QoS Classes and Pod Priority

Node pressure hits at 3 AM. Memory is exhausted. The kubelet starts evicting pods. Which pods survive, and which die first? Most engineers assume it comes down to priority, or which pod is using the most memory, or some manual annotation. The actual answer is more deterministic: it's governed by QoS class, a classification Kubernetes assigns at pod admission time based entirely on how you write your resource spec.

QoS class isn't a field you set. It's a derived property: a side effect of the relationship between resources.requests and resources.limits across every container in a pod. Get it wrong and your DNS server becomes the first casualty when a batch job exhausts node memory. Get it right and your critical infrastructure survives node pressure while transient workloads absorb the evictions.