ML Engineering and MLOps
Everything it takes to get a model out of a notebook and keep it working in the real world: data pipelines, serving, hardware, monitoring and experiments.
Alsomachine learning operations · ML in production
Training a model is like baking one great cake. MLOps is running the bakery: the ingredients arrive on time, every cake comes out right, and you notice quickly when customers stop liking them.
Where it fits
Not shown: 6 more kinds and parts. See the full list below.
+All relationships as a list
Overview
Most of the effort in real ML systems goes into everything around the model. The main pieces:
- Data and features: reliable pipelines and feature engineering, so training and production see the same inputs.
- Serving: model serving turns a trained model into a fast, reliable service, trading off latency, throughput and cost.
- Hardware and efficiency: GPUs and other accelerators do the heavy math; quantization shrinks models so they are cheaper and fit on phones.
- Monitoring: the world changes after deployment, and data drift silently erodes accuracy.
- Safe rollouts: A/B tests check that a new model actually helps real users before everyone gets it.
Everyday use cases
- Case 01
Keeping recommendations fresh
Retraining on new data regularly and checking each new model before it reaches users.
- Case 02
Fast responses at scale
Serving a model to millions of users within milliseconds and at reasonable cost.