Operations, HA and Managed Flink Quiz
Check collection paths, memory/recovery boundaries and managed-service support.
- Which resource does the reviewed Kubernetes HA leader election use?
Show Answer
Answer: Fabric8 ConfigMapLock, relying on Kubernetes control-plane availability.
Explanation: It does not add a separate leader-election server or invariably use Lease resources. ZooKeeper HA is also a supported alternative.
- Do HA ConfigMaps, HA storageDir and checkpoint directories store the same data?
Show Answer
Answer: ConfigMaps hold leader information/references, HA storageDir holds JM recovery metadata/job graphs, and checkpoint directories preserve checkpoint state.
Explanation: One HA path setting does not prepare every state file or recovery permission.
- Does an HA-only ConfigMap Role provide all Native Flink permissions?
Show Answer
Answer: No. Native resource management also requires pod/service and other appropriate permissions.
Explanation: Check each caller's needs and inspect API errors, logs and restarts; failures are not always silent.
- What do the Flink and node autoscalers primarily adjust?
Show Answer
Answer: Vertex parallelism and node capacity, respectively.
Explanation: Pressure, quotas, state recovery and placement couple the loops; Flink does not universally act first.
- Must Karpenter create a node for every Pending pod?
Show Answer
Answer: No. Distinguish unschedulable pods from image-pull, PVC, admission and other causes.
Explanation: Constraints, resource requests, instance availability and quotas can prevent provisioning.
- Does a consolidation delay longer than Flink stabilization prevent interruptions?
Show Answer
Answer: No guarantee follows; node failure, Spot reclamation and drift can occur independently.
Explanation: Test provisioning, restoration, backlog catch-up and disruption/PDB policies together.
- Does managed Flink 2.3 support imply availability of every upstream feature?
Show Answer
Answer: No. Check the Java 17/Python 3.12 combination and service-specific exclusions.
Explanation: Restrictions include Java 21, ForSt, Native S3, custom telemetry and Studio. Separate AWS infrastructure HA from application/state responsibilities.
- How should managed KPU costs be compared with EKS Spot?
Show Answer
Answer: Compare total cost and operations at equivalent throughput, latency and recovery objectives.
Explanation: Include orchestration and related storage/network costs for KPUs and interruption/recovery costs for Spot.
- Do RocksDB managed memory and network buffers share one pool?
Show Answer
Answer: They are separate budget regions.
Explanation: In the checked 4GiB configuration, changing managed fraction from 0.4 to 0.5 left network unchanged and reduced task heap. Budgets differ from measured RSS.
- What should be checked when a PodMonitor exists but has no targets?
Show Answer
Answer: Prometheus monitor/namespace selectors, real pod labels, named ports, target namespaces and discovery permissions.
Explanation: This chart's Operator pod lacked a default instance label. Deployment metadata is not automatically a pod label.
- Is block-cache-usage a ratio, and is a Flink Counter always a Prometheus Counter?
Show Answer
Answer: No. Cache usage is bytes, and the reviewed reporter exports Flink Counters as Gauges.
Explanation: Interpret cache capacity/hit/miss too. Histograms map to Summaries; inspect exported TYPE and labels.
- Should kubernetes.cluster-id be set directly in an Operator-managed CR?
Show Answer
Answer: No. The 1.15 validator forbids it; identity is managed from CR name/namespace.
Explanation: Do not blindly copy low-level CLI configuration. kubernetes.namespace and high-availability.cluster-id are also restricted.
- Do two JM replicas and a completed checklist guarantee instant, uninterrupted recovery?
Show Answer
Answer: No. Validate election, node/AZ placement, storage, state restoration, replay time and actual results.
Explanation: Keeping defaults can be valid. Measured SLO/recovery outcomes and operating ownership matter beyond settings.