Scheduling and placement

Stage 4 · Flight Control

The Kubernetes kube-scheduler assigns unscheduled Pods (spec.nodeName empty) to feasible cluster nodes.

Placement decisions follow a deterministic multi-stage algorithm rather than arbitrary round-robin assignment.

The three-phase scheduling lifecycle

Diagram RL-04 — the scheduler filters infeasible nodes before scoring feasible ones; binding writes nodeName.

1. Filtering (Predicates):
- Eliminates all nodes incapable of hosting the Pod (insufficient CPU/RAM, untolerated taints, missing node affinity labels).
2. Scoring (Priorities):
- Ranks remaining feasible nodes using algorithms such as NodeResourcesLeastAllocated or topology spread weights.
3. Binding:
- Atomically writes the winning node's name to spec.nodeName, triggering the node's local kubelet to pull images and boot containers.

A Pending Pod indicates placement constraints have not been met:

Symptom / Reason	Underlying Cause	Remediation
`Insufficient cpu/memory`	Total node requests exceed allocatable capacity	Add node capacity or scale down workloads
`Untolerated taint`	Node has taint (e.g. `dedicated=search:NoSchedule`)	Add matching toleration to Pod spec
`Node affinity mismatch`	Pod requires label not present on any node	Label the target nodes
`Unbound PVC`	Storage provisioner waiting on first consumer	Investigate PVC binding status

Taints & Tolerations:
- Taint on Node: Repels Pods (kubectl taint nodes node1 dedicated=db:NoSchedule).
- Toleration on Pod: Allows the Pod to schedule on tainted nodes.
Node Affinity:
- Attracts Pods to nodes with matching labels (kubernetes.io/os=linux).
- requiredDuringSchedulingIgnoredDuringExecution: Strict hard requirement.
- preferredDuringSchedulingIgnoredDuringExecution: Soft preference.
Topology Spread Constraints:
- Distributes replicas evenly across failure domains (racks, availability zones, hostnames) to prevent single-point-of-failure concentration (maxSkew: 1).

1. Scheduling failure events: Inspect why the scheduler rejected nodes:

kubectl describe pod <pod-name> -n apollo-airlines-apps | grep -A 10 Events:

2. Node allocation breakdown: Review allocatable resources per node:

kubectl describe node apollo11-worker | grep -A 10 "Allocated resources"

3. Active node taints:

kubectl get nodes -o custom-columns='NAME:.metadata.name,TAINTS:.spec.taints'