KubeRay Operator Quiz
Multiple Choice Questions
- What does installing the KubeRay operator start?
- A) Every Ray workload automatically
- B) The reconciler; workload CRs still need to be created
- C) Every GPU node
- D) A Deployment for every worker
Show Answer
Answer: B
Operator installation is separate from creating RayCluster, RayJob, or RayService resources.
- What does chart 1.7.0 install, and what is RayCronJob’s default state?
- A) Only three CRDs exist
- B) An installed RayCronJob CRD always enables scheduling
- C) Four CRDs, with the RayCronJob feature gate disabled by default
- D) No v1 API exists
Show Answer
Answer: C
Distinguish RayCluster, RayJob, RayService, and RayCronJob.
- What is RayJob’s default cleanup behavior?
- A) Omitted shutdownAfterJobFinishes enables cleanup
- B) TTL 0 deletes every EC2 instance and PVC
- C) shutdownAfterJobFinishes defaults false; configure cleanup and preservation
- D) External clusters are always deleted
Show Answer
Answer: C
Check TTL, deletionStrategy, and ownership of shared versus created clusters.
- What does RayService incremental upgrade require?
- A) A feature gate alone guarantees no downtime
- B) Strategy, Gateway API/implementation, capacity, readiness, and draining
- C) Only editing one existing Pod image
- D) A mandatory RayJob
Show Answer
Answer: B
It shifts traffic between clusters; it is not simply an in-place Pod rolling update.
- Which scaling sequence is appropriate?
- A) Ray autoscaler directly provisions all EC2 capacity
- B) Ray demand → KubeRay Pod reconciliation → Kubernetes placement/capacity provisioning
- C) Karpenter calls Ray actor methods
- D) Both controllers own the identical resource
Show Answer
Answer: B
Image, PVC, or authorization-related Pending states are not fixed solely by adding nodes.
- How should idleTimeoutSeconds 60 be interpreted?
- A) Every worker disappears exactly 60 seconds later
- B) Reviewed global default; consider group overrides, bounds, activity, and drain conditions
- C) Whole-RayJob TTL
- D) Karpenter’s fixed provisioning time
Show Answer
Answer: B
Configuration duration and actual deletion completion are different.
- Which GPU resource precedence statement is accurate?
- A) Only Pod limits are used; overrides are ignored
- B) Structured group resources and rayStartParams can override limits
- C) Every Pod always has one GPU
- D) Logical settings create more physical GPUs
Show Answer
Answer: B
Align Ray logical resources with device plugins, drivers, and visible hardware.
- Does a Helm chart upgrade automatically update existing CRD schemas?
- A) It always upgrades and deletes them
- B) No; check stored CR compatibility and the separate CRD update procedure
- C) CRDs are ordinary Pods
- D) Always deleting CRDs first is safe
Show Answer
Answer: B
Understand Helm crds/ lifecycle limits and the impact of CRD deletion.
Short Answer Questions
- Why are worker replica count and Ray Pod count not always equal?
Show Answer
With numOfHosts, one group replica can correspond to multiple hosts/Pods. Compare the actual spec with generated resources.
- Why does enabling Ray token authentication not complete security configuration?
Show Answer
It is separate from TLS and does not replace authentication/authorization for every application endpoint. Review access paths, secret delivery, version support, and organizational policy.