Why are monitoring metrics stored in a time-series database instead of a normal relational database?

Metrics are append-only, timestamped and queried by time range, and a TSDB compresses and scans exactly that workload far better than a relational store.

A monitoring system writes the same handful of metric names millions of times a day, never updates a past value, and asks questions of the form "CPU for these hosts over the last six hours, averaged per minute". That workload is nothing like a business database: values arrive in timestamp order and change slowly, so a TSDB stores them with delta and compression schemes that shrink them enormously, and indexes by time and labels so a range scan is cheap. Downsampling old data (per-second detail for a day, per-hour averages for a year) is built in, which is how a year of history stays affordable.

Go deeper:

• Prometheus — Storage — how a real TSDB lays out blocks, compresses samples and retains data.
• Time series database — Wikipedia — why the workload is different and the main products.