☸️ Your Pods Are Healthy One Minute, Evicted the Next – No Crash, No Error
Pods running perfectly fine, passing every health check, suddenly show as Evicted with no application error, no crash log, nothing in the pod’s own logs explaining why – because the decision to remove them was made entirely at the NODE level, for reasons the pod itself never got a chance to see coming.
🐞 The Problem
$ kubectl get pods NAME READY STATUS RESTARTS checkout-api-7f9d8b 0/1 Evicted 0 $ kubectl describe pod checkout-api-7f9d8b Status: Failed Reason: Evicted Message: The node was low on resource: memory.
🔍 Why This Happens
Kubernetes' kubelet actively monitors each node's available memory (and other resources), and when a node comes under genuine memory pressure, it evicts pods to protect the NODE itself from becoming completely unresponsive - starting with pods using more memory than their configured 'requests' (an easy target for eviction since they're already exceeding their reserved amount) or pods with no resource requests/limits configured at all, which the scheduler had no accurate accounting for in the first place.
✅ The Fix: Set Realistic Resource Requests and Limits on Every Pod
- Set a memory REQUEST that accurately reflects what the pod actually needs under normal load – the scheduler uses this to decide which node has room, and an unrealistically low request lets too many pods land on the same node, setting up exactly this kind of pressure.
- Set a memory LIMIT as well, so a pod with a genuine memory leak or unexpected spike gets OOMKilled on its own (a clear, attributable failure) rather than silently contributing to node-wide pressure that evicts OTHER, innocent pods alongside it.
- Monitor actual memory usage against configured requests/limits over time (via metrics-server, Prometheus, or your cloud provider’s monitoring) – both to catch pods that are chronically under-provisioned and to right-size requests instead of guessing.
⚠️ Why ‘No Error in the Pod’s Own Logs’ Makes Sense Here
- Eviction is a decision made by the kubelet ABOUT the pod, from the outside, based on node-wide conditions the pod’s own application code has no visibility into whatsoever – there’s no error to log from inside the pod because, from the application’s perspective, nothing about ITS own execution ever went wrong.
- Check
kubectl describe node <node-name>for the node’s own conditions and recent events – this is where the actual pressure signal and eviction decision are recorded, not in any individual pod’s logs.
An evicted pod was never accused of doing anything wrong — it was simply the one the node picked to protect itself, and the real fix is giving the scheduler accurate numbers so it never has to make that choice under pressure.

