Name	Name	Last commit message	Last commit date
parent directory ..
README.md	README.md
kep.yaml	kep.yaml

KEP-2268: Non graceful node shutdown

Release Signoff Checklist
Summary
Motivation
Proposal
- Handle the Return of Shutdown Node
- Risks and Mitigations
Design Details
Production Readiness Review Questionnaire
Implementation History
Alternatives
- Node fencing
- SafeDetach Option

Release Signoff Checklist

Items marked with (R) are required prior to targeting to a milestone / release.

Note: Any PRs to move a KEP to implementable or significant changes once it is marked implementable should be approved by each of the KEP approvers. If any of those approvers is no longer appropriate than changes to that list should be approved by the remaining approvers and/or the owning SIG (or SIG-arch for cross cutting KEPs).

Note: This checklist is iterative and should be reviewed and updated every time this enhancement is being considered for a milestone.

Summary

In this KEP, we are proposing a way that allows stateful workloads to failover to a different node successfully after the original node is shutdown or in a non-recoverable state such as the hardware failure or broken OS.

Motivation

The Graceful Node Shutdown KEP introduced a way to detect a node shutdown and handle it gracefully. However, a node shutdown action may not be detected by Kubelet's Node Shutdown Mananger, either because the command does not trigger the inhibitor locks mechanism used by Kubelet or because of a user error, i.e., the ShutdownGracePeriod and ShutdownGracePeriodCriticalPods are not configured properly.

When a node is shutdown but not detected by Kubelet's Node Shutdown Manager, the pods that are part of a StatefulSet will be stuck in terminating status on the shutdown node and cannot move to a new running node. This is because Kubelet on the shutdown node is not available to delete the pods so the StatefulSet cannot create a new pod with the same name. If there are volumes used by the pods, the VolumeAttachments will not be deleted from the original shutdown node so the volumes used by these pods cannot be attached to a new running node. As a result, the application running on the StatefulSet cannot function properly. If the original shutdown node comes up, the pods will be deleted by Kubelet and new pods will be created on a different running node. If the original shutdown node does not come up, these pods will be stuck in terminating status on the shutdown node forever.

In this KEP, we are proposing a way to handle node shutdown cases that are not detected by Kubelet's Node Shutdown Manager. The pods will be forcefully deleted in this case, trigger the deletion of the VolumeAttachments, and new pods will be created on a new running node so that application can continue to function.

Similarly, this approach can also be applied to the case when a node is in a non-recoverable state such as the hardware failure or broken OS.

Use Cases

If user wants to intentionally shutdown a node, he/she can validate whether graceful node shutdown feature works. If Kubelet is able to detect node is shutting down, it will gracefully delete pods, and new pods will be created on another running node.
If graceful shutdown is not working or node is in non-recoverable state due to hardware failure or broken OS, etc., user now can enable this feature, and add node.kubernetes.io/out-of-service=nodeshutdown:NoExecute taint which will be explained in detail below to trigger non-graceful shutdown behavior.

Goals

Increase the availability of stateful workloads when a node is shutdown, or in other non-recoverable cases such as the hardware failure or broken OS.
Requires user intervention to detect node shutdown/failure cases, because currently there is no good way of detecting whether node is in the middle of restarting.
- Once the feature in this KEP is developed and tested, we plan to work on automatic way of detecting and fencing node for shutdown/failure use cases and re-evaluating some approaches listed in the Alternatives section.

Non-Goals

Node/control plane partitioning other than a node shutdown is not covered by the proposal, but will be addressed in the future and built on top of this design.
- We do not have a way to determine whether it is safe to detach in the case of node partitioning. In the Alternatives section, we discussed the node fencing approach and explained why that is not selected for this KEP but will be considered after this KEP is implemented.
Implement in-cluster logic to handle node/control plane partitioning.

Proposal

In this section, we are describing the proposed changes to enable workloads to failover to a running node when the original node is shutdown but not detected by Kubelet's Node Shutdown Manager, or when the node is in other non-recoverable cases such as the hardware failure or broken OS.

Note: The proposal in the current KEP involves user manually add/remove a taint. In the future, we will look into a more automatic approach that does not require these manual steps.

When users verify that a node is already in shutdown or power off state (not in the middle of restarting), either user intentionally shut it down or node is down due to hardware failures, OS issues, they can taint the node with out-of-service with NoExecute effect.

Existing logic:

When a node is not reachable from the control plane, the health check in Node lifecycle controller, part of kube-controller-manager, sets Node v1.NodeReady Condition to False or Unknown (unreachable) if lease is not renewed for a specific grace period. Node Status becomes NotReady.
After 300 seconds (default), the Taint Manager tries to delete Pods on the Node after detecting that the Node is NotReady. The Pods will be stuck in terminating status.

Proposed logic change:

[Proposed change] This proposal requires a user to apply a out-of-service taint on a node when the user has confirmed that this node is shutdown or in a non-recoverable state due to the hardware failure or broken OS. Note that user should only add this taint if the node is not coming back at least for some time. If the node is in the middle of restarting, this taint should not be used.
[Proposed change] In the Pod GC Controller, part of the kube-controller-manager, add a new function called gcTerminating. This function would need to go through all the Pods in terminating state, verify that the node the pod scheduled on is NotReady. If so, do the following:
Upon seeing the out-of-service taint, the Pod GC Controller will forcefully delete the pods on the node if there are no matching tolation on the pods. This new out-of-service taint has NoExecute effect, meaning the pod will be evicted and a new pod will not schedule on the shutdown node unless it has a matching toleration. For example, node.kubernetes.io/out-of-service=out-of-service=nodeshutdown:NoExecute or node.kubernetes.io/out-of-service=out-of-service=hardwarefailure:NoExecute. We suggest to use NoExecute effect in taint to make sure pods will be evicted (deleted) and fail over to other nodes.
We'll follow taint and toleration policy. If a pod is set to tolerate all taints and effects, that means user does NOT want to evict pods when node is not ready. So GC controller will filter out those pods and only forcefully delete pods that do not have a matching toleration. If your pod tolerates the out-of-service taint, then it will not be terminated by the taint logic, therefore none of this applies.
[Proposed change] Once pods are selected and forcefully deleted, the attachdetach reconciler should check the out-of-service taint on the node. If the taint is present, the attachdetach reconciler will not wait for 6 minutes to do force detach. Instead it will force detach right away and allow volumeAttachment to be deleted.
This would trigger the deletion of the volumeAttachment objects. For CSI drivers, this would allow ControllerUnpublishVolume to happen without NodeUnpublishVolume and/or NodeUnstageVolume being called first. Note that there is no additional code changes required for this step. This happens automatically after the Proposed change in the previous step to force detach right away.
When the external-attacher detects the volumeAttachment object is being deleted, it calls CSI driver's ControllerUnpublishVolume.

Note: [Existing] If the out-of-service:NoExecute taint is applied to a node that is healthy and running, the Taint Manager will try to delete the pods without the matching toleration. Pods will be deleted by Kubelet because the node is healthy and running. New pods will be created on a different running node and application will continue to run.

Handle the Return of Shutdown Node

If the nodes that were previously shutdown with pods forcefully deleted come back, we need to make sure they are cleaned up before pods are scheduled on them again.

[Proposed Change] We require the user to manually remove the out-of-service taint after the pods are moved to a new node and the user has checked that the shutdown node has been recovered since the user was the one who originally added the taint. This makes sure no new pods are scheduled on the node before the out-of-service taint is removed. In the future, we can enhance the Pod GC Controller to automatically remove the out-of-service taint, after verifying that no pods are attached to the node, and no volume attachments on the node.

Note that there is no additional code change required for the return of shutdown node.

[Existing Logic] The health check in Node lifecycle controller, part of kube-controller-manager, sets Node v1.NodeReady Condition to True if lease is renewed again. Node Status becomes Ready.

[Existing Logic] Note that if the out-of-service taint was not added to the shutdown node, the pods will be stuck in terminating status forever if the shutdown node remains down. If the shutdown node comes up, however, the original volumeAttachment will be deleted and the pods will be moved to a new node successfully.

Risks and Mitigations

This KEP introduces changes, and if not tested correctly could result in data corruption for users. To mitigate this we plan to have a high test coverage and to introduce this enhancement as an alpha

Design Details

Test Plan

[X] I/we understand the owners of the involved components may require updates to existing tests to make this code solid enough prior to committing the changes necessary to implement this enhancement.

Prerequisite testing updates

There are existing tests for Pod GC controller: https://github.com/kubernetes/kubernetes/blob/master/test/e2e/node/pod_gc.go

There are existing tests for attach detach controller. Creating a pod that uses PVCs which will test attach: https://github.com/kubernetes/kubernetes/blob/master/test/e2e/framework/pod/create.go Deleting a pod will trigger PVC to be detached: https://github.com/kubernetes/kubernetes/blob/master/test/e2e/framework/pod/delete.go

Unit tests

Add unit tests to affected components in kube-controller-manager:
- Add tests in Pod GC Controller for the new logic to clean up pods and the out-of-service taint.
- Add tests in Attachdetach Controller for the changed logic that allow volumes to be forcefully detached without wait.

Integration tests

After reviewing the tests, we decided that the best place to add a test for this feature is under test/e2e/storage: https://github.com/kubernetes/kubernetes/blob/master/test/e2e/storage/non_graceful_node_shutdown.go

E2E tests

Added E2E tests to validate workloads move successfully to another running node when a node is shutdown: https://github.com/kubernetes/kubernetes/blob/master/test/e2e/storage/non_graceful_node_shutdown.go
- Feature gate for NodeOutOfServiceVolumeDetach is enabled. Add out-of-service taint after node is shutdown:
  - Verify workloads are moved to another node successfully.
  - Verify the out-of-service taint is removed after the shutdown node is cleaned up.
Add stress and scale tests before moving from beta to GA.

We also plan to test this with different version Skews.

Graduation Criteria

This KEP will be treated as a new feature, and will be introduced with a new feature gate, NodeOutOfServiceVolumeDetach.

This enhancement will go through the following maturity levels: alpha, beta and stable.

Alpha

Initial feature implementation in kube-controller-manager.
Add basic unit tests.

Alpha -> Beta Graduation

Gather feedback from developers and surveys.
Unit tests and e2e tests outlined in design proposal implemented.
Tests are in Testgrid and linked in KEP

Beta -> GA Graduation

Feature is deployed in production and have gone through at least one Kubernetes upgrade.
More rigorous forms of testing—e.g., downgrade tests and scalability tests.
Allowing time for feedback.

Upgrade / Downgrade Strategy

once enabled by default, the upgrades should be smooth as the default behaviour is off by default. when downgrading/rollbacking with this behaviour enabled, there should be no issues as the control plane with the logic is still there.

Production Readiness Review Questionnaire

Feature Enablement and Rollback

How can this feature be enabled / disabled in a live cluster?

Feature gate (also fill in values in kep.yaml)
- Feature gate name:
  - NodeOutOfServiceVolumeDetach
- Components depending on the feature gate:
  - kube-controller-manager
Other
- Describe the mechanism:
- Will enabling / disabling the feature require downtime of the control plane?
  - Yes.
- Will enabling / disabling the feature require downtime or reprovisioning of a node?
  - No.

Does enabling the feature change any default behavior?

Yes. If this feature is enabled, the pod gc controller in kube-controller-manager would need to ensure that out-of-service taint is applied on a node before forcefully deleting pods.

Can the feature be disabled once it has been enabled (i.e. can we roll back the enablement)?

Yes. If this feature is disable once it has been enabled, it falls back to default behavior before the feature is enabled, meaning the pod gc controller will not check the out-of-service taint and make decision depending on that.

What happens if we reenable the feature if it was previously rolled back?

If we reenable the feature if it was previously rolled back, the pod gc controller will again check the out-of-service taint before forcefully deleting pods.

Are there any tests for feature enablement/disablement?

We will add unit test for feature enablement/disablement.

Rollout, Upgrade and Rollback Planning

How can a rollout or rollback fail? Can it impact already running workloads?

The rollout should not fail. Feature gate only needs to be enabled on kube-controller-manager. So it is either enabled or disabled. In an HA cluster, assume a 1.N kube-controller-manager has feature gate enabled and takes the lease first, and the GC Controller has already forcefully deleted the pods. Then it loses it to a 1.N-1 kube-controller-manager which as feature gate disabled. In this case, the Attach Detach Controller won't have the new behavior so it will wait for 6 minutes before force detach the volume. If the 1.N kube-controller-manager has feature gate disabled while the 1.N-1 kube-controller-manager has feature gate enabled, the GC Controller will not forcefully delete the pods. So the Attach Detach Controller will not be triggered to force detach. In the later case, it will still keep the old behavior.

What specific metrics should inform a rollback?

If for some reason, the user does not want the workload to failover to a different running node after the original node is shutdown and the out-of-service taint is applied, a rollback can be done. I don't see why it is needed though as user can prevent the failover from happening by not applying the out-of-service taint. Since use of this feature requires applying a taint manually by the user, it should not specifically require rollback.

Were upgrade and rollback tested? Was the upgrade->downgrade->upgrade path tested?

We will manually test upgrade from 1.25 to 1.26 and rollback from 1.26 to 1.25.

Is the rollout accompanied by any deprecations and/or removals of features, APIs, fields of API types, flags, etc.?

No.

Monitoring Requirements

How can an operator determine if the feature is in use by workloads?

An operator, the person who operates the cluster, can check if the NodeOutOfServiceVolumeDetach feature gate is enabled and if there is an out-of-service taint on the shutdown node. The usage of this feature requires the manual step of applying a taint so the operator should be the one applying it.

How can someone using this feature know that it is working for their instance?

API .status If it works, pods in the stateful workload should be re-scheduled to another running node. Phase in Pod Status should be Running for a new Pod on the other running node. If not, check the pod status to see why it does not come up.

What are the SLIs (Service Level Indicators) an operator can use to determine the health of the service?

Metrics
- Metric name:
  - New metrics are added in Pod GC Controller:
    - force_delete_pods_total{reason="out-of-service|terminated|orphaned|unscheduled"}, the number of pods that are being forcefully deleted since the Pod GC Controller started.
    - force_delete_pod_errors_total{reason="out-of-service|terminated|orphaned|unscheduled"}, the number of errors encountered when forcefully deleting the pods since the Pod GC Controller started.
  - For Attach Detach Controller, the following metric will be recorded if a force detach is performed because the node has the out-of-service taint or a timeout happens:
    - attachdetach_controller_forced_detaches{reason="out-of-service|timeout"}, the number of times the Attach Detach Controller performed a forced detach.
  - There is also a kube_node_spec_taint in kube-state-metrics that is a metric for the taint of a Kubernetes cluster node.
- [Optional] Aggregation method:
- Components exposing the metric: kube-controller-manager
Other (treat as last resort)
- Details:
  - Check whether the workload moved to a different running node after the original node is shutdown and the out-of-service taint is applied on the shutdown node.

What are the reasonable SLOs (Service Level Objectives) for the enhancement?

At a high level, this usually will be in the form of "high percentile of SLI per day <= X". It's impossible to provide comprehensive guidance, but at the very high level (needs more precise definitions) those may be things like:

per-day percentage of API calls finishing with 5XX errors <= 1%
99% percentile over day of absolute value from (job creation time minus expected job creation time) for cron job <= 10%
99,9% of /health requests per day finish with 200 code The failover should always happen if the feature gate is enabled, the taint is applied, and there are other running nodes. We can also check the deleting_pods_total, deleting_pods_error_total metrics in Pod GC Controller and the attachdetach_controller_forced_detaches and attachdetach_controller_forced_detaches_taint metric in the Attach Detach Controller.

Are there any missing metrics that would be useful to have to improve observability of this feature?

Dependencies

Does this feature depend on any specific services running in the cluster?

This feature relies on the kube-controller-manager being running. If the workload is running on a StatefulSet, it also depends on the CSI driver.

For each of these, fill in the following—thinking about running existing user workloads and creating new ones, as well as about cluster-level services (e.g. DNS):

[Dependency name] CSI driver
- Usage description:
  - Impact of its outage on the feature: If the CSI driver is not running, the pod cannot use the persistent volume any more so the workload will not be running properly.
  - Impact of its degraded performance or high-error rates on the feature: Workload does not work properly if the CSI driver is down.

Scalability

Without this feature, a user can forcefully delete the pods after they are in terminating state and new pods will be re-scheduled to another running node after 6 minutes. With this feature, new pods will be re-scheduled to another running node without the 6 minute wait after the user has applied the out-of-service taint. It speeds up the failover but should not affect the scalability.

Will enabling / using this feature result in any new API calls?

Nothing new.

Will enabling / using this feature result in introducing new API types?

No.

Will enabling / using this feature result in any new calls to the cloud provider?

Volume detach/attach could trigger cloud provider calls.

Will enabling / using this feature result in increasing size or count of the existing API objects?

No.

Will enabling / using this feature result in increasing time taken by any operations covered by existing SLIs/SLOs?

No.

Will enabling / using this feature result in non-negligible increase of resource usage (CPU, RAM, disk, IO, ...) in any components?

No.

Can enabling / using this feature result in resource exhaustion of some node resources (PIDs, sockets, inodes, etc.)?

No.

Troubleshooting

How does this feature react if the API server and/or etcd is unavailable?

If API server or etcd is not available, we can't get accurate status of node or pod. However the usage of this feature is very manual so an operator can verify before applying the taint.

What are other known failure modes?

For each of them, fill in the following information by copying the below template:

[Failure mode brief description]
- Detection: How can it be detected via metrics? Stated another way: how can an operator troubleshoot without logging into a master or worker node?
  
  After applying the out-of-service taint, if the workload does not move to a different running node, that is an indicator something might be wrong. Note that since we removed the 6 minute wait in the Attach Detach controller, force detach will happen if the feature gate is enabled and the taint is applied without delay.
- Mitigations: What can be done to stop the bleeding, especially for already running user workloads?
  
  So if the workload does not failover, it behaves the same as when this feature is not enabled. The operator should try to find out why the failover didn't happen. Check messages from GC Controller and Attach Detach Controller in the kube-controller-manager logs. Check if there are following messages from GC Controller: "failed to get node : " "Garbage collecting pods that are terminating on node tainted with node.kubernetes.io/out-of-service". Check if there is the following message from Attach Detach Controller: "failed to get taint specs for node : "
- Diagnostics: What are the useful log messages and their required logging levels that could help debug the issue? Not required until feature graduated to beta.
  
  Set log level to at least 2. For example, the following message is in GC Controller if the feature is enabled and out-of-service taint is applied. If the pods are forcefully deleted by the GC Controller, this message should show up. klog.V(2).Infof("garbage collecting pod %s that is terminating. Phase [%v]", pod.Name, pod.Status.Phase) There is also a message in Attach Detach Controller that checks the taint. If the taint is applied and feature gate is enabled, it force detaches the volume without waiting for 6 minutes. klog.V(2).Infof("node %q has out-of-service taint", attachedVolume.NodeName)
- Testing: Are there any tests for failure mode? If not, describe why.
  
  We have unit tests that cover different combination of pod and node statuses.

What steps should be taken if SLOs are not being met to determine the problem?

In that case, we need to go through the logs and find out the root cause. Check messages from GC Controller and Attach Detach Controller in the kube-controller-manager logs as described in the above section.

Implementation History

2019-06-26: Initial KEP published.
2020-11-10: KEP updated to handle part of the node partitioning
2021-08-26: The scope of the KEP is narrowed down to handle a real node shutdown. Test plan is updated. Node partitioning will be handled in the future and it can be built on top of this design.
2021-12-03: Removed SafeDetach flag. Requires a user to add the out-of-service taint when he/she knows the node is shutdown.
Kubernetes v1.24: moved to alpha.
Kubernete v1.26: moved to beta.
Kubernete v1.28: moved to stable.

Alternatives

Node fencing

Node fencing was brought up as an alternative during KEP reviews. This approach depends on the user to enter what commands are required to shut down a node as the commands differ in different environment.

- kind: ConfigMap
  apiVersion: v1
  metadata:
   name: fence-method-fence-rhevm-node1
   namespace: default
  data:
   method.properties: |
          agent_name=fence-rhevm
          namespace=default
          ip=ovirt.com  # address to the rhevm management
          username=admin@internal
          password-script=/usr/sbin/fetch_passwd
          ssl-insecure=true
          plug=vm-node1  # the vm name
          action=reboot
          ssl=true
          disable-http-filter=true

This approach seems too intrusive as a general in-tree solution. Since the scope of this KEP is narrowed down to handle a real node shutdown scenario, we do not need the node fencing approach. I think what has been proposed in this KEP also does not prevent us from supporting the node partitioning case in the future. We can reconsider this approach after the proposed solution is implemented.

SafeDetach Option

This proposal introduces a new option in CSIDriver. If set to true, it would trigger volume detach if a node is shutdown.

In the CSIDriverSpec, add a SafeDetach boolean to indicate if it is safe to detach.

// CSIDriverSpec is the specification of a CSIDriver.
type CSIDriverSpec struct {
  ...

  // SafeDetach indicates this CSI volume driver is able to perform detach
  // operations. When set to true, the CSI volume driver needs to ensure that
  // ControllerUnpublishVolume() implementation is checking whether a volume
  // can be detached safely
  // +optional
  SafeDetach *bool
}

Since we have modified the KEP and require the node to be shutdown, we no longer need this SafeDetach flag. Currently no CSI driver can truly detect whether it is safe to detach itself. We can re-visit this option later when handling network partitioning case.

Files

2268-non-graceful-shutdown

Directory actions

More options