oxygen914/kubeoncall/node-runtime-pressure-triage

Node Runtime Pressure Triage

Diagnostic objective

Separate clock synchronization failure and descriptor exhaustion from monitoring error, transient load, process leaks, and host-wide runtime degradation.

Required live evidence

For clock drift, collect offset trend, synchronization state, source reachability, recent suspend/reboot, and peer-node offsets. For descriptors, collect allocated/max trend, per-process descriptor growth, socket/file classes, workload placement, error logs, and recent releases. Always confirm the alert with an independent current signal.

Decision branches

• Clock offset rises while peers remain synchronized: inspect the node time service, upstream reachability, virtualization clock, and recent reboot history.
• Descriptor use grows with one process and open-file errors: identify the leaking workload and correlate growth with traffic or a release.
• Metric is high but resource state and peer signals disagree: classify as monitoring uncertainty and keep the conclusion unverified.

Stop conditions

Do not recommend a restart, process kill, time step, limit change, or Pod eviction automatically. Stop at an evidence-backed cause, affected scope, confidence, and the bound versioned runbook.