"Three replicas" says nothing about where they run, so tell the scheduler to spread them.

Without a rule all three can land on one node; one node failure then takes the whole service down while kubectl get pods still said 3/3.

topologySpreadConstraints

topologySpreadConstraints:
- maxSkew: 1
  topologyKey: topology.kubernetes.io/zone
  whenUnsatisfiable: DoNotSchedule   # or ScheduleAnyway
  labelSelector: {matchLabels: {app: checkout}}

Prefer spread constraints for balance

Anti-affinity is a pairwise "avoid that Pod" rule and can leave zones lopsided. Spread constraints reason about the whole distribution, which is nearly always the real goal. Use both when you need hard co-location rules and balance.

Spread replicas across failure domains

Replica count guarantees that N Pods exist, not where they run. Without a placement rule the scheduler may put several, or all, replicas on one node. The service looks perfectly healthy in kubectl get pods right up until that node fails, and then everything goes down at once. Spreading replicas across failure domains is what turns replicas into actual redundancy.

Pod anti-affinity says: do not schedule me on a node or zone that already runs a Pod matching this selector. It prevents exact co-location but says nothing about balance, so you can still end up with all replicas in one zone. Topology spread constraints directly express the goal: keep the difference between the most and least loaded zone within maxSkew. With three replicas and maxSkew 1 across three zones the only valid layout is one per zone. Reveal the three panels with the right arrow.

whenUnsatisfiable mirrors the hard and soft split used elsewhere: DoNotSchedule is a filter, ScheduleAnyway is a score. For critical services use zone spread with DoNotSchedule, accepting that a Pod may wait for capacity in the right zone, and combine with a PodDisruptionBudget.

On GKE regional clusters and EKS multi-AZ node groups, the zone label is automatically set on nodes. Remember storage: a zonal disk pins a Pod to its zone, so spread and volume topology must agree (Module 7).