引導式測驗 · 中等 等級
盛宴開源特徵商店測驗
測試對 Feast 的離線/線上商店拆分、時間點連接及其解決的訓練服務偏差問題的理解。
問題 1 的 8
What problem can consistent feature definitions and retrieval through Feast help reduce?
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。
- What problem can consistent feature definitions and retrieval through Feast help reduce?
- Training-serving skew caused by inconsistent feature logic across training and production
- Slow model training due to large datasets
- Running out of storage space for raw data
- Overfitting caused by too many features
- In Feast, what is the offline store typically used for?
- Serving features with millisecond latency in production
- Retrieving historical feature values for training
- Storing model weights
- Logging errors from the online store
- What does Feast's get_online_features call do?
- Builds a point-in-time correct training set from historical data
- Registers a new feature definition
- Fetches a customer's current feature values from a low-latency store for live prediction
- Deletes stale features from the offline store
- What does the materialize step in Feast do?
- Deletes old feature definitions
- Trains a new model using the latest features
- Converts CSV files into Parquet format
- Copies the latest feature values from the offline store into the online store
- According to the guide, does Feast typically compute feature engineering transformations itself?
- Usually Feast retrieves or materializes values transformed elsewhere, although supported on-demand and streaming workflows can run transformations
- Yes, Feast always performs the transformations from raw data
- Yes, but only for numeric features
- No, Feast cannot connect to any data source
- What does a Feast FeatureView group together?
- A list of all models in production
- Related features along with a time-to-live and a reference to their data source
- Only the online store configuration
- A single row of training data
- Why does the online store hold only the latest value per entity key rather than a full history?
- Because the online store cannot store more than one value
- Because historical values are deleted for privacy reasons
- Because serving needs current state, not history
- Because Redis and DynamoDB do not support timestamps
- What broader trend does the guide describe for the future of feature stores like Feast?
- A move away from offline stores entirely
- Replacing all online stores with flat CSV files
- Feature stores becoming obsolete as models get larger
- Tighter integration with streaming systems so online feature freshness approaches real time