A dashboard compresses thousands of observations into a few panels. Every query, aggregation, scale, and time window removes context. Read the dashboard as an argument to verify—not as a photograph of reality.
1. Read the Panel Definition First
Identify metric, unit, source, query, aggregation, filters, timezone, refresh interval, and delay. Confirm whether the panel shows a rate, total, gauge, percentile, or estimate.
- Does “CPU” mean one core, all cores, a container limit, or host capacity?
- Does “latency” include retries, failures, and client-side waiting?
- Is “errors” a count or percentage?
- Do “connections” include idle sessions, pool waiters, or only active queries?
2. Inspect Time Range and Resolution
A one-hour view can make a brief spike dominant; a 30-day view can average it away. Dashboards often increase query step for longer ranges, turning peaks into lower averages. Zoom in and out and note whether the latest bucket is incomplete.
Compare an equivalent baseline such as the same weekday and hour, not only the previous period. Keep deployment, failover, backup, and maintenance annotations visible.
3. Challenge the Axis and Encoding
A truncated axis exaggerates small changes; a wide axis hides important ones. Logarithmic scales show multiplicative change. Dual axes can imply a relationship between unrelated series.
- Show units on every axis.
- Prefer zero baselines for bars; clearly mark non-zero time-series ranges.
- Avoid 3D and area encodings that hinder comparison.
- Use consistent colors for the same state.
4. Question Aggregation
A fleet average can hide one failed node. A total can rise because traffic rose even when error rate improved. A maximum can be one noisy sample. A p95 without sample count can be unstable.
Review rate and volume, center and tail, global and segmented views. For Prometheus counters, graph rate() or increase() over an appropriate window rather than the raw cumulative value. Do not average precomputed percentiles.
5. Look for Missing and Surviving Data
No data is not zero. A host that stops exporting may vanish from an average and make the fleet look healthier. Fast failures can improve latency while error rate worsens. Client timeouts may be absent from server-side duration.
Check target health, newest sample time, missing series, error rate, traffic, and inventory coverage. Pair metrics with logs or traces and verify important claims against an independent source.
6. Use a Repeatable Reading Order
- Confirm scope, timezone, freshness, and filters.
- Check traffic and sample count before ratios or percentiles.
- Review errors and user impact.
- Inspect the latency distribution, not only the mean.
- Check saturation and queues.
- Segment by failure domain.
- Correlate with changes, then verify raw evidence.
A good dashboard accelerates investigation. It does not replace validation, context, or judgment.