Required affinity is a filter; preferred affinity is a score.

Use it when a Pod needs, or would like, nodes with certain labels: SSD disks, a zone, an instance family, a GPU type.

nodeAffinity

affinity:
  nodeAffinity:
    requiredDuringSchedulingIgnoredDuringExecution:  # HARD
      nodeSelectorTerms:
      - matchExpressions:
        - {key: disktype, operator: In, values: [ssd]}
    preferredDuringSchedulingIgnoredDuringExecution: # SOFT
    - weight: 80
      preference:
        matchExpressions:
        - {key: topology.kubernetes.io/zone,
           operator: In, values: [europe-west1-b]}

required...

Acts like a filter. No matching node means the Pod stays Pending.

preferred...

Acts like a score (weight 1 to 100). The Pod still schedules elsewhere if needed.

...IgnoredDuringExecution

Checked only at scheduling time. Changing a node label later does not evict a running Pod.

Operators

In, NotIn, Exists, DoesNotExist, Gt, Lt. Use Exists when only the label's presence matters.

Node affinity: Pods choosing their nodes

Node affinity lets a Pod express which nodes it can or prefers to run on, based on node labels. The simplest form is nodeSelector, an exact label match; node affinity generalises it with operators and a soft form.

The long field names encode two decisions. requiredDuringScheduling is a hard rule that works like an extra filter: if no node matches, the Pod stays Pending. preferredDuringScheduling is a soft rule that works like an extra score with a weight from 1 to 100: the scheduler favours matching nodes but will use others if it must. The suffix IgnoredDuringExecution means the rule is evaluated only when scheduling. If the node's labels change afterwards, a running Pod is not evicted.

Common built-in labels are kubernetes.io/hostname, topology.kubernetes.io/zone and region, node.kubernetes.io/instance-type and kubernetes.io/arch. Cloud platforms add their own: GKE and EKS label nodes with node pool or node group names and capacity types such as Spot, which you can use to steer workloads to cheaper capacity.

Gotcha: required affinity combined with too few matching nodes is a frequent cause of mysteriously Pending Pods. Read the FailedScheduling event, which reports node affinity mismatches explicitly.