# Optional Kubernetes exercise, not executed by the portable console
Use only a cluster/namespace you own and explicitly select. Build the supplied Dockerfile, load its local image into your own local cluster or replace the image reference with your own published digest. The manifests use apps/v1 Deployment, v1 Service, autoscaling/v2 HPA and policy/v1 PDB where supplied. HPA needs a functioning resource metrics API such as Metrics Server and CPU requests on every measured container. These resources do not install Metrics Server.
The normal /work endpoint is an asynchronous delay workload, so CPU utilization need not rise with queueing. This CPU HPA example is not a promise that the delay workload will scale; a queue-based custom metric requires a separate metrics adapter and policy. Do not increase load beyond the driver bounds to force a screenshot.
The SQLite path is per-pod ephemeral /tmp storage. It is NOT a cross-pod shared database, not durable storage and not a production idempotency service. The two-process shared-file lesson is a separate owned desktop experiment. No PVC, distributed database, ingress, TLS or credentials are provisioned.
A PDB constrains qualifying voluntary evictions; it does not prevent node loss or set a Deployment rolling-update budget. That budget is expressed separately in maxUnavailable/maxSurge. Startup/readiness/liveness and graceful termination still require target-environment tests.
Client YAML parsing or client-side dry-run is not proof that a server accepted resources or that autoscaling/eviction behavior executed. Do not apply to an existing public or production cluster as part of this lesson. Record which checks actually ran.