"Pod disruption" and "eviction" are often used as if they mean the same thing. They don't. A disruption is any event that removes a pod. An eviction is one specific way a pod gets removed. Knowing which one you face decides whether a PodDisruptionBudget (PDB) can protect your application.
Short answer
- Pod disruption is the broad category: voluntary or involuntary events that terminate pods.
- Eviction is a mechanism. It is either API-initiated (respects PDBs) or node-pressure (ignores PDBs).
Types of pod disruption
Involuntary (unavoidable): hardware failure, a deleted VM, a kernel panic, a network partition that removes a node, or eviction because a node runs out of resources.
Voluntary: deleting a Deployment or pod, updating a pod template, draining a node for repair, upgrade or scale-down, or removing a pod to make room for another.
A PDB limits only voluntary disruptions, and only when the action uses the Eviction API. Deleting pods or Deployments directly bypasses PDBs. Involuntary disruptions cannot be prevented by a PDB, but they count against the budget.
API-initiated vs node-pressure eviction
| API-initiated eviction | Node-pressure eviction | |
|---|---|---|
| Triggered by | An Eviction API call, such as kubectl drain | The kubelet, when memory, disk, inodes or PIDs fall below thresholds |
| Disruption type | Voluntary | Involuntary |
| Respects PDB | Yes | No |
| Respects terminationGracePeriodSeconds | Yes | No (hard threshold: immediate; soft threshold: eviction-max-pod-grace-period) |
| Pod result | Gracefully terminated | Phase set to Failed |
How to tell which one happened
- Check node conditions. Run kubectl describe node <node-name>. If MemoryPressure, DiskPressure or PIDPressure is True, suspect node-pressure eviction.
- Inspect the pod. Run kubectl describe pod <pod-name> and read the status and message for the eviction reason.
- Check your PDBs. Run kubectl get pdb -A. If ALLOWED DISRUPTIONS is 0, API-initiated evictions for that app are blocked, so a drain will keep retrying.
- Review recent events. Run kubectl get events --sort-by=.lastTimestamp to match pod terminations to a drain, upgrade or resource event.
How to protect against each
Voluntary disruptions: create a PDB. Use either minAvailable or maxUnavailable, not both.
apiVersion: policy/v1kind: PodDisruptionBudgetmetadata: name: web-pdbspec: minAvailable: 2 selector: matchLabels: app: webWarning: a budget that allows no disruptions (for example, minAvailable equal to the replica count) can block node drains and cluster upgrades.
Node-pressure eviction: set accurate resource requests and limits, keep spare node capacity, and run multiple replicas spread across nodes or zones. Change kubelet eviction thresholds only after testing, because the defaults protect node stability.
Verification
- kubectl get pdb shows the expected ALLOWED DISRUPTIONS value.
- On a non-production node, kubectl drain <node-name> --ignore-daemonsets evicts pods one at a time without dropping below your budget. Run kubectl uncordon <node-name> afterwards. Draining is disruptive, so avoid doing this on production nodes during peak hours.