How Scheduling Decisions Happen

When a new Pod is created without a node already assigned, the scheduler (introduced back in the foundations module) is the component responsible for picking one, and it does that in two clear stages. First, filtering: it eliminates every node that flatly cannot run this Pod - not enough free CPU or memory to satisfy its resource requests, a required port already in use, a volume that is only available in a different zone, a label selector or affinity rule that does not match, or a taint the Pod does not tolerate (taints are covered in their own section below). Second, scoring: among the nodes that survive filtering, it ranks them by preference - spreading load, respecting soft affinity preferences, and similar heuristics - and binds the Pod to the winner. This matters practically because a Pod stuck in Pending is a symptom, not a diagnosis: the scheduler always emits an event explaining exactly which predicate(s) it failed, and reading that event is almost always faster than guessing.

Analogy: Kubernetes is a thermostat, not a remote control. You declare the temperature; independent controllers measure reality and keep acting. Debugging means finding which sensor, rule or actuator prevents convergence.

Read the object as evidence

kubectl get RESOURCE NAME -n NAMESPACE -o yaml
kubectl describe RESOURCE NAME -n NAMESPACE
kubectl get events -n NAMESPACE --sort-by=.metadata.creationTimestamp
kubectl explain RESOURCE.spec

Do not memorize these as a ritual. The first command exposes desired and observed state, the second connects conditions and events, the third supplies a timeline, and the fourth checks the server's schema. Compare what the controller was asked to do with what it reports doing. Then test the narrowest hypothesis.

Scenario: A Pod declares a resource request of 4 CPU cores, and the cluster's total free capacity across all its nodes adds up to 10 CPU cores - so it looks like there is plenty of room. But every individual node in that fleet only has 3 CPU cores free, and a single Pod's request is indivisible - it must fit entirely on one node, it cannot be split across two. Aggregate cluster capacity is a misleading number for scheduling; the real question is always whether any single node satisfies every constraint the Pod has, one node at a time.

Read FailedScheduling literally

Use kubectl describe pod and group scheduler reasons: insufficient requested resources, untolerated taints, selector or affinity mismatch, volume topology, host ports, or quota. Capacity is multidimensional and per node; totals can mislead. The scheduler uses requests, not current usage, so a quiet but fully requested node is unavailable to new guaranteed intent.

Scenario: Eight nodes each have 500m CPU allocatable, yet a Pod requesting 2 CPU remains Pending. Four aggregate cores cannot be combined for one Pod. Resize the request, use larger nodes or split the work.
Tip: Compare kubectl top node with kubectl describe node allocated requests. Usage answers current load; allocation answers scheduling commitment.

Production reasoning

Ask four questions: Who owns this object? What dependency must become ready next? Which controller reports the blocking condition? What evidence would disprove my current theory? This prevents symptom-driven changes. Record the context, namespace, object generation, image digest and recent rollout before mutation; a recreated Pod may erase the evidence you needed.

Warning: Running is not the same as ready, healthy, durable or correct. Kubernetes status is layered. Confirm the application-level outcome as well as the object state.
Goal: Put this model into practice in the Kubernetes lab resource-requests-oom. Open /labs/kubernetes, choose resource-requests-oom, predict the failure path before changing anything, then use the simulator's check command to validate the finished state.

Deliberate practice

Before the lab, write the expected object relationship and the first three commands you will run. Afterward, explain why the fix converged and name one tempting change that would only mask the symptom. Repeat using an explicit namespace and a structured output format. This prediction-observation-explanation loop is what turns command familiarity into production judgment.

45-minute investigation

  1. Map (5 min): draw the owner-to-child chain and mark every namespace, selector, identity and dependency involved.
  2. Predict (5 min): write one expected status condition, one likely event and one log or metric signal before opening the lab.
  3. Observe (10 min): collect YAML, describe output and ordered events. Do not mutate state. Record which observation disproves your first theory.
  4. Repair (15 min): make the smallest declarative correction, watch the responsible controller converge, and verify the user-facing path rather than stopping at Running.
  5. Stress (10 min): change one relevant constraint - replica count, label, readiness, resource value or placement rule - predict the outcome, observe it, then restore the known-good declaration.
Tip: Keep a short incident note with symptom, evidence, hypothesis, change, result. Across four sections this produces a reusable runbook instead of a pile of remembered commands.