গাইড লিঙ্ক ক্যুইজ · মাঝারি স্তর

Kubernetes কুইজে অটোস্কেলিং মডেল ইনফারেন্স

মডেল-সার্ভিং অটোস্কেলিং সংকেত কনফিগার করুন এবং GPU শিডিউলিং, মডেল স্টার্টআপ বিলম্ব, সারি এবং স্কেল-টু-জিরো ট্রেডঅফের জন্য অ্যাকাউন্ট।

প্রতিটি প্রশ্নের পূর্বরূপ দেখুন8 প্রশ্ন

আপনি কুইজ নেওয়ার সাথে সাথে উত্তর এবং ব্যাখ্যাগুলি উপস্থিত হয়।

  1. 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
  2. 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
  3. 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
  4. 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
  5. 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
  6. 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
  7. 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
  8. 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