quiz bu lëkkale ak tegtal · Diggante Niveau
Laajte ak tontu ci jëmmal sistem ML
Jëfandikool ab jego bu lalu ci jëmmal sistem bu tàmbalee ci mébetu produit bi ak ay done jaare ko ci jàngat, serwiis, saytu, delloosi ak def kompromis ci liggéey bi.
Laaj 1 de 8
What should be clarified before selecting a model architecture in a system design?
Xoolaatal laaj bu nekk8 laaj
Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.
- What should be clarified before selecting a model architecture in a system design?
- Start by choosing a cloud provider and GPU family
- Select a neural architecture before defining the task
- Product goal, users, constraints, and success measures
- Choose a framework based on the largest package catalog
- Why separate candidate retrieval from ranking in a recommendation design?
- Ranking creates the training data for all user events
- Retrieval narrows a large set before a richer model orders candidates
- Retrieval makes offline evaluation unnecessary
- The two stages must use identical scoring models
- Which split can reduce leakage when requests have time structure?
- A random row split even when future behavior is the target
- Training on future observations to mimic random traffic
- Splitting users so the same user is present in every set
- A chronology-respecting split
- What does a serving path typically include?
- Feature retrieval, preprocessing, model inference, and response handling
- Training jobs, hyperparameter search, and offline evaluation only
- Dataset curation, labeling, and model architecture only
- Model registry, pipeline scheduler, and dashboard without request handling
- Why check feature availability between training and serving?
- Features used in training are automatically available in production
- Online features can be computed with different definitions without impact
- A missing or differently computed serving feature can create skew
- Feature consistency affects only model licensing
- Which metric pair covers service behavior and model behavior?
- Training loss and validation accuracy alone
- Latency and error rates plus input or output quality signals
- Number of API calls and model-file size only
- Service uptime alone without model-behavior checks
- What should a rollback plan specify?
- Re-train from the newest data without comparing versions
- Restore weights while leaving incompatible preprocessing and serving configuration
- Remove health checks to prevent deployment alerts
- Conditions and steps for returning to a known-good model or service version
- What makes a system design explanation persuasive?
- Tie architectural choices to requirements, tradeoffs, and tests
- Memorize a cloud architecture template regardless of constraints
- Describe model layers without the surrounding data and serving path
- List services without explaining their measurements or failure handling