CSE Roadmap

Core tier · about 3 hours · 8 parts

Part 6 of 8 · 1 min read

Good chart habits

Layout rules that make an anomaly easy to spot.

UDS lists practices. They are grouped below.

Keep dashboards in code. Define charts in a domain-specific language and keep them in version control. Then the pull request that changes the code can change the dashboard too, and staging and production stay in step without error-prone manual edits.

Make charts comparable. - Put the most important charts at the top, since dashboards render top to bottom. - Use one default time zone, such as UTC, so people in different countries can talk about the same moment. - Give every chart in a dashboard the same resolution and range, so anomalies line up by eye. For an incident, one hour at one-minute resolution suits. For capacity planning, one year at one-day resolution.

Keep charts readable. - Few metrics and few points per chart. Too many slow the page and hide the anomaly. - Put metrics with similar ranges together. A metric with a huge range flattens the others. Split the percentiles: for example, the 10th percentile, average and 90th in one chart, and the 0.1th, 99.9th, minimum and maximum in another.

Annotate. Give each chart a short note that links to the runbook, nearby dashboards and who to call next. Draw a horizontal line at each alert threshold and a vertical line at each deployment.

Avoid gaps. A metric emitted only when an error happens leaves wide empty stretches, and the reader wonders if the service stopped reporting. Emit 0 when there is no error and 1 when there is.

Check yourself: 1. Why use the same time range on all charts of a dashboard? 2. Why mark deployments on a chart? 3. How does emitting 0 avoid confusion?