Introduction
Everything in this track works on one laptop — until it doesn't. Datasets grow to terabytes, models to billions of parameters, and traffic to thousands of predictions per second. At that point you need scalable ML systems: spreading the work across many machines and GPUs, and organising data so many models can use it consistently.
Think of moving house: one person with a car can manage a flat, but a whole office needs a team, several trucks and a plan for who carries what.
Two main things to scale:
- Training — too much data or too large a model for one device.
- Serving & data — too many requests, and features that must be identical in training and production.