Library branch kiosks need continuous catalog-vision inference. Which Kubernetes workload type should ops deploy?
Select an answer to reveal the explanation.
Short Explanation
Kiosks need the lights left on—a Deployment plus Service keeps catalog-vision answering all day. A one-shot Job that exits after one classify leaves patrons staring at a blank screen.
Full Explanation
Continuous catalog-vision for library branch kiosks needs a steady Deployment plus Service: always-on replicas, health probes, and a stable virtual IP/DNS name the kiosks call all day. The Deployment controller replaces failed pods; the Service load-balances across Ready backends so patrons keep getting classifications without operator babysitting.
A one-shot Job that exits after the first successful classification leaves the kiosk with no backend for the rest of the day. A CronJob that only wakes during annual inventory week matches a batch calendar, not continuous branch hours. An empty Namespace with no pods until patrons complain is monitoring-by-outrage, not operations.
Exam caveat: choose replica count and PDB for branch hours; consider scale-down after close if budget-constrained, but do not confuse that with Job/CronJob semantics. GPU requests must match node labels, and image pull + model load time should be absorbed by readiness gates before traffic is admitted.
Operational check: from a kiosk test client, hit the Service continuously during a rolling restart and confirm zero or near-zero failed requests. Alert on Ready replicas < desired and on inference error rate so empty namespaces never become the failure mode.