# Metrics

Numeric time series sent over OTLP: request rates, latency, queue depth, and any custom measurement you record.

## Selecting a series

Pick a metric name, then narrow by its labels. Keep the time range close to the question you are asking: a wide window shows trend, a narrow one shows an incident.

## What to record

- **Rate**: how much work is arriving.
- **Errors**: how much of it is failing.
- **Duration**: how long it takes.
- **Saturation**: how full the resource is.

Label cardinality matters. Labels with unbounded values, such as user ids, request ids or raw URLs, create a separate series per value and quickly become unusable.

Metric values are also what a metric-threshold alert rule reads. See [Alerts](/alerts).