---
name: ML Pipeline
description: Reproducducible training and serving.
---
# ML Pipeline

A model in a notebook is a prototype. A pipeline that trains, version, and serves it reproducibly is what reaches production.

## Components
- **Reproducible data prep.** Same input → same features, every run. Snapshot the input data version alongside the model.
- **Versioned models.** Every trained model gets a version tied to its code, data, and config. Without this, you can't roll back or diagnose drift.
- **Train/serve parity.** The feature engineering at training time must be identical at serving time. Drift between them is silent degradation.
- **Monitoring, not just metrics.** Track prediction distributions, input distributions, and live performance over time. A model degrades as the world changes.
- **Retraining trigger.** Define what triggers a retrain (schedule, drift threshold, metric drop), not just "when someone notices".