Models are not ship-once artifacts. Lifecycle covers versioning, deployment, monitoring, retraining, and retirement—often called MLOps when automated at scale.
Lifecycle stages
- Experiment — notebooks, tracked metrics
- Register — versioned artifact with metadata
- Deploy — batch or online serving
- Monitor — latency, errors, data drift
- Retrain — scheduled or triggered by drift
- Retire — deprecate unsafe or obsolete models
Drift types
- Data drift — input distribution changes
- Concept drift — relationship X→y changes
- Upstream schema changes — broken pipelines
Version tag example
model_card = {
"name": "churn_v3",
"trained_on": "2025-09-01",
"metrics": {"auc": 0.82},
"owner": "ml-platform",
}
print(model_card["name"])Practice: Optional snippets use pandas-style pseudocode—run with Pandas locally if you want tactile practice.
Important interview questions and answers
- Q: Model registry?
A: Central store of approved model versions with lineage and metrics. - Q: Concept drift example?
A: Pandemic shifts shopping behavior—old churn patterns fail.
Self-check
- List four lifecycle stages after experiment.
- Define data drift vs concept drift.
Tip: Register model version + training date before any production deploy.
Interview prep
- Data drift?
- Input feature distribution changes versus training.
- Concept drift?
- Relationship between features and target changes over time.
- Model registry?
- Versioned store of approved models with lineage and metrics.