{
  "lesson": "10-cold-start-mitigation",
  "title": "Cold Start Mitigation for Serverless LLMs",
  "questions": [
    {
      "stage": "pre",
      "question": "Roughly how long does a cold start typically take for a 70B model on a fresh node without mitigations?",
      "options": [
        "3-8 minutes",
        "Over 1 hour",
        "Under 10 seconds",
        "30-60 seconds"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which AWS-side feature does the lesson recommend for pre-seeding container images so step-2 image pull disappears?",
      "options": [
        "Bottlerocket dual-volume architecture referenced from EC2NodeClass",
        "Spot fleet placement",
        "EBS volume snapshots only",
        "ECS task definitions"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which Modal feature provides the closest thing to a \"warm GPU boot in seconds\" by deserializing post-load state directly into HBM?",
      "options": [
        "GPU memory snapshots (checkpoints)",
        "Tiered NVMe-to-DRAM loading",
        "Run:ai Model Streamer",
        "Live migration"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Why does live migration transfer input tokens rather than KV cache between nodes?",
      "options": [
        "KV cache is encrypted and cannot move",
        "Live migration is required by GDPR",
        "Input tokens have larger entropy",
        "Recomputing KV on the destination is cheaper than transferring GB of KV cache over the network"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which serverless layer trades direct GPU cost for predictable readiness by keeping at least one replica live?",
      "options": [
        "Tiered loading",
        "Live migration",
        "Warm pool with min_workers >= 1",
        "Bottlerocket pre-seeding"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Why does the lesson say cold-start mitigation must be stacked across layers rather than picked as a single tool?",
      "options": [
        "Modal owns the entire stack",
        "All five layers are bundled in vLLM",
        "It is a regulatory requirement",
        "No single layer eliminates every step (node provision, image pull, weights load, engine init); stacking layers compresses each step"
      ],
      "correct": 3,
      "explanation": ""
    }
  ]
}