What tow changes on your own machines, and how it's undone

On machines you list yourself (provider: byoh), tow records every change it makes in /etc/tow/prepared.json on each machine, readable by root only. Each entry says which cluster made the change, what it did, how to undo it (undo, for you), and the exact script tow runs to undo it (revert):

sudo cat /etc/tow/prepared.json

A machine’s packages or settings that were there before tow are never recorded, so they’re never undone.

What tow delete does

CommandKubernetesThe preparation
tow delete --name labremovedstays; in a terminal, delete asks first (default: no)
tow delete --name lab --resetremovedundone
tow delete --name lab --yesremovedstays: undoing it unattended takes --reset
tow delete --name lab --forgetstaysstays: nothing is touched
  • Before anything changes, delete reads each machine’s record and lists, machine by machine, what goes and what stays.
  • A machine it can’t reach stops it before anything changes. --forget is for a machine that’s gone for good.
  • When an undo fails, the cluster is kept. Run delete again: what was undone left the record and isn’t done twice.
  • Volumes on the machines’ disks (local-path, OpenEBS, Longhorn) hold your data, so delete refuses a cluster that has some, and lists them. delete -v deletes them first, through Kubernetes, so each storage driver deletes its own data.

What’s recorded, and how each is undone

Kubernetes (removed by every delete but --forget):

ChangeUndone by
k3sk3s’s own uninstaller: k3s-uninstall.sh on control planes, k3s-agent-uninstall.sh on workers. It also removes k3s’s network interfaces and rules, and its data, /var/lib/rancher/k3s.
kubeadm’s init or joinkubeadm reset -f, then /etc/cni/net.d (kubeadm leaves it, its docs say)
kubelet, kubeadm, kubectlpurged with cri-tools and kubernetes-cni, then their apt repository and key (on RHEL-family machines: dnf remove, then /etc/yum.repos.d/kubernetes.repo)
containerd and runcstopped, binaries and service removed, /var/lib/containerd deleted; a config.toml that was there before tow comes back
CRI-Ostopped, purged, its repository and settings removed

The preparation (undone by --reset):

ChangeUndone by
packages (conntrack, open-iscsi…)apt-get remove (on RHEL-family machines, dnf remove) of those packages only: no purge, no autoremove
kernel modulestow’s lines removed from /etc/modules-load.d/k8s.conf; the file goes when it’s empty
kernel settingstow’s lines removed from /etc/sysctl.d/k8s.conf; the file goes when it’s empty
swapthe lines tow commented out of /etc/fstab (# tow: ) come back, swap units are unmasked, dphys-swapfile is enabled again, swapon -a
SELinux (kubeadm, RHEL family)enforcing again, now (setenforce 1) and in /etc/selinux/config, as far as it enforced before
ufw rulesthe rules commented tow:<cluster> are deleted. ufw stays on or off as it was: tow never turns it on.
firewalldits zone tow-nodes and policy tow-admin are deleted (with firewall-offline-cmd when firewalld isn’t running). firewalld stays on or off as it was.
NetworkManager (Calico)/etc/NetworkManager/conf.d/calico.conf removed, NetworkManager reloaded
multipathd (Longhorn)/etc/multipath/conf.d/longhorn.conf removed, multipathd restarted

Each undone change leaves the record; once nothing is left in it, the file and /etc/tow go too.

What stays until the machine restarts

  • Kernel modules and kernel settings: they stay loaded and set; the next boot no longer applies them.
  • On kubeadm, kube-proxy’s iptables rules and the network’s interfaces (cni0, flannel.1, vxlan.calico…). kubeadm’s docs say the rules are harmless left as they are. k3s’s uninstaller removes its own.
  • kube-vip’s virtual IP, if kube-vip was stopped before it could let the address go.

Doing it by hand

For a machine tow can no longer reach, or after --forget, use the undo of each entry in the machine’s record, newest first, Kubernetes first. As root:

# k3s
/usr/local/bin/k3s-uninstall.sh          # control plane
/usr/local/bin/k3s-agent-uninstall.sh    # worker

# kubeadm
kubeadm reset -f && rm -rf /etc/cni/net.d /etc/tow/kubeadm-done
apt-mark unhold kubelet kubeadm kubectl
apt-get purge kubelet kubeadm kubectl cri-tools kubernetes-cni
rm /etc/apt/sources.list.d/kubernetes.list /etc/apt/keyrings/kubernetes-apt-keyring.gpg
# (RHEL family instead: dnf remove kubelet kubeadm kubectl cri-tools kubernetes-cni && rm /etc/yum.repos.d/kubernetes.repo)
systemctl disable --now containerd
rm -f /usr/local/bin/containerd* /usr/local/bin/ctr /usr/local/sbin/runc /usr/local/lib/systemd/system/containerd.service
rm -rf /etc/containerd /var/lib/containerd && systemctl daemon-reload

# the preparation: see each entry's "undo"; for the firewall
ufw status numbered      # then 'ufw delete <number>' for each rule marked tow:<cluster>, highest number first
firewall-cmd --permanent --delete-zone=tow-nodes && firewall-cmd --permanent --delete-policy=tow-admin && firewall-cmd --reload

# last
rm -rf /etc/tow

Volume data folders are left for you to delete once you’re sure you don’t need them: /opt/local-path-provisioner (local-path on kubeadm), /var/openebs/local (OpenEBS), /var/lib/longhorn (Longhorn).

Last modified November 16, 2026: tow: first public version (5b35cf3)