes 1.35, the kubelet will fail to start by default if cgroup v2 is not detected on the node. In order to run Kubernetes with cgroup v1, bypass the k0s pre-flight checks via --ignore-pre-flight-checks and set failCgroupV1: false in the kubelet configuration via the k0s worker profiles:

...
spec:
  workerProfiles:
    # By default, workers will pick up the default profile.
    - name: default
      values:
        failCgroupV1: false

Migrate to cgroup v2 now!

The above escape hatch will not be available indefinitely, so plan accordingly. KEP-5573 states:

We will commit to removing cgroup v1 code but there is not yet a timeline for removal. The removal will be done no earlier than 1.38 to maintain the k8s deprecation policy.

Required cgroup controllers:

  • cpu
  • cpuacct
  • cpuset
  • memory
  • devices
  • freezer
  • pids

Optional cgroup controllers:

No integration with Name Service Switch (NSS) APIs#

The k0s Linux binaries, including the ones distributed on the GitHub releases pages, are by default statically linked and do not depend on the host's C library. This ensures that k0s can run seamlessly across a wide range of Linux environments. The k0s executable and most of the embedded binaries are built without cgo and use the pure Go resolver, which only consults the files and dns sources of /etc/nsswitch.conf. The Go executables that do require cgo, such as containerd and runc, as well as all non-Go executables, are statically linked against musl libc, which doesn't implement NSS at all. In all cases, k0s cannot use glibc's NSS APIs, which require dynamic linking.

This limitation is particularly relevant when a system uses NSS plugins, such as nss-myhostname, for resolving network names like localhost. Systems lacking a dedicated stub resolver capable of handling localhost DNS queries specifically will encounter issues running k0s. To mitigate this, users are advised to either activate a stub DNS resolver, such as systemd-resolved, or to manually add localhost entries to the /etc/hosts file as shown below:

127.0.0.1 localhost
::1 localhost

The same limitation applies to systems that resolve all names via NSS, such as through nss-resolve on hosts running systemd-resolved. Programs linked against glibc can resolve names just fine on such systems, even if /etc/resolv.conf doesn't list any nameservers. In contrast, k0s and its embedded components only ever consult /etc/resolv.conf. If that file has no nameserver entries, they fall back to querying a nameserver on localhost, which usually doesn't exist. This leaves k0s without working name resolution, although the host itself appears fine. Make sure that /etc/resolv.conf lists reachable nameservers. When using systemd-resolved, this typically means making /etc/resolv.conf a symbolic link to /run/systemd/resolve/stub-resolv.conf, as described in the systemd-resolved documentation. Pods can't reach that stub resolver on the host's loopback interface, which is why k0s configures the kubelet to use systemd-resolved's uplink file /run/systemd/resolve/resolv.conf for pods instead. The k0s pre-flight checks detect this situation by resolving a name under the reserved invalid top-level domain, which any working resolver answers with NXDOMAIN. If that lookup fails, k0s issues a warning. If /etc/resolv.conf doesn't list any nameservers on top of that, worker nodes refuse to start. Controllers only warn in that case.

Kernel Parameters#

Inotify Instance Limit#

Inotify API allows applications to monitor filesystem changes.

Workloads that run many pods or watch large numbers of files can exhaust the node's default inotify limits, particularly fs.inotify.max_user_instances, which limits how many inotify instances one user (UID) can have.

A value of at least 1024 is recommended; 8192 is commonly used in Kubernetes environments:

fs.inotify.max_user_instances = 8192

  1. Check the current value:

    cat /proc/sys/fs/inotify/max_user_instances

  2. To apply the new value persistently, create a sysctl configuration file, for example, /etc/sysctl.d/99-inotify.conf, and place the parameter setting from above in it.

  3. Load the new kernel values:

    sudo sysctl --system

External hard dependencies#

There are very few external tools that are needed or used.

mount/umount#

When setting up pods, kubelet will call mount binary on the host. Similarly when destroying pods it will call umount. mount and umount are only needed on worker nodes where kubelet runs.

External soft dependencies#

There are a few external tools that may be needed or used under specific circumstances:

containerd and AppArmor#

In order to use containerd in conjunction with AppArmor, it must be enabled in the kernel and the /sbin/apparmor_parser executable must be installed on the host, otherwise containerd will disable AppArmor support.

iptables#

iptables may be executed to detect if there are any existing iptables rules and if those are in legacy or nft mode. If iptables is not found, k0s will assume that there are no pre-existing iptables rules.

useradd / adduser#

During k0s install the external tool useradd will be used on the controllers to create system user accounts for k0s. If this doesn't exist it will fall-back to busybox's adduser.

userdel / deluser#

k0s reset will execute either userdel or deluser to clean up system user accounts.

modprobe#

On k0s worker modprobe will be executed to load missing kernel modules if they are not detected.

id#

External id will be executed as a fallback if local user lookup fails, in case NSS is used.

Windows specific#

TBD.