引導式測驗 · 中等 等級
回填功能和訓練資料測驗
測試對回填為何重要、重新計算期間的時間點正確性以及原始資料保留要求的理解。
問題 1 的 8
何時在特徵定義變更後可以使用回填?
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。
- 何時在特徵定義變更後可以使用回填?
- When historical training examples need a consistent, intentionally updated feature definition
- Whenever any code changes, regardless of model use
- Only after deleting all raw data
- To make every feature timestamp equal to the current time
- 如果模型在訓練時使用的資料中,有些記錄採用了舊的特徵定義,而其他記錄採用了新的特徵定義,且沒有進行回填,會發生什麼情況?
- The model performs identically regardless
- The model learns from an unlabeled shift in what the feature means, behaving like noise or bias tied to record age
- The training pipeline automatically detects and fixes the inconsistency
- Nothing changes since feature definitions never affect training
- 在特徵邏輯變更後,重新從頭計算歷史特徵需要哪些資料?
- A GPU cluster
- A feature that has never changed
- Retained source events or other input data with enough detail for the recomputation
- Only the most recent feature values, without any raw event history
- 為什麼回填必須保留與原始計算相同的事件時間語意?
- Because otherwise the file format would be incompatible
- So Kafka can process the data faster
- Event-time semantics don't need to be preserved
- So the backfilled value doesn't use future data unavailable at that historical timestamp, preserving point-in-time correctness
- 每一次特徵定義變更是否都應自動觸發完整歷史的回填?
- No; choose the affected history based on the new definition, training window, and serving needs
- Yes, even when the model never uses the feature
- Yes, and overwrite all prior versions immediately
- No, historical features should never be recomputed
- How can a team make a feature backfill auditable and traceable to a specific definition?
- By deleting all previous feature definitions
- Record the feature-code/config version and write the result to a versioned output or partition
- By preventing any changes to feature logic
- By requiring manual re-entry of all historical data
- Why can rebuilding a streaming feature from historical data require careful validation?
- Streaming features cannot be backfilled under any circumstances
- Streaming and batch features always use the same code automatically
- Replay and batch paths can differ in ordering, late-event handling, or implementation, so compare them against intended historical semantics
- Kafka does not support historical data of any kind
- What should teams balance when setting raw-data retention for future backfills?
- Deleting raw events immediately after aggregation to save space
- Ignoring retention policies entirely
- Storing only aggregated values and discarding all raw events by default
- The history needed for planned recomputations against privacy, deletion, security, and contractual limits