引導式測驗 · 中等 等級

批次與即時推論測驗

透過生產機器學習系統中的新鮮度、延遲、吞吐量、成本和操作來比較批量、串流和線上推理。

相關引導路徑批次與即時推論
問題 1 的 8

Which workload is a natural fit for batch inference?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. Which workload is a natural fit for batch inference?
    • Scoring next week's inventory forecasts overnight
    • Blocking each payment until a score returns
    • Responding to every chat token synchronously
    • Updating a decision before a user clicks
  2. Why use real-time inference for a payment authorization decision?
    • It has no latency requirement
    • The decision is needed before the transaction proceeds
    • It can always be computed the night before
    • It avoids all service operation
  3. What distinguishes streaming inference from a nightly batch job?
    • It always returns an HTTP response to a user
    • It only runs once per calendar year
    • It processes arriving events continuously with bounded delay
    • It uses no event ordering
  4. Which design combines bulk precomputation with fresh request context?
    • It removes feature freshness concerns
    • It eliminates all model-serving infrastructure
    • It guarantees identical rankings for every user
    • It combines efficient bulk work with fresh request context
  5. What contributes to end-to-end real-time latency?
    • Only the model's matrix multiplication time
    • Network, feature retrieval, preprocessing and postprocessing as well as model execution
    • Training dataset size only
    • The number of offline evaluation metrics
  6. Why can a real-time endpoint cost more than a scheduled job?
    • Real-time models cannot be quantized
    • Batch jobs always require more hardware
    • Always-available capacity may sit idle between requests
    • Requests have no network overhead
  7. What should a streaming system handle for late events?
    • Static batch partition names only
    • Only HTTP status codes
    • The model's final training epoch
    • Event-time windows, ordering and replay behavior
  8. Which principle should drive choosing among inference modes?
    • Freshness, latency, throughput, cost and operational needs
    • Always choose the lowest single-request latency
    • Use the most complex infrastructure available
    • Pick the mode with the newest model name