Fencing implementation for Kubernetes
Image | Build Status |
---|---|
kube-fencing-controller | |
kube-fencing-switcher | |
kube-fencing-agents |
This project designed to solve the problem of cleaning resources from the failed nodes that's blocks any further operation and recovery.
Fencing is neccesary if you want to have redundancy for your StatefulSet pods.
If any node falls, kube-fencing will guaranteed kill it via fence-agent, afterwards it will clear the node of all resources, that's make Kubernetes possible to schedule pods on the rest nodes.
Kube-fencing includes three containers:
The main controller which watches for the node states, and if one of them becomes to the NotReady
due NodeStatusUnknown
reason, runs fencing procedure.
This is small container which can be deployed as daemonset, it will enable fencing during start, and disable fencing when node is gracefully shutdowns or reboots.
This container contains installed fence-agents
package.
When fencing procedure is called fecning-controller creates Job which can use fecning-agents image to execute specific fencing agent. If fencing was successful it will celanup (or delete) the node from the kubernetes.
The next fencing agents are included:
fence_ack_manual fence_brocade fence_dummy fence_idrac fence_ilo4_ssh fence_ipmilan fence_ovh fence_rsb fence_vmware
fence_alom fence_cisco_mds fence_eaton_snmp fence_ifmib fence_ilo_moonshot fence_ironic fence_powerman fence_sanbox2 fence_vmware_soap
fence_amt fence_cisco_ucs fence_emerson fence_ilo fence_ilo_mp fence_kdump fence_pve fence_sbd fence_wti
fence_apc fence_compute fence_eps fence_ilo2 fence_ilo_ssh fence_ldom fence_raritan fence_scsi fence_xenapi
fence_apc_snmp fence_docker fence_hds_cb fence_ilo3 fence_imm fence_lpar fence_rcd_serial fence_tripplite_snmp fence_zvmip
fence_azure_arm fence_drac fence_hpblade fence_ilo3_ssh fence_intelmodular fence_mpath fence_rhevm fence_vbox
fence_bladecenter fence_drac5 fence_ibmblade fence_ilo4 fence_ipdu fence_netio fence_rsa fence_virsh
kubectl apply -f https://github.com/kvaps/kube-fencing/raw/master/deploy/kube-fencing.yaml
# Simple notify example (with after-hook)
kubectl apply -f https://github.com/kvaps/kube-fencing/raw/master/deploy/examples/after-hook.yaml
# HP iLO example
kubectl apply -f https://github.com/kvaps/kube-fencing/raw/master/deploy/examples/hp-ilo.yaml
Prepare your own fencing PodTemplate using the examples above.
Fencing-controller will spawn this PodTemplate every time when node going to unknown state.
It also appends fencing/node
and fencing/id
annotations to the pod, thus allows you to use this information in your fencing command.
The specified command must ends with 0
exit-code when fencing was successful and return 1
exit-code when failed.
You can create multiple PodTemplates for different nodes, but fencing
will be used by default.
All configuration is reduced to the specific annotations.
You can specify the needed annotations for specific node or commonly for PodTemplate, hovewer node annotations take precedence.
Annotation | Description | Default |
---|---|---|
fencing/enabled |
Fencing-switcher automatically sets this annotation to enable or disable fencing for the node. (can be specified only for node, usually you don't need to configure it). | false |
fencing/id |
Specify the device id which will be used to fence the node. | same as node name |
fencing/template |
Specify PodTemplate which be used to fence the node. | fencing |
fencing/mode |
Specify cleanup mode for the node:
|
flush |
fencing/after-hook |
Specific PodTemplate which will be spawned after successful fencing. | unspecified |
fencing/timeout |
Timeout in seconds to wait for the node recovery before starting fencing procedure. | 0 |
fencing/ttl |
TTLSecondsAfterFinished parameter for the created Jobs. |
100 |