M5.3 module quiz
Ten questions. 80% to pass (8 of 10). Open book, unlimited retakes.
Answers are collapsed under each question. Answer first, then check.
Q1
Kubernetes cost is driven by:
A. Resource requests, which reserve node capacity whether used or not
B. Pod count per namespace
C. Actual measured utilization
D. The size of the container image as it is stored on the node disk
Show answer
Correct: A. Which is why a cluster at 20% utilization can be at 95% requested capacity, and why the two numbers tell completely different stories.
Q2
Requests set too high cause:
A. Pod eviction whenever the node comes under sustained memory pressure
B. CPU throttling at the limit
C. Scheduling failures only
D. Reserved capacity nobody uses, so nodes are added for phantom demand
Show answer
Correct: D. The cluster autoscaler is doing its job correctly; it is responding to a request figure that does not reflect reality.
Q3
Requests set too low cause:
A. Wasted reserved capacity on nodes
B. Slower pod scheduling
C. Eviction and throttling under load
D. Higher cost per node
Show answer
Correct: C. Which is why the target is a p95-to-p99 sweet spot rather than the minimum that works on a quiet day.
Q4
An HPA reporting ScalingLimited with reason TooFewReplicas means:
A. The autoscaler wants to scale below minReplicas
B. The HPA has hit its configured maximum value
C. The deployment is paused
D. Metrics are unavailable
Show answer
Correct: A. It is the strongest over-provisioning signal available, and it is what the HPA-signal rightsizing rules read.
Q5
Two pods instead of one reduce failure probability by roughly:
A. 100,000×
B. 2×
C. 1,000×
D. 10×
Show answer
Correct: C. On the standard independence assumption, and correlated failures such as node loss, zone loss or a bad deploy erode even that.
Q6
Which idle workload shape has no corresponding rule?
A. Stopped deployment
B. Suspended CronJob
C. Service with no endpoints
D. A pod stuck Pending
Show answer
Correct: D. Pending and CrashLoopBackOff are pod-level states, and the rule framework evaluates workload objects. Detect those with kubectl or cluster alerting.
Q7
The orphan K8s rules cover:
A. Unbound PVCs, released PVs, and services with no endpoints
B. All PV states, bound included
C. Failed PVs plus any old unbound PVs across the cluster
D. Only PVCs, never volumes
Show answer
Correct: A. Exactly three. There is no Failed-PV rule and no old-unbound-PV rule, and the orphan category is exempt from the low-savings floor so a small PVC still surfaces.
Q8
K8s security signals correlate with cost because:
A. Security incidents are expensive to detect, to resolve, and to report to the regulator afterwards
B. Privileged, root and host-network workloads are frequently legacy designs that are also over-provisioned
C. They require larger nodes to run on
D. Security scanning itself costs money
Show answer
Correct: B. Modernising them tends to reduce resource needs as a side effect, which is what makes the combined ROI argument work.
Q9
Orphan PVCs accumulate because:
A. Kubernetes never deletes volumes
B. Cloud providers charge a fee for each volume that is deleted
C. The default reclaim policy retains volumes after the claim is gone
D. They are hidden from kubectl output
Show answer
Correct: C. Setting the reclaim policy deliberately is the prevention; cleanup is the remediation for what already accumulated.
Q10
Single-replica production workloads are acceptable when:
A. Never, under any circumstances
B. The service tier genuinely tolerates a 1-to-5 minute outage on pod failure
C. The workload is entirely stateless and holds no persistent data at all
D. A PVC is attached to the pod
Show answer
Correct: B. It is a tier decision rather than a technical one, and stating the tolerated outage explicitly is what turns it from an accident into a choice.
What’s next
Back to K8s cost discipline.