Practise the vocabulary for ML deployment patterns: shadow mode, champion/challenger, canary, feature stores, and rollback.
0 / 6 completed
1 / 6
Shadow mode deployment for a new ML model means:
Shadow mode (also called dark launch) is risk-free validation: the production model still serves users, but the new model's predictions are captured and compared offline to ensure quality before promotion.
2 / 6
Champion/challenger testing in model deployment refers to:
Champion/challenger routes a portion of live traffic to the challenger model — e.g. 10% challenger, 90% champion — to compare business metrics (not just offline metrics) in real conditions.
3 / 6
A 'point-in-time correct' feature store is important because:
Point-in-time correctness prevents training-serving skew and data leakage: when training on historical data, only features that were available before each event timestamp are used.
4 / 6
Model rollback in a production ML system means:
Rollback is a critical operational capability: if a newly deployed model causes metric regression, the system should quickly revert to the previous champion model while the issue is investigated.
5 / 6
Training-serving skew means:
Training-serving skew is a common production failure: if training uses raw data but serving uses a preprocessed version (or vice versa), the model is being applied to a different data distribution than it was trained on.
6 / 6
A canary deployment for an ML model means:
Canary releases limit the blast radius of a bad deployment — if the 5% canary shows degraded business metrics or errors, rollout is halted before 95% of users are affected.