Don't Panic: Grafana Fundamentals

Grafana is the bit between a pile of operational data and the person trying to decide what to do about it. It does not collect the metrics, logs, traces, database records, or cloud-service data. It asks the systems that do collect them a question, then turns their answer into something a human can scan before the coffee gets cold.

The important boundary is pleasantly unforgiving. A data source is the backend that holds the data and understands its own query language. Grafana sends a query to that backend and renders the result. Put Prometheus, Loki, a SQL database, or a cloud service behind the same dashboard if that helps, but they have not become one big friendly database merely because they are sitting next to each other.

A panel is one question plus one way of showing the answer. A time series is for change over time. A stat is for a current aggregate. A table, logs view, or trace view is for detail when the overview has pointed at something suspicious. A dashboard collects those panels for one audience and one decision. The sensible reading order is current state first, diagnosis later. This saves readers from starting at the engine room when they only needed to know whether the ship was moving.

The surprising part is that a handsome panel has no special relationship with truth. A chart can have excellent colors, a dramatic slope, and precisely the wrong query. Units, time range, labels, and scope are part of the claim. If they are unclear, the panel is not cautious. It is ambiguous with good typography.

Variables let one dashboard select a service, environment, cluster, or region instead of multiplying copies. They help when the question stays the same. They do not rescue a question that was unclear before the dropdown arrived. Keep their values narrow enough that a reader can understand what changed.

Alert rules are a separate promise. They combine queries, a condition, an evaluation interval, a pending duration, labels, annotations, routing, and an explicit decision about no data or errors. A red panel can wait for investigation. An alert asks somebody to respond. If nobody can name the first action, it belongs in the dashboard for now.

Read the Intro for the full working model and where Grafana stops. Use Slides for the request path and the dashboard decisions. Keep the Cheatsheet nearby when building or reviewing panels. The practice reference and exercise turn the model into a small TestData dashboard, where no production service can complain about the experiment.