Declare the state, then watch the loop fight you.

Kubernetes is taught as a field reference, and the fields are not the hard part. The hard part is a control loop nobody watches: what the ReplicaSet controller does when readiness never goes true, what the scheduler does when your requests do not fit, and when exactly traffic stops arriving at a pod that is shutting down.

strategy:
  type: RollingUpdate
  rollingUpdate:
    maxSurge: 1
    maxUnavailable: 25%      # replicas: 4
readinessProbe:
  httpGet: { path: /, port: 8080 }
  periodSeconds: 5

  0.0s apply  deployment/checkout   accepted, resourceVersion 8841
  0.4s ScalingReplicaSet  checkout-7d4f  →  1
  0.4s Killing  checkout-9b21-x7k     grace 30s
  1.1s GET /checkout   503   no endpoints
  2.6s Started  checkout-7d4f-q2m
  3.0s Ready    checkout-7d4f-q2m
  3.2s GET /cart       500   not actually ready
  3.4s GET /checkout   200

Every failed request there carries the reason it failed, stamped at the moment it was emitted rather than guessed at afterwards — which is what makes this gradeable instead of a simulation toy. The defect is not in the syntax. It is in the interaction between a controller, a probe and a clock, and no amount of reading the manifest reveals it. So there is a model of the loop here.