Generative AI on Kubernetes
Eight chapters, drawn and narrated, for the platform engineers who now have to run language models: how a model server is deployed, where the weights live, what the inference metrics actually measure, how GPUs become schedulable, and what production asks for.
After the book by Roland Huß and Daniele Zonca — Operationalizing Large Language Models (O'Reilly). An original animated retelling — every picture, every question and every word of narration is written here. It follows the structure of the authors’ book and is not a copy of it, not endorsed by them, and not a substitute for reading it.
Written for people who already run Kubernetes. It assumes pods, nodes, controllers and YAML, and explains everything about language models from scratch.
8 of 8 chapters ready · 2 hours · 230 scenes · 63 questions
Getting a Model onto the Cluster
Everything before the first token: why this is hard, the server that answers, and how the weights reach the pod.
- Introduction — 9 min, 19 scenes, 7 questions
- Deploying Models — 15 min, 28 scenes, 8 questions
- Model Data — 17 min, 31 scenes, 8 questions
Running Inference Well
What only matters once it runs: measuring the inference, making GPUs schedulable, and staying up under load.
- Model Observability — 15 min, 26 scenes, 8 questions
- Kubernetes and GPUs — 16 min, 32 scenes, 8 questions
- Running In Production — 16 min, 32 scenes, 8 questions
Tuning Models and Building on Them
Leaving serving behind: changing the model itself, and the applications built on top of it.
- Model Customization — 13 min, 24 scenes, 8 questions
- AI-Driven Applications — 18 min, 38 scenes, 8 questions
Before you start
- The source is an O’Reilly early release — its fifth, of 21 November 2025. Two chapters it lists in its own contents, on job-scheduling optimization and on running agentic applications in production, were not in that release and so are not here either.
- Its errata were checked rather than copied: 62 verified errors in the source are recorded in the build notes, and this telling teaches the corrected version. Where the source’s own cross-references are broken, chapters are named by title here and never by number.
- The narration is synthesised speech, not a human reading. Captions are a keystroke away in the player.