Cloud Native Operations

English

한국어

English

한국어

Appearance

EKS Auto Mode Scaling Behavior Quiz ​

Related Document: Scaling Behavior

Multiple Choice Questions ​

1. What does consolidationPolicy: WhenEmptyOrUnderutilized permit? ​

  • A) Only empty-node removal
  • B) Cost-saving consolidation of eligible empty or non-empty nodes
  • C) Keeping all nodes indefinitely
  • D) Removal only at a fixed clock time
Show Answer

Answer: B) Cost-saving consolidation of eligible empty or non-empty nodes

Explanation: It permits removal or cheaper replacement when the workloads remain schedulable under the relevant constraints. Resource requests, PDBs, budgets and annotations matter; it is not simply a measured-CPU threshold.

yaml
# Fragment inside a NodePool; eligibility delay, not a deletion deadline
disruption:
  consolidationPolicy: WhenEmptyOrUnderutilized
  consolidateAfter: 1m

The timer resets when Pods are added or removed. One minute is an eligibility/stability delay, not a guarantee that deletion or replacement finishes in one minute. The current AWS API also supports Balanced, which considers disruption cost alongside savings.

2. Which command lists NodeClaims? ​

  • A) kubectl get nodes --show-claims
  • B) kubectl get nodeclaims
  • C) kubectl describe karpenter claims
  • D) kubectl get ec2-nodes
Show Answer

Answer: B) kubectl get nodeclaims

Explanation: NodeClaims describe desired node configuration and provider lifecycle/status, not just the instant of provisioning. Inspect conditions and the linked node name; Drifted is a condition, not the old fictitious karpenter.sh/drift-hash node annotation.

bash
kubectl --context "$CLUSTER_NAME" --request-timeout=15s get nodeclaims -o wide
: "${NODECLAIM_NAME:?Select a NodeClaim from the list}"
kubectl --context "$CLUSTER_NAME" --request-timeout=15s describe nodeclaim "$NODECLAIM_NAME"

3. When can an unschedulable Pod lead Auto Mode to provision capacity? ​

  • A) Every Pod has spent five minutes Pending
  • B) The Pod cannot fit existing capacity and a compatible eligible pool/capacity can satisfy its constraints
  • C) Node count is below a fixed global threshold
  • D) Only after an operator runs a manual scale command
Show Answer

Answer: B) The Pod cannot fit existing capacity and a compatible eligible pool/capacity can satisfy its constraints

Explanation: The scheduler's inability to place the Pod, matching NodePool/NodeClass constraints, quotas and actual capacity all matter. A scheduled Pod waiting for an image or init container can be Pending without needing a new node. Auto Mode does not guarantee immediate allocation or the earlier unverified 40–90-second estimate.

4. Which can block consolidation of a candidate node? ​

  • A) An applicable do-not-disrupt annotation
  • B) A PDB blocks required workload eviction
  • C) The consolidateAfter stability interval has not elapsed
  • D) All of the above
Show Answer

Answer: D) All of the above

Explanation: Each can prevent or delay a consolidation action. The former answer incorrectly included DaemonSet-only nodes as a blocker: such nodes can qualify as empty.

yaml
# Metadata fragment on the intended Node or Pod, not NodePool metadata
metadata:
  annotations:
    karpenter.sh/do-not-disrupt: "true"

This is not a guarantee against every disruption reason. Expiration, interruptions and termination-grace limits must be considered separately. Remove temporary protection after its intended maintenance need ends.

5. Which can contribute to drift or replacement? ​

  • A) Relevant desired NodeClass configuration no longer matches the node
  • B) AWS selects a new managed Auto Mode AMI
  • C) Resolved subnet/security-group selection changes
  • D) All of the above can contribute
Show Answer

Answer: D) All of the above can contribute

Explanation: Inspect actual NodeClaim conditions and provider decisions. Not every edit causes drift: widening requirements can leave existing instances compatible, and weight/limits/disruption settings are behavioral. Editing rules on the same security group is not the same as resolving a different group. Graceful replacement is subject to scheduling and budgets and can be parallel; it is not always a sequential one-node-at-a-time process.

6. How should you improve Auto Mode startup latency? ​

  • A) Switch amiFamily to AL2023
  • B) Measure the workload path and test supported capacity, image and storage changes
  • C) Assume the smallest possible volume is always fastest
  • D) Treat 40–90 seconds as a guaranteed end-to-end SLO
Show Answer

Answer: B) Measure the workload path and test supported capacity, image and storage changes

Explanation: AWS selects the managed Bottlerocket image. Auto Mode does not offer the old amiFamily comparison knob. Separate node registration, image/initialization time and application readiness, and include failures/timeouts in measurements.

The previous quiz's AL2023 20–40 seconds and Bottlerocket 15–25 seconds are preserved as unverified historical teaching estimates. They are not measured Auto Mode results and do not establish a current boot-time advantage or supported AMI choice.