Model versioning на собеседовании Data Scientist

Зачем

Production ML — multiple versions exist concurrently:

• Production v1 (current).
• Staging v2 (testing).
• Experiments v3, v4 (R&D).

Need clear identification.

Model registry

Stores models с metadata.

MLflow Model Registry.

mlflow.register_model("runs:/<run_id>/model", "fraud_detector")
client.transition_model_version_stage(
    name="fraud_detector", version=3, stage="Production"
)

Tools: MLflow, Weights & Biases, Vertex AI Model Registry, SageMaker Model Registry.

Naming convention

Semver-like.

fraud_detector v2.1.0
  major: incompatible API change.
  minor: new features (recall improved 5%).
  patch: bug fixes / stability.

Or timestamp-based: fraud_detector_2026-05-07_abc123.

Каждое version имеет:

• Code commit hash.
• Training data version.
• Metrics on benchmarks.
• Author / approvers.

Stages

None (initial) → Staging (validate) → Production (live) → Archived (retired).

Transitions tracked. Audit trail.

В strict regulated — manual approval для Production.

Rollback

При issue с new version — rollback.

Mechanisms:

• Feature flag — toggle between versions.
• Routing tier — switch traffic.
• Re-deploy old image.

Speed. < 5 min. Automated.

Database considerations. Model schema changes — migrations forwarded compatible когда possible.

Связанные темы

• MLflow и DVC для DS
• Canary и shadow deployment ML для DS
• MLOps мониторинг для DS
• Feature Store для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на индустриальных MLOps practices.

Тренируйте Data Science — откройте тренажёр с 1500+ вопросами для собесов.