Introduction
A trained model is just a file until it's serving predictions to users or other systems. Deployment is about two questions:
- How will predictions be produced and delivered — all at once in a batch, or one request at a time?
- How do we replace an old model with a new one without breaking things?
Think of a restaurant changing its menu. You wouldn't throw out the old menu and serve untested dishes to every customer on day one. You'd test new dishes in the kitchen, offer them to a few tables first, and keep the old dishes ready in case something goes wrong.