Skip to main content
Metrics answer how much, how often, how slow — cheaply, and over windows that would be impractical to ask of logs or traces.
Before you start. You need metrics arriving from at least one source: an OpenTelemetry SDK, a Collector, or Prometheus remote-write.

​
Finding a metric

The metric catalogue lists every metric name seen in your organization, discovered from what has arrived rather than declared in advance. A metric appears the first time a sample is written and stays listed while data remains inside retention. Metric names arrive in two shapes and both work:
Nothing is normalised between them. If the same measurement reaches you from both a Prometheus server and an OpenTelemetry SDK, you will see two metrics with two names and two sets of labels. Pick one path per measurement rather than reconciling them afterwards.

​
Charting

Select a metric, filter on its labels, and choose an aggregation. The chart respects the time-range picker, which is the setting that most often explains a surprising result.

​
Resolution changes with the range

A wide range is answered from pre-aggregated data. Metrics are continuously rolled up to one-minute and one-hour resolution, and a query reads whichever fits:
  • Recent, narrow ranges → full resolution
  • Longer ranges → the one-minute rollup
  • Months → the one-hour rollup
So a 20-second spike is visible on a 15-minute chart and averaged away on a 30-day one. It is not missing — it is summarised. No amount of zooming a wide range recovers it; narrow the range instead.
There is an opposite effect on very fresh data. A sample written seconds ago is in the raw table before the rollups covering it are complete, so it can be visible on “last 15 minutes” and not yet on “last 7 days”. Wait a minute rather than widening the range. See Data retention.

​
How long metrics are kept

Metrics outlive logs and traces by a year, which is what makes them the right signal for capacity planning and trend analysis.

​
Cardinality

Every distinct combination of label values is a separate stored series, and that — not the number of samples — is what drives cost.
A label carrying a user id, request id or raw URL path turns one metric into millions of series. The effect is silent: charts get slower, ingest grows, and nothing reports an error.Per-request values belong on spans and log records. See cardinality.

​
Metrics you get without instrumenting anything

The last is worth knowing about: RED metrics are derived from your traces, so instrumenting for traces gives you the service-level metrics for free.

​
Querying from your own tools

Metrics are readable through the API, and there is a Prometheus-compatible query API so Grafana can use aiAxonIQ as a datasource without an adapter — including its limits, which are worth reading before you rely on it. See Grafana and Prometheus clients.

​
Alerting on a metric

A threshold rule compares an aggregate against a bound on a fixed 60-second loop. Two things catch people out — only three comparators exist, and a rule saved with an unimplemented one is created without complaint and never fires. See Alerting.

​
Next

Build a dashboard

Panels, templates, versions and shareable snapshots.

Set an SLO

Error budgets and burn-rate alerting.

Forecast and anomalies

Where the AI features read metrics.