Kubernetes scheduling explainer

Filters — for every node, each filter in the scheduler’s order: NodeUnschedulable (cordoned), NodeName, TaintToleration (NoSchedule and NoExecute; tolerationSeconds noted), NodeAffinity (nodeSelector and required terms with In/NotIn/Exists/DoesNotExist/Gt/Lt and matchFields), NodePorts, NodeResourcesFit (requests — including sidecar and init containers, pod-level resources and RuntimeClass overhead — against allocatable minus what bound pods request, and the pod count), PodTopologySpread (skew per domain, minDomains, nodeAffinityPolicy, nodeTaintsPolicy, matchLabelKeys) and InterPodAffinity (required affinity and anti-affinity, and other pods' anti-affinity against this one). The outlined cell is the first failure — the one the scheduler reports.

Scores — for nodes that pass: TaintToleration ×3 (PreferNoSchedule taints), NodeAffinity ×2 (preferred terms), PodTopologySpread ×2 (ScheduleAnyway constraints, or the built-in spreading over hostname and zone when the pod has none and a Service or controller selects it), InterPodAffinity ×2 (preferred terms both ways, and others' required affinity), NodeResourcesFit ×1 (least allocated, counting 100m/200Mi for containers without requests), NodeResourcesBalancedAllocation ×1 (since v1.36: how much the pod improves the node’s CPU/memory balance) and ImageLocality ×1 (status.images on the node).

When nothing fits: the FailedScheduling message exactly as kubectl describe pod shows it, and preemption — which nodes could make room by evicting lower-priority pods, which pods would go (most important spared first, PodDisruptionBudgets respected where possible), and why that node was chosen (fewest PDB violations, lowest victim priority, smallest priority sum, fewest victims, latest start time). preemptionPolicy: Never is honoured.

Replicas — each replica placed in turn, seeing the ones before it, so you can watch spreading, anti-affinity and resource exhaustion play out. Click a row for its tables.