AI Infrastructure on EKS Quiz
15questions on current APIs and operational boundaries.
1. What is JARK, and is it automatically complete?
Answer and explanation
An integration pattern of JupyterHub/Argo Workflows/Ray/Karpenter. Explicitly connect identities, submissions, execution, Pod placement, node supply, storage and authorization.
2. What must be configured for JupyterHub with Cognito?
Answer and explanation
Callback/token/userInfo URLs, scopes, stable username claims and allow policies. Read the client secret from a prepared file; configure MFA/federation in the provider separately. Successful authentication does not imply universal access.
3. What are the roles of the Ray head and Karpenter?
Answer and explanation
GCS is Global Control Service; scheduling interacts with raylets. A head advertising CPUs may run work. Karpenter provisions nodes based on worker-Pod demand and Kubernetes scheduling state.
4. How should DRA and device plugins be compared?
Answer and explanation
DRA provides structured requests, attributes and allocation through DeviceClass/ResourceSlice/ResourceClaim. Device plugins also support MIG/time-slicing; they are not universally exclusive-only. Features depend on drivers, hardware and gates.
5. What matters about MIG profiles and isolation?
Answer and explanation
3g.20gb describes one profile, not three20GB devices. MIG partitions hardware but does not replace host/driver/authorization isolation. MPS/time-slicing are not security boundaries.
6. Can one GPU Operator version establish all DRA support?
Answer and explanation
No. Check Kubernetes APIs, drivers/hardware/CDI and individual gates. Operator26.7 GPUCluster is mutually exclusive with ClusterPolicy; standalone0.5 has an opt-in guard against device-plugin collisions.
7. What should be checked when integrating Langfuse?
Answer and explanation
Verify current SDK/backend dependencies, DB/ClickHouse/object storage, file credentials, tracing and sensitive-data retention. Old2.x Deployments or trace() calls are not current4.x APIs.
8. How should EFS, FSx and Mountpoint be selected?
Answer and explanation
Compare I/O, authorization, namespace/PVC scope, capacity and filesystem semantics. EFS requests are not quotas; Mountpoint CSI2.8 uses static PVs for existing buckets and is not fully POSIX.
9. How should EFA bandwidth and interface counts be interpreted?
Answer and explanation
Distinguish aggregate instance bandwidth from per-interface values; do not multiply an already aggregate figure. Verify same-AZ placement, interfaces, drivers/libfabric/NCCL, security groups and Pod resources. RAID0 does not enable EFA.
10. How should GPU memory and XID metrics be interpreted?
Answer and explanation
FB_USED/FREE are MiB gauges; a high ratio is not necessarily OOM. XID_ERRORS is the last code gauge, not a counter for increase(). Investigate actual errors, caches, allocation failures and device-specific limits.
11. Does Karpenter consolidation directly use GPU utilization metrics?
Answer and explanation
It uses workload requests, scheduling feasibility, prices and disruption constraints rather than a DCGM20% threshold. Limits/budgets are not absolute cost/failure guarantees; verify termination and recovery.
12. Who publishes ResourceSlices, and what do they represent?
Answer and explanation
Drivers publish real device inventory with typed attributes/capacity. Creating an arbitrary slice does not create GPUs. CEL and matchAttribute must follow the actual published schema.
13. Does requesting a GPU for Milvus complete RAG?
Answer and explanation
No. Match supported images/indexes, embedding dimensions/revisions, metrics/parameters, tenant filtering and update/deletion lifecycle. Validate result authorization and handle absent evidence.
14. How should pending GPU Pods and node failures be investigated?
Answer and explanation
Counting pending GPU-requesting Pods does not diagnose the cause. Inspect events, PVCs, affinity, taints, quotas, claims and images, and test checkpoint/resume. Multiple requesting containers count once per Pod.
15. Does MCP provide a standard Kubernetes auto-discovery gateway?
Answer and explanation
No. It defines operations such as tool listing/calling. Choose an actual server/gateway release, transport and authorization design. Invented images/labels/config or a URL environment variable do not implement discovery/execution.