ML / AI System Design
ML pipeline, feature stores, model serving, online vs batch, when to use deep learning
Technical Reference & Key Concepts
ML Pipeline Stages
- Data Ingestion: Collect raw data from logs, databases, event streams.
- Feature Engineering: Transform raw data into features. Normalization, encoding, embedding.
- Feature Store: Centralized registry for features (Feast, Tecton). Reusable, consistent train/serve.
- Training: Model selection, hyperparameter tuning, experiment tracking (MLflow).
- Evaluation: Offline metrics (precision, recall, AUC) + online A/B testing.
- Serving: Batch (scheduled predictions), online (low-latency API), or edge (on-device).
- Monitoring: Data drift, model drift, prediction distributions. Retrain triggers.
When to Use Deep Learning
- Yes: Unstructured data (images, audio, text), complex patterns, huge datasets.
- No: Structured/tabular data with <100K rows. Gradient-boosted trees (XGBoost) often win.