You already know: containers and images (D11 Platform tools); probes, autoscaling and rollouts as concepts (D11 Model deployment, D14 Reliability); consensus and etcd (D14 Fundamentals); gang-scheduled multi-GPU training (D15); instance types and pricing (Cloud fundamentals).
Why this matters for Atlas. Kubernetes is the standard way to run containers across a fleet of machines: you declare what should run ("5 replicas of the tutor server, each with 1 GPU"), and controllers continuously make reality match. For ML it provides a shared pool of GPUs that training, inference, evaluation and pipelines can all draw from, with health checks, rollouts and autoscaling. It also has well-known gaps for ML (it schedules one pod at a time, so a 64-GPU job can wait forever), filled by add-ons like Kueue, Volcano and operators for Ray and PyTorch. This lesson builds the core ideas: the reconciliation loop, resource requests and bin-packing GPUs, and gang scheduling with queues.
1. Declarative state and the reconciliation loop
The question. What does Kubernetes actually do when you "apply" a Deployment?
Start from nothing. You submit desired state (YAML objects) to the API server, which stores it in etcd (a Raft-replicated store, D14). Controllers run endless loops: observe the actual state, compare with the desired state, take actions to reduce the difference, repeat. A Deployment controller makes sure the right number of pods (one or more containers scheduled together) exist; the scheduler assigns pending pods to nodes; the kubelet on each node starts containers and reports health. Crashes, node failures and manual deletions are all just "actual ≠ desired" and get corrected the same way. That's the property that makes Kubernetes resilient: no step-by-step scripts, only convergence.