DEV Community

Muhtalip Dede
Muhtalip Dede

Posted on

Stop paging on every kubelet Event: Incidents + confidence gates

#kubernetes #ai #devops #sre

Always-on Kubernetes watch only helps if you correlate into Incidents and gate by severity and confidence — not one Slack message per BackOff.

Originally published at https://kprompt.ai/blog/observe-agent-alert-fatigue.

Wiring kubectl get events into Slack feels productive for about a day. Then CrashLoop storms reopen the same fingerprint every few seconds, ImagePull retries spam the channel, and nobody trusts the bot. Always-on watch is only useful if the product artifact is a correlated Incident — not a raw Event dump.

The failure mode: Event → page

  • One message per kubelet Event → thread death under restart storms
  • LLM called on every raw API event → cost and latency spikes with no extra truth
  • No durable “open incident” → you cannot ask why later or mark false positive
  • No severity/confidence gate → low-signal noise trains the team to ignore alerts

Observe pipeline (gated)

kprompt’s optional Observe agent watches Pods/Events (and optional workloads) in one namespace, correlates into durable Incidents, optionally analyzes with your BYOK LLM, then notifies Discord, Slack, or a webhook only after a severity + confidence gate. Default Observe never applies, patches, or deletes (ADR-0013).

kprompt agent run -n payments \
  --emit-initial --analyze --fetch-logs --health \
  --min-severity medium --min-confidence 0.7
Enter fullscreen mode Exit fullscreen mode

Levers that cut fatigue

Lever What it does
--heuristic Zero token spend; deterministic detectors for demos and many day-2 signals
--min-severity / --min-confidence Defaults medium / 0.7 — raise both to cut noise and LLM burn
Incident batching One analysis per evidence fingerprint, not per raw Event
--memory / --patterns Namespace facts + “seen before” bias confidence — never sole proof of root cause
Slack threads Prefer bot + channel update-in-thread over a new webhook post per pulse

Rough expectation from the ops runbook: busy namespace with LLM on and gate at medium/0.7 → a handful of completions per real incident, not hundreds per hour. If spend spikes, check the gate and whether CrashLoop storms reopen fingerprints.

Memory and patterns are bias, not proof

Namespace memory and pattern learning can boost confidence when a signature looks familiar, and Slack false-positive feedback can dampen repeat noise. They must not invent root cause alone. If Prom, OTel, or GitOps evidence is missing, Observe degrades honestly — it does not fabricate metrics. More on the memory bet: cluster memory.

Autopilot stays propose-only

Optional --autopilot-propose emits PlanResult-shaped remediations with Applied false. Silent heal is not the goal. Apply remains policy + allowlist + explicit approve under ADR-0015. See also Autonomous not yet.

Try it without paging anyone 

git clone https://github.com/kprompt/kprompt-examples.git
cd kprompt-examples
make walkthrough
Enter fullscreen mode Exit fullscreen mode

Heuristic mode needs no API key. For a Slack path, add --slack (or Discord webhook) only after the gate settings match your tolerance. Walkthrough details: Break a kind cluster on purpose. Experimental — prefer kind / non-prod first.

Try: kprompt.ai · GitHub · brew install kprompt/tap/kprompt
Muhtalip

Top comments (0)