Build a Monitoring Dashboard Without Writing Queries

A small product needs a handful of panels, not a wall of graphs. Which numbers matter, why every chart needs a 'compared with before' and deploy markers, and how to build it without learning PromQL, NRQL or SQL.

A useful monitoring dashboard for a small product has about six to ten panels: a few headline numbers that show their change against the period before, p95 latency and error rate for the APIs your customers depend on, the conversion of your main journey, the routes failing most, and uptime or error budget. Put every chart on one shared time range, mark deploys and incidents on the time charts, and delete any panel nobody would act on.

Most dashboards fail in one of two ways. Either they never get built, because the tool wants PromQL, NRQL or SQL before it shows anything, or they get built once with forty panels and nobody opens them again. The fix for both is to decide what the dashboard must answer before choosing any chart.

Step 1: Write Down the Questions First

A dashboard should answer, in under a minute:

  1. Are customers getting through the thing they pay for?
  2. Is anything slower or failing more than usual?
  3. Did something change (a deploy, an incident) that explains it?

Every panel must help answer one of those. "Total requests since launch" answers none of them. Neither does a CPU graph for a server that has never been the problem.

Step 2: Headline Numbers With a Change vs Before

A number on its own means little. "1,284 orders" is good or bad only next to yesterday or last week. Show every headline number with its change against the same length of time before:

function change(now: number, before: number) {
  if (before === 0) return now === 0 ? '0 %' : 'new'
  const pct = ((now - before) / before) * 100
  return `${pct >= 0 ? '▲' : '▼'} ${Math.abs(pct).toFixed(0)} %`
}

change(1284, 1189) // "▲ 8 %"

Good headline numbers for a small product: visitors, sign-ups, orders or other paid actions, failed requests, open incidents. Four or five is enough.

Step 3: p95 Latency and Error Rate for Key APIs

Averages hide pain. If 95 requests take 100 ms and 5 take 6 seconds, the average looks fine while one in twenty customers waits. Chart the 95th percentile (p95) instead, and only for the APIs that matter: sign-in, checkout or order creation, and the main read your app does on every page.

Next to each, chart the error rate (5xx responses divided by requests), not the error count. A count rises with traffic; a rate rises only when something is wrong. For how to find which route is slow, see finding a slow API endpoint.

Step 4: Conversion of Your Main Journey

Infrastructure can look healthy while revenue falls. A form validation bug or a broken button returns no 5xx at all. One panel showing the conversion of your main journey (visit → sign up, or cart → order) catches what server metrics cannot. If it drops and your error rate does not, look at the frontend; silent frontend failures covers the usual causes.

Step 5: Top Failing Routes and Uptime

A ranked list beats a line chart for "where is it breaking": the five routes with the most failed requests, with the failure count next to each. Add one status panel for uptime monitors and, if you have promises (SLOs), how much error budget is left. An error budget turns "is 0.4 % errors bad?" into a yes or no. For outside checks, see knowing when your site is down.

Step 6: One Time Range, Deploy and Incident Markers

Two rules make a dashboard readable:

  • One time range for every panel. If one chart shows 24 hours and the next 7 days, you will misread both. Put a single time control at the top, with a "compare with the period before" switch.
  • Mark changes on time charts. Most regressions start at a deploy. A vertical mark at each deploy and each incident start answers "what changed?" without opening another tool. For deploy marks to work, every release needs a version your telemetry can see; with OpenTelemetry that is the standard service.version resource attribute:
// Set from your CI commit or release tag
resource: resourceFromAttributes({
  'service.name': 'api',
  'service.version': process.env.RELEASE ?? 'dev',
})

Step 7: Remove Vanity Panels

Go through each panel and ask: "If this moved, what would I do?" If the answer is nothing, delete it. Common ones to cut: lifetime totals, raw request counts without a rate, average latency, every host's CPU on a product that has never been CPU-bound, and duplicate charts of the same metric. A shorter dashboard gets opened; a long one gets ignored, like too many alerts.

Building It in Relia Without Queries

Relia builds dashboards by picking, not by writing queries. In the panel picker you choose What (a metric), For (a service, API, page or other thing Relia has seen, busiest first), Split by, and Show as, with a live preview. Panel types are charts, big numbers with the change vs before, top lists, status (monitors, Relia's checks, error budget left), open incidents and notes.

You can start blank, from a template (Traffic, Checkout health, Launch day, AI agents, Changes and incidents), or with "Copy and edit" on one of Relia's own service dashboards. One time bar sets the range for every panel, with compare and auto-refresh, and filters narrow the whole dashboard by service, environment, country or device.

Time charts carry deploy and incident marks, and a computed one-line context under the chart, such as "p95 is 57 % higher since the 15:30 deploy of api 1.6.1". An Ask box on the dashboard takes requests like "make it 7 days" or "add p95 of the checkout API"; Relia proposes checked changes, you apply them, then Save.

Dashboards can be team or private, and you can star favourites. TV mode and PNG / CSV export are on Growth and Pro. Free includes 2 dashboards, Growth 10 and Pro 30. Relia never changes your code.

FAQ

How many panels should a small product's dashboard have?

Six to ten. If you need more, split by audience (a product dashboard and a server dashboard) rather than making one long page.

Should I chart averages or percentiles for latency?

Percentiles. p95 shows what your slower customers experience; the average hides it. Keep the median (p50) only if you also want the typical case.

Do I need SQL or PromQL to build a good dashboard?

No. You need to know which questions it must answer and which few metrics answer them. A picker over a fixed set of metrics covers what a small product needs, and keeps every chart correct and scoped to your data.

All guides