Pods evicted by node pressure (MemoryPressure, DiskPressure, PIDPressure)
When a node runs low on memory, disk or process IDs the kubelet evicts pods. Read the node conditions and set thresholds, requests and limits to match.
What you see
- Pods are terminated on a node and their phase is
Failed - The node reports
MemoryPressure,DiskPressureorPIDPressure
What it means
The kubelet watches resources such as memory.available, nodefs.available, imagefs.available and pid.available. When one crosses an eviction threshold it first tries to reclaim node resources, for example by removing unused container images when disk is short. If that is not enough it evicts pods.
During a node-pressure eviction the kubelet sets the phase of the chosen pods to Failed and terminates them. It does not respect a PodDisruptionBudget or the pod's terminationGracePeriodSeconds. For hard thresholds it uses a 0s grace period, so shutdown is immediate. Pods managed by a Deployment or StatefulSet are replaced by the control plane.
How to confirm the cause
- 1Check the node's conditions
kubectl describe node NODE_NAME
Look at
ConditionsforMemoryPressure,DiskPressureandPIDPressure. - 2See what is running on the node and what it requests
kubectl describe node NODE_NAME
The
Non-terminated PodsandAllocated resourcesblocks show who is using the node.
Causes and fixes
Memory pressure
- HOW TO TELL
MemoryPressureis True. The signal ismemory.available.- FIX
- Set memory requests and limits so pods fit, spread load across more nodes, or add memory. Pods using more than they requested are evicted before those within their requests.
Disk pressure
- HOW TO TELL
DiskPressureis True. Signals arenodefs.available,nodefs.inodesFree,imagefs.availableand similar.- FIX
- The kubelet first garbage collects dead pods and containers, then unused images. If that is not enough, find what is filling the node's filesystem, using the Linux disk-full runbook.
PID pressure
- HOW TO TELL
PIDPressureis True. The signal ispid.available.- FIX
- Find the workload creating too many processes and cap it.
How to stop it coming back
- Set requests and limits on every pod so eviction order is predictable.
- Alert on node conditions before the kubelet starts evicting.
- Give critical pods a high priority class.
Keep going
- Runbook: OOMKilled and exit code 137
- Runbook: Pod stuck in Pending with FailedScheduling
- Runbook: No space left on device
- Cheat sheet: kubectl
- Lesson: the kubernetes topic in Learn
Sources
Kubernetes documentation (kubernetes.io), read 18 Nov 2026. The docs version selector listed v1.33 to v1.37.