The idea
A Pod is a Kubernetes scheduling unit containing one or more containers with shared context. A Deployment manages a desired set of Pods through controllers. Pod instances can be replaced and their addresses can change. Desired replicas express a target, not a guarantee of healthy capacity under every condition.
Worked example
A fictional Deployment requests three replicas of a web service. If one Pod fails, a controller works toward the desired state. The replacement is not the same instance, so data kept only in the old container filesystem should not be treated as durable.
Try it
Draw a Deployment and three Pods, then replace one Pod. Label desired state versus observed state. Explain which identity or address may change and why three requested replicas do not prove the service can withstand every outage.