The Scheduler

The scheduler decides which node each new Pod runs on. It works in two phases — filter the nodes that could host the Pod, then score the survivors and pick the best — and you steer it with affinity, taints, tolerations, and spread constraints.

Most of the time the scheduler just works and you never think about it. When you need replicas spread across zones, Pods kept off certain nodes, or workloads pinned to specialized hardware, these are the controls.

Filter, Then Score

Scheduling is two passes. Filtering eliminates nodes that cannot run the Pod — not enough requestable resources, a taint the Pod doesn't tolerate, a node selector that doesn't match. Scoring ranks the remaining feasible nodes by spreading, resource balance, and affinity preferences, and the Pod goes to the highest scorer. If filtering leaves no node, the Pod stays Pending — which, with the Cluster Autoscaler, is the signal to add a node (Topic 28).

Node Affinity and Selectors

nodeSelector is the simple form: run only on nodes with these labels. Node affinity is the expressive form, with requiredDuringScheduling (a hard filter) and preferredDuringScheduling (a soft preference that influences scoring). Use these to pin GPU workloads to GPU nodes, keep workloads in a region, or prefer cheaper node types when available.

Pod Affinity and Anti-Affinity

Where node affinity is about node labels, pod affinity is about other Pods. Pod affinity co-locates Pods (place this cache near that app); pod anti-affinity spreads them apart (never put two replicas of this database on the same node, or in the same zone). Anti-affinity is the classic tool for availability — but required anti-affinity at large scale is computationally expensive, so prefer the soft (preferred) form or topology spread for big deployments.

spec:
  affinity:
    podAntiAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        - labelSelector:
            matchLabels:
              app: web
          topologyKey: kubernetes.io/hostname

Taints, Tolerations, and Spread

Taints repel Pods from a node unless the Pod carries a matching toleration — the inverse of affinity. Control-plane nodes are tainted so ordinary workloads stay off; GPU nodes are often tainted so only GPU workloads land there. Separately, topology spread constraints are the modern, declarative way to distribute Pods evenly across zones or nodes, with a maxSkew that bounds imbalance — cleaner than hand-rolled anti-affinity for the common "spread my replicas" case.