Practise the standard verbs for keeping a feature store consistent between training and serving.
0 / 5 completed
1 / 5
Fill in: 'We ___ point-in-time correctness in the feature store so training data never leaks information from after the prediction moment.'
We 'enforce correctness' — the standard, established feature-store collocation for guaranteeing no future information leaks into training data. The other options aren't the recognised term here.
2 / 5
Fill in: 'Computing features differently in training and serving pipelines can ___ a subtle skew that quietly degrades live model accuracy.'
We say a training-serving mismatch will 'leave' a skew in place — the standard, natural collocation for the resulting problem. The other options aren't idiomatic here.
3 / 5
Fill in: 'We ___ a single feature definition between the training and serving paths so the two pipelines can never silently drift apart.'
We 'share a definition' — the standard, simple collocation for reusing one computed logic across two consumers. The other options are less idiomatic here.
4 / 5
Fill in: 'We ___ historical values for a newly added feature so a model can be retrained on the same feature it will see in production.'
We 'backfill' values — the standard, established collocation for computing historical data for a feature added after the fact. The other options aren't the recognised term here.
5 / 5
Fill in: 'We ___ feature freshness timestamps regularly, so a stale upstream feed doesn't quietly feed outdated values into live predictions.'
We 'audit' timestamps — the standard, simple collocation for periodically verifying data is as recent as expected. The other options are less idiomatic here.