intermediate · ai
MLOps: models that survive production
Everything between a notebook that works and a model that keeps working with real users on it.
·
By Amit Chakraborty — Kolkata, India · Remote worldwide
What you will be able to do
Version, deploy, monitor and roll back a model.
Syllabus — 12 chapters
- 1. Why models fail in production and never in the notebook — Name the gap between an experiment and a running system.
- 2. Data versioning — Reproduce a training run from six months ago.
- 3. Features: stores, and the simpler thing that usually works — Avoid training and serving skew without adopting a platform you do not need.
- 4. Reproducible training pipelines — Turn a notebook into something that runs unattended.
- 5. Experiment tracking and honest comparison — Compare two runs without fooling yourself.
- 6. Registry, packaging and deployment targets — Move a chosen model from a laptop to somewhere it can be called.
- 7. Serving: batch, real time and the latency budget — Choose a serving mode from the requirement rather than the fashion.
- 8. Monitoring: drift, skew and knowing before the users do — Detect a model that has quietly stopped working.
- 9. Evaluation: offline metrics, online metrics and the gap — Understand why the offline number improved and the product did not.
- 10. CI for models — Write automated checks that block a bad model from shipping.
- 11. Cost: hardware, batching and quantisation — Cut serving cost without silently cutting quality.
- 12. Rollback and shadow deployment — Ship a model with a defined way back.
0 of 12 published. A new chapter every Thursday night. Free to read.
Learn it free, elsewhere too
The documentation, videos, books and practice sites worth your time on this subject. All free. None of it is mine.