ML Systems Design

• ML system lifecycle: problem framing → data → training → evaluation → deployment → monitoring
• Data management: feature stores, data versioning (DVC), labelling pipelines, data quality checks
• Training infrastructure: distributed training (data parallel, model parallel), experiment tracking (MLflow, W&B)
• Model evaluation: offline metrics, A/B testing, shadow deployment, interleaving experiments
• Model serving: batch vs real-time inference, model registry, model versioning
• Feature engineering: online vs offline features, feature freshness, feature serving latency
• ML pipelines: orchestration (Airflow, Kubeflow, Metaflow), reproducibility
• Monitoring: data drift, concept drift, model degradation, alerting