5 exercises on MLOps — feature stores, model registries, drift detection, and serving patterns.
0 / 5 completed
1 / 5
What is a feature store in MLOps?
Feature store: it solves training-serving skew — the same feature logic runs in the offline store (for training) and the online store (for real-time inference). Tools like Feast, Tecton, and Databricks Feature Store manage feature definitions and ensure consistency.
2 / 5
What is a model registry?
Model registry: MLflow Model Registry, SageMaker Model Registry, and Vertex AI Model Registry let teams track which model version is in production, compare runs, and manage promotion workflows from experimentation to production deployment.
3 / 5
What is model drift and what are its two main types?
Drift detection: data drift (covariate shift) means the feature distributions seen in production differ from training data. Concept drift means the underlying pattern has changed (e.g. user behaviour shifts). Both cause model performance to degrade silently without alerting.
4 / 5
What triggers model retraining in a production MLOps system?
Retraining triggers: a mature pipeline monitors data quality, input distributions (via PSI or KS tests), and downstream business metrics. When drift or degradation is detected beyond a threshold, the pipeline automatically kicks off data collection, retraining, evaluation, and promotion.
5 / 5
What is model serving and what are two common patterns?
Model serving: online serving (e.g. TensorFlow Serving, Triton, BentoML) targets millisecond latency for per-request predictions. Batch scoring processes millions of records offline using Spark or a job scheduler, suitable for nightly recommendation pre-computation.