kube-scheduler

kube-scheduler watches pods whose spec.nodeName is empty, picks a node for each and writes it to the pod (binding). The selection has two phases:

1. Filtering: nodes that can't take the pod are eliminated (insufficient resources, taints, nodeSelector, affinity, port conflicts).
2. Scoring: the remaining ones are scored (balanced resource use, the image being on the node, spreading) and the highest-scoring is chosen.

$ kubectl get events --field-selector reason=Scheduled | tail -1
Normal  Scheduled  pod/api-7f8c9  Successfully assigned magaza/api-7f8c9 to worker-2

The pod spec gives the scheduler hints. podAntiAffinity spreads replicas of the same app across different nodes; topologySpreadConstraints balances replicas across zones:

    spec:
      affinity:
        podAntiAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            - labelSelector:
                matchLabels: {app: api}
              topologyKey: kubernetes.io/hostname
      topologySpreadConstraints:
        - maxSkew: 1
          topologyKey: topology.kubernetes.io/zone
          whenUnsatisfiable: ScheduleAnyway
          labelSelector:
            matchLabels: {app: api}