Disruptions

This guide is for application owners who want to build highly available applications, and thus need to understand what types of disruptions can happen to Pods.

It is also for cluster administrators who want to perform automated cluster actions, like upgrading and autoscaling clusters.

Voluntary and involuntary disruptions

Pods do not disappear until someone (a person or a controller) destroys them, or there is an unavoidable hardware or system software error.

We call these unavoidable cases involuntary disruptions to an application. Examples are:

  • a hardware failure of the physical machine backing the node
  • cluster administrator deletes VM (instance) by mistake
  • cloud provider or hypervisor failure makes VM disappear
  • a kernel panic
  • the node disappears from the cluster due to cluster network partition
  • eviction of a pod due to the node being out-of-resources.

Except for the out-of-resources condition, all these conditions should be familiar to most users; they are not specific to Kubernetes.

We call other cases voluntary disruptions. These include both actions initiated by the application owner and those initiated by a Cluster Administrator. Typical application owner actions include:

  • deleting the deployment or other controller that manages the pod
  • updating a deployment's pod template causing a restart
  • directly deleting a pod (e.g. by accident)

Cluster administrator actions include:

  • Draining a node for repair or upgrade.
  • Draining a node from a cluster to scale the cluster down (learn about Cluster Autoscaling ).
  • Removing a pod from a node to permit something else to fit on that node.

These actions might be taken directly by the cluster administrator, or by automation run by the cluster administrator, or by your cluster hosting provider.

Ask your cluster administrator or consult your cloud provider or distribution documentation to determine if any sources of voluntary disruptions are enabled for your cluster. If none are enabled, you can skip creating Pod Disruption Budgets.

Dealing with disruptions

Here are some ways to mitigate involuntary disruptions:

The frequency of voluntary disruptions varies. On a basic Kubernetes cluster, there are no automated voluntary disruptions (only user-triggered ones). However, your cluster administrator or hosting provider may run some additional services which cause voluntary disruptions. For example, rolling out node software updates can cause voluntary disruptions. Also, some implementations of cluster (node) autoscaling may cause voluntary disruptions to defragment and compact nodes. Your cluster administrator or hosting provider should have documented what level of voluntary disruptions, if any, to expect. Certain configuration options, such as using PriorityClasses in your pod spec can also cause voluntary (and involuntary) disruptions.

Pod disruption budgets

FEATURE STATE: Kubernetes v1.21 [stable]

Kubernetes offers features to help you run highly available applications even when you introduce frequent voluntary disruptions.

As an application owner, you can create a PodDisruptionBudget (PDB) for each application. A PDB limits the number of Pods of a replicated application that are down simultaneously from voluntary disruptions. For example, a quorum-based application would like to ensure that the number of replicas running is never brought below the number needed for a quorum. A web front end might want to ensure that the number of replicas serving load never falls below a certain percentage of the total.

Cluster managers and hosting providers should use tools which respect PodDisruptionBudgets by calling the Eviction API instead of directly deleting pods or deployments.

For example, the kubectl drain subcommand lets you mark a node as going out of service. When you run kubectl drain, the tool tries to evict all of the Pods on the Node you're taking out of service. The eviction request that kubectl submits on your behalf may be temporarily rejected, so the tool periodically retries all failed requests until all Pods on the target node are terminated, or until a configurable timeout is reached.

A PDB specifies the number of replicas that an application can tolerate having, relative to how many it is intended to have. For example, a Deployment which has a .spec.replicas: 5 is supposed to have 5 pods at any given time. If its PDB allows for there to be 4 at a time, then the Eviction API will allow voluntary disruption of one (but not two) pods at a time.

The group of pods that comprise the application is specified using a label selector, the same as the one used by the application's controller (deployment, stateful-set, etc).

The "intended" number of pods is computed from the .spec.replicas of the workload resource that is managing those pods. The control plane discovers the owning workload resource by examining the .metadata.ownerReferences of the Pod.

Involuntary disruptions cannot be prevented by PDBs; however they do count against the budget.

Pods which are deleted or unavailable due to a rolling upgrade to an application do count against the disruption budget, but workload resources (such as Deployment and StatefulSet) are not limited by PDBs when doing rolling upgrades. Instead, the handling of failures during application updates is configured in the spec for the specific workload resource.

When a pod is evicted using the eviction API, it is gracefully terminated, honoring the terminationGracePeriodSeconds setting in its PodSpec.

PodDisruptionBudget example

Consider a cluster with 3 nodes, node-1 through node-3. The cluster is running several applications. One of them has 3 replicas initially called pod-a, pod-b, and pod-c. Another, unrelated pod without a PDB, called pod-x, is also shown. Initially, the pods are laid out as follows:

node-1node-2node-3
pod-a availablepod-b availablepod-c available
pod-x available

All 3 pods are part of a deployment, and they collectively have a PDB which requires there be at least 2 of the 3 pods to be available at all times.

For example, assume the cluster administrator wants to reboot into a new kernel version to fix a bug in the kernel. The cluster administrator first tries to drain node-1 using the kubectl drain command. That tool tries to evict pod-a and pod-x. This succeeds immediately. Both pods go into the terminating state at the same time. This puts the cluster in this state:

node-1 drainingnode-2node-3
pod-a terminatingpod-b availablepod-c available
pod-x terminating

The deployment notices that one of the pods is terminating, so it creates a replacement called pod-d. Since node-1 is cordoned, it lands on another node. Something has also created pod-y as a replacement for pod-x.

(Note: for a StatefulSet, pod-a, which would be called something like pod-0, would need to terminate completely before its replacement, which is also called pod-0 but has a different UID, could be created. Otherwise, the example applies to a StatefulSet as well.)

Now the cluster is in this state:

node-1 drainingnode-2node-3
pod-a terminatingpod-b availablepod-c available
pod-x terminatingpod-d startingpod-y

At some point, the pods terminate, and the cluster looks like this:

node-1 drainednode-2node-3
pod-b availablepod-c available
pod-d startingpod-y

At this point, if an impatient cluster administrator tries to drain node-2 or node-3, the drain command will block, because there are only 2 available pods for the deployment, and its PDB requires at least 2. After some time passes, pod-d becomes available.

The cluster state now looks like this:

node-1 drainednode-2node-3
pod-b availablepod-c available
pod-d availablepod-y

Now, the cluster administrator tries to drain node-2. The drain command will try to evict the two pods in some order, say pod-b first and then pod-d. It will succeed at evicting pod-b. But, when it tries to evict pod-d, it will be refused because that would leave only one pod available for the deployment.

The deployment creates a replacement for pod-b called pod-e. Because there are not enough resources in the cluster to schedule pod-e the drain will again block. The cluster may end up in this state:

node-1 drainednode-2node-3no node
pod-b terminatingpod-c availablepod-e pending
pod-d availablepod-y

At this point, the cluster administrator needs to add a node back to the cluster to proceed with the upgrade.

You can see how Kubernetes varies the rate at which disruptions can happen, according to:

  • how many replicas an application needs
  • how long it takes to gracefully shutdown an instance
  • how long it takes a new instance to start up
  • the type of controller
  • the cluster's resource capacity

Pod disruption conditions

FEATURE STATE: Kubernetes v1.26 [beta]

When enabled, a dedicated Pod DisruptionTarget condition is added to indicate that the Pod is about to be deleted due to a disruption. The reason field of the condition additionally indicates one of the following reasons for the Pod termination:

PreemptionByKubeScheduler
Pod is due to be preempted by a scheduler in order to accommodate a new Pod with a higher priority. For more information, see Pod priority preemption.
DeletionByTaintManager
Pod is due to be deleted by Taint Manager (which is part of the node lifecycle controller within kube-controller-manager) due to a NoExecute taint that the Pod does not tolerate; see taint-based evictions.
EvictionByEvictionAPI
Pod has been marked for eviction using the Kubernetes API .
DeletionByPodGC
Pod, that is bound to a no longer existing Node, is due to be deleted by Pod garbage collection.
TerminationByKubelet
Pod has been terminated by the kubelet, because of either node pressure eviction or the graceful node shutdown.

When the PodDisruptionConditions feature gate is enabled, along with cleaning up the pods, the Pod garbage collector (PodGC) will also mark them as failed if they are in a non-terminal phase (see also Pod garbage collection).

When using a Job (or CronJob), you may want to use these Pod disruption conditions as part of your Job's Pod failure policy.

Separating Cluster Owner and Application Owner Roles

Often, it is useful to think of the Cluster Manager and Application Owner as separate roles with limited knowledge of each other. This separation of responsibilities may make sense in these scenarios:

  • when there are many application teams sharing a Kubernetes cluster, and there is natural specialization of roles
  • when third-party tools or services are used to automate cluster management

Pod Disruption Budgets support this separation of roles by providing an interface between the roles.

If you do not have such a separation of responsibilities in your organization, you may not need to use Pod Disruption Budgets.

How to perform Disruptive Actions on your Cluster

If you are a Cluster Administrator, and you need to perform a disruptive action on all the nodes in your cluster, such as a node or system software upgrade, here are some options:

  • Accept downtime during the upgrade.
  • Failover to another complete replica cluster.
    • No downtime, but may be costly both for the duplicated nodes and for human effort to orchestrate the switchover.
  • Write disruption tolerant applications and use PDBs.
    • No downtime.
    • Minimal resource duplication.
    • Allows more automation of cluster administration.
    • Writing disruption-tolerant applications is tricky, but the work to tolerate voluntary disruptions largely overlaps with work to support autoscaling and tolerating involuntary disruptions.

What's next

Last modified February 22, 2023 at 9:09 AM PST: 更新编辑 (f4a7975)