Pod Crashes & Events¶

Three monitors cover pod-level health.

pod_crashes¶

Fires critical whenever a container's waiting/terminated reason matches the configured list.

Default reasons: CrashLoopBackOff, OOMKilled, ImagePullBackOff, ErrImagePull, CreateContainerConfigError, FailedScheduling, FailedMount, Evicted.

monitors:
  pod_crashes:
    enabled: true
    reasons:
      - CrashLoopBackOff
      - OOMKilled
      - ImagePullBackOff
      - ErrImagePull
      - CreateContainerConfigError
      - FailedScheduling
      - FailedMount
      - Evicted
    include_recent_logs: true   # attach last 50 log lines (when supported by the notifier)
    max_log_lines: 50

Trigger it on purpose:

kubectl run oom --image=polinux/stress --restart=Never \
  --requests='memory=64Mi' --limits='memory=64Mi' -- \
  stress --vm 1 --vm-bytes 100M --vm-hang 30

The pod should OOMKill within seconds and kpulse should fire.

pod_restarts¶

Fires warning when a container restarts more than threshold times within window.

monitors:
  pod_restarts:
    enabled: true
    threshold: 5
    window: 15m

kpulse keeps a sliding window of restart counts per container. Useful for "this pod is flapping" cases that don't trip CrashLoopBackOff (e.g. liveness probe restarts).

warning_events¶

Fires info for every Warning-type Event whose reason is not in reasons_ignore. The "everything else weird" catch-all.

monitors:
  warning_events:
    enabled: true
    reasons_ignore:
      - FailedGracefulShutdown
      - Unhealthy
      - Failed                # owned by pod_crashes
      - BackOff               # owned by pod_crashes
      - BackoffLimitExceeded  # owned by job_failed
      - InvalidDiskCapacity   # k3d / Docker Desktop kubelet quirk

The defaults strip the most common noise:

Probe flaps: FailedGracefulShutdown, Unhealthy fire constantly during normal probe transitions.
Already covered: Failed and BackOff are emitted by the kubelet for the same crashes that pod_crashes already alerts on; BackoffLimitExceeded duplicates job_failed. Including them here avoids two alerts per real failure.
k3d quirk: InvalidDiskCapacity fires on k3d/Docker Desktop nodes because the kubelet can't stat the host-backed image filesystem. Not actionable.

If you see other recurring noise, add the reason here.