สร้างแดชบอร์ด Grafana

Prometheus stores your metrics, but raw numbers are hard to read. Grafana turns them into charts you can actually understand at a glance. 📊

Grafana does not store data itself. It queries a data source like Prometheus and draws whatever that source returns.

First step in Grafana is connecting Prometheus as a data source. You give it the Prometheus server URL and test the connection.

A dashboard is a screen of panels. Each panel is one chart driven by one query, so you compose a full view from small pieces.

Panels ask Prometheus questions in PromQL, its query language. This is where you turn stored metrics into a meaningful line.

The rate function shows per-second change of a counter over a window. It is the classic way to plot traffic to your model.

Divide error rate by total rate to plot an error ratio. A rising line here means more requests are failing.

From your latency histogram, compute the 99th percentile so you see the slow tail, not just the average.

For single numbers like current error rate, a stat panel shows one large value and can flash red past a threshold. ⚠️

Add a dashboard variable for model version, then one dashboard works for every model instead of copying it per version.

Every dashboard can be exported to JSON. Commit that file to Git so your dashboards are versioned and reproducible like the rest of your stack.

You want a Grafana panel showing per-second request traffic to your model from a counter. Which PromQL fits?

You connected Grafana to Prometheus, built panels with PromQL for rate, errors, and p99 latency, and saved the dashboard as JSON in Git. ✅