How many dashboards per service is sane?

Two. A triage dashboard (8–12 panels max, ordered by leading indicators and SLOs) and a deep-dive dashboard for specialists. Everything else is a report, not a dashboard.

What if we don’t have SLOs yet?

Start with rough thresholds pegged to hard limits (pool size, queue depth, GC pause). Then set SLOs based on historical performance and business impact. Wire burn-rate alerts once you know your error budget.

We’re not on Kubernetes—does this still apply?

Yes. The indicators don’t care: backlog growth, retries, connection saturation, and burn-rate translate to VMs and functions just fine. Swap HPA for ASG pending capacity and ELB 5xx for ingress 5xx.

How do we handle AI/LLM services?

Track model confidence, token/sec, cache hit rate, and cost per request. Alert on confidence dips with stable traffic and on sudden cost spikes. Gate prompt/weight rollouts with the same Argo/Flagger metrics you use for APIs.

What about traces—aren’t they leading indicators?

Traces are fantastic for attribution and pathologies, but they’re heavy as leading indicators. Use metrics to predict and traces to explain. Link directly from the panel to the representative trace sample.

Reliability-observability · Dec 3, 2025 · 9 minute read

Kill the Chart Zoo: Dashboards That Make Decisions in 60 Seconds

Stop shipping slide decks to on-call. Build instruments that predict incidents and trigger rollouts before customers feel pain.

Back to all posts

Kill the Chart Zoo: Dashboards That Make Decisions in 60 Seconds

Key takeaways

Implementation checklist