The simplest way to steer a Pod: label the nodes, then ask for a label.

Label the nodes

Chapter 2 labelled a node once. Nodes are cluster-scoped objects and take labels like Pods do:

kubectl label node peaks-node-2 disk=ssd
kubectl label node peaks-node-3 disk=hdd gpu=true
kubectl get nodes --show-labels
kubectl get nodes -l disk=ssd

Some labels are already there. kubernetes.io/hostname is the node's name, kubernetes.io/os and kubernetes.io/arch describe the machine. The rest are yours: zone, disk, gpu, whatever the fleet needs.

Ask for a label

spec:
  nodeSelector:
    disk: ssd
  containers:
    - name: db
      image: postgres:16

Every key in nodeSelector must match a label on the node, with that exact value. Extra labels on the node are fine. A node without the label is filtered out.

When nothing matches

The Pod stays Pending and the scheduler says so:

0/3 nodes are available: 3 node(s) didn't match Pod's node affinity/selector.

Nothing is wrong with the Pod. Label a node and it schedules a moment later; the scheduler retries Pending Pods whenever the cluster changes.

kubectl label node peaks-node-2 disk=nvme --overwrite

Where it fits

A nodeSelector goes in the Pod spec, so for a Deployment it lives under spec.template.spec. Every replica then obeys it, and a rollout moves them if you change it.