गाइड से जुड़ी प्रश्नोत्तरी · मध्यम स्तर

बैच बनाम वास्तविक समय अनुमान प्रश्नोत्तरी

उत्पादन एमएल सिस्टम में ताजगी, विलंबता, थ्रूपुट, लागत और संचालन द्वारा बैच, स्ट्रीमिंग और ऑनलाइन अनुमान की तुलना करें।

संबंधित मार्गदर्शक पथबैच बनाम वास्तविक समय अनुमान
प्रश्न 1 का 8

Which workload is a natural fit for batch inference?

प्रत्येक प्रश्न का पूर्वावलोकन करें8 प्रश्न

जैसे ही आप प्रश्नोत्तरी लेते हैं उत्तर और स्पष्टीकरण दिखाई देते हैं।

  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