Resource contention occurs when workloads compete for finite resources such as memory or CPU on shared Kubernetes infrastructure. Kubernetes protects the node by applying the requests, limits, and eviction rules declared for Pods. section 5.1.3 section 5.1.1
During contention, Kubernetes responds according to the resource and workload configuration:
- The scheduler places Pods using resource requests, while limits constrain usage after scheduling. section 5.1.1
- A Pod exceeding its memory limit is restarted, while one exceeding its CPU limit is throttled. section 5.1.1
- Pods using more memory than requested may be evicted; managed workloads such as Deployments can reschedule evicted Pods. section 5.1.3
- Best-effort Pods are most vulnerable, while Guaranteed Pods with equal requests and limits are protected from eviction. section 5.1.3 section 5.1.2
- Frequent evictions indicate that requests may be too low or that the cluster needs additional capacity. section 5.1.3
A protected critical workload
A Deployment can request 200m CPU and 250Mi memory, set higher limits, and assign a high-priority class. Setting requests equal to limits instead makes the Pod Guaranteed; enabling preemption can allow the priority class to remove lower-priority Pods when capacity is unavailable.
