THE CLUSTER

Real: the three control loops. The Horizontal Pod Autoscaler sets each app's replicas to ceil(current × CPU utilisation ÷ 60%), with Kubernetes' ±10% tolerance, its scale-up limit (double, or four more) and a scale-down window. The scheduler filters out nodes that are not ready, tainted (the control plane), full, or short of CPU or memory for the pod's request, then scores the rest — least allocated, balanced CPU/memory, and spreading an app across nodes — and binds to the highest; the panel shows the scores it compared. The Cluster Autoscaler adds a node when pods stay unschedulable and removes one when the cluster is mostly idle and its pods fit elsewhere. A rolling update surges 25% new pods and keeps zero unavailable.

Simulated: the traffic and the CPU it costs, the three apps (invented), and time — a node that stops responding is evicted after 3 seconds here instead of Kubernetes' 5 minutes, and an idle node goes after 6 seconds instead of 10 minutes. Drag to turn it, pinch or scroll to zoom, double-tap to reset, tap a node to pick the one to unplug.