Part 4: Performance and Cost Tuning Quiz
Baseline: Spark 4.2.0 / Karpenter 1.14 · September 2026.
1. Are R-series and NVMe always fastest for shuffle?
Show answer
No. Measure task/stage time, spill, shuffle fetch wait, skew, and CPU/memory/disk/network bottlenecks. Compare available capacity and cost per successful job; other families and EBS are not inherently unsuitable.
2. What should be checked when evaluating Graviton?
Show answer
Validate arm64 support in the Spark image, JNI, codecs, BLAS, Python wheels and plugins, then benchmark representative work. An existing tested multi-architecture image can avoid building your own.
3. Is it safe to format an unmounted nvme device?
Show answer
No. EBS also appears as NVMe on Nitro, and an unmounted disk can contain data or partitions. Device names and mount status do not establish that it is disposable instance store.
4. What does instanceStorePolicy: RAID0 configure on AL2023?
Show answer
Karpenter uses NodeConfig to configure instance-store RAID0 for kubelet/containerd ephemeral storage and advertises allocatable capacity. RAID0 is not backup; a complete NodeClass still needs its AMI, IAM, subnet and other settings.
5. Why can this lab use NVMe scratch without hostPath?
Show answer
The prepared NodeClass backs the kubelet filesystem with instance store, so disk-backed emptyDir uses it. Different node configuration or tmpfs changes the backing storage.
6. Which scratch-volume name does Spark recognize?
Show answer
A name such as spark-local-dir-scratch, with the spark-local-dir- prefix and a suffix. The old spark-local-dir does not match and can cause an additional emptyDir at the same mount path.
7. Does emptyDir sizeLimit reserve disk space?
Show answer
No. Consider requests, node allocatable, other pods, logs and writable layers. The node can fill first; memory-backed tmpfs consumes the memory budget.
8. What does the On-Demand-driver/Spot-executor split not guarantee?
Show answer
It does not eliminate driver failure/drift/expiration or make executor loss harmless. A Spot-only selector does not allow On-Demand fallback and may leave pods Pending during shortages.
9. How do taints/tolerations and node selectors differ?
Show answer
A toleration permits a taint; a selector chooses matching nodes. Other pods may have the same toleration, and NoSchedule does not evict existing pods.
10. Do Spot warnings and decommission flags guarantee completed migration?
Show answer
No. Warnings are best-effort and hibernation has no two-minute window. EventBridge/SQS/interruption handling, drain, hooks/signals, time, peer capacity and networking all matter. A 120s pod grace does not extend cloud lifetime.
11. Does any remote storage automatically become shuffle fallback?
Show answer
No. Configure spark.storage.decommission.fallbackStorage.path and working filesystem/permissions. RDD-cache and shuffle recovery differ; recomputation, fetch failures and retry limits can still fail the job.
12. Which signals drive DRA and Karpenter?
Show answer
DRA adjusts executor count from Spark task backlog and idle/cache/shuffle state. Karpenter provisions nodes from pod requests and scheduling constraints. metrics-server is not required; allocation.batch.size controls Kubernetes allocator creation rate.
13. Does consolidateAfter greater than executorIdleTimeout guarantee DRA releases first?
Show answer
No. The timers start from different events, and shuffle/cache can retain executors. WhenEmpty reduces consolidation of ordinary workload pods, but drift, expiration and Spot termination remain separate.
14. Are 4g JVM heap and cores=2 the entire pod memory and CPU limit?
Show answer
No. This example adds 409MiB default overhead for 4505MiB total memory. Cores affect the default CPU request and Spark concurrency; set the CPU limit explicitly with spark.kubernetes.executor.limit.cores. Review PySpark, off-heap and explicit overhead separately.