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
- Dealing with Disruptions
- How Disruption Budgets Work
- PDB Example
- Separating Cluster Owner and Application Owner Roles
- How to perform Disruptive Actions on your Cluster
- What's next
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.
Caution: Not all voluntary disruptions are constrained by Pod Disruption Budgets. For example, deleting deployments or pods bypasses Pod Disruption Budgets.
Dealing with Disruptions
Here are some ways to mitigate involuntary disruptions:
- Ensure your pod requests the resources it needs.
- Replicate your application if you need higher availability. (Learn about running replicated stateless and stateful applications.)
- For even higher availability when running replicated applications, spread applications across racks (using anti-affinity) or across zones (if using a multi-zone cluster.)
The frequency of voluntary disruptions varies. On a basic Kubernetes cluster, there are no voluntary disruptions at all. 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.
Kubernetes offers features to help run highly available applications at the same time as frequent voluntary disruptions. We call this set of features Disruption Budgets.
How Disruption Budgets Work
An Application Owner can create a
PodDisruptionBudget object (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 Pod Disruption Budgets by calling the Eviction API instead of directly deleting pods or deployments. Examples are the
kubectl drain command and the Kubernetes-on-GCE cluster upgrade script (
When a cluster administrator wants to drain a node they use the
kubectl drain command. That tool tries to evict all the pods on the machine. The eviction request may be temporarily rejected, and the tool periodically retries all failed requests until all pods 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 pods controller. The controller is discovered from the pods using the
.metadata.ownerReferences of the object.
PDBs cannot prevent involuntary disruptions from occurring, but 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 controllers (like deployment and stateful-set) are not limited by PDBs when doing rolling upgrades – the handling of failures during application updates is configured in the controller spec. (Learn about updating a deployment.)
When a pod is evicted using the eviction API, it is gracefully terminated (see
terminationGracePeriodSeconds in PodSpec.)
Consider a cluster with 3 nodes,
node-3. The cluster is running several applications. One of them has 3 replicas initially called
pod-c. Another, unrelated pod without a PDB, called
pod-x, is also shown. Initially, the pods are laid out as follows:
|pod-a available||pod-b available||pod-c 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-x. This succeeds immediately. Both pods go into the
terminating state at the same time. This puts the cluster in this state:
|pod-a terminating||pod-b available||pod-c available|
The deployment notices that one of the pods is terminating, so it creates a replacement called
node-1 is cordoned, it lands on another node. Something has also created
pod-y as a replacement for
(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:
|pod-a terminating||pod-b available||pod-c available|
|pod-x terminating||pod-d starting||pod-y|
At some point, the pods terminate, and the cluster looks like this:
|pod-b available||pod-c available|
At this point, if an impatient cluster administrator tries to drain
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:
|pod-b available||pod-c available|
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-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 drained||node-2||node-3||no node|
|pod-b available||pod-c available||pod-e pending|
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
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.