MLOps
Experiment tracking, model registries, deployment strategies, monitoring and drift detection, and feature stores — running ML as a reliable production system.
Questions on the operational discipline around ML models once they leave a notebook — tracking what was actually trained on what data, registering and versioning models, deploying changes safely, detecting when a live model has quietly stopped being right, and keeping training and serving features consistent.
Subtopics
- Experiment tracking and model registriesSeniorWhy 'the notebook that produced it' isn't reproducibility, what a tracking system must capture, model registries and lineage, and reproducibility in practice.6 concepts
- Model deployment strategiesSeniorShadow deployment, canary rollouts, A/B testing, blue-green for models, and the rollback plan every deployment needs before it ships.6 concepts
- Monitoring and drift detectionSeniorData drift vs concept drift, why offline accuracy can't be monitored directly in production, proxy metrics, and building alerts that catch silent degradation.6 concepts
- Feature stores and training/serving consistencySeniorTraining-serving skew, online vs offline feature stores, point-in-time correctness, and why feature pipelines need the same rigor as model code.6 concepts