Kvíz spojený s průvodcem · Střední úroveň
Autoscaling Model Inference na Kubernetes Quiz
Nakonfigurujte signály automatického škálování pro modelování a zohledněte plánování GPU, zpoždění spouštění modelu, fronty a kompromisy škálování na nulu.
Otázka 1 z 8
Which signal may represent inference pressure better than CPU alone?
Náhled každé otázky8 otázky
Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.
- Which signal may represent inference pressure better than CPU alone?
- Queue depth or request age
- The model's training accuracy
- The number of YAML files
- The checkpoint's creation date
- Why can a GPU pod remain pending after HPA requests more replicas?
- HPA automatically adds physical GPUs
- No node may have available GPU capacity matching the pod request
- The model has no output signature
- Queue depth is always zero
- What can a large model's cold start include?
- A training data label join
- Only a request counter update
- Image and weight download plus model initialization
- A decrease in node capacity
- What can scaling an inference workload to zero trade for idle cost savings?
- No need for readiness checks
- Higher GPU memory at all times
- Guaranteed zero queueing
- Lower idle cost but potentially longer first-request latency
- What does KEDA commonly add to Kubernetes scaling?
- A GPU driver in each pod
- Event-source or external-metric triggers for workloads
- Model calibration metrics
- A replacement for container images
- Why configure stabilization windows or scale policies?
- To remove GPU resource requests
- To force every pod to load identical model weights
- To reduce oscillation and abrupt replica changes
- To guarantee low latency
- What should readiness indicate for a model-serving pod?
- The cluster has no queued work
- The container process has merely started
- The training pipeline is complete
- The model has loaded and the pod can handle requests
- Why set a maximum replica or GPU limit?
- To bound cost and respect hardware quota
- To guarantee capacity during all traffic spikes
- To prevent any scale-up
- To change model prediction values