Evaluation behavior on Grafana Labs

How evaluation works

Mon, 10 Aug 2026 13:40:40 -0500

How evaluation works

The following diagram shows how an alert moves through states.

State	Condition is:
Normal	Not met
Pending	Met, pending period has not elapsed
Alerting	Met past the pending period. Notification sent.
Recovering	No longer met, keep firing for period active

For more details on alert states and how they function, refer to View the state and health of alert rules.

How checking works

During evaluation, Grafana checks your alert rule at regular intervals. Three settings control when the rule fires:

Evaluation group: Sets how often Grafana checks the alert rules in it. You create one (or reuse an existing one), and give it a name and an interval. That way related rules can share the same schedule, for example a group of related rules checked once a minute.
Pending period: How long the condition must persist before the alert fires.
Keep firing for: How long the alert stays firing after the condition resolves.

Pending period and keep firing for

Mon, 10 Aug 2026 13:40:40 -0500

Pending period and keep firing for

The pending period prevents false positives

Temporary data spikes can briefly cross the threshold without indicating a real problem. That means you don’t need an alert to fire. The pending period requires the condition to persist across multiple checks before firing.

For example, a 5m pending period with a 1m evaluation interval means the condition must stay above your threshold for five consecutive checks before the alert fires.

Data often arrives a little late, so the newest data point can be incomplete when Grafana runs a check. Because the pending period waits for several checks in a row, one reading based on partial data won’t fire the alert on its own.

Reduce alert fatigue through evaluation behavior

When a condition hovers near the threshold, the alert can fire and resolve repeatedly. Keep firing for holds the firing state briefly after the condition resolves, so you only receive a resolved notification when the problem is genuinely gone.

No-data and error handling

Mon, 10 Aug 2026 13:40:40 -0500

No-data and error handling

Each time an alert rule runs, it evaluates its query and decides whether to fire or stay quiet. Sometimes the rule runs but can’t decide whether to fire for two reasons:

The query returns no data. In this case, Grafana assigns a state of No Data. This happens when:
- No series match the query due to an incorrect metric or label filter
- The data source stops reporting
- The time window has no samples, often because the evaluation window or query range is too short for how often the data arrives
The query fails to run. If this occurs, Grafana assigns a state of Error. This can happen with:
- Evaluation timeouts
- Repeated query failures from a data source that’s down
- An authentication problem

Those auto-generated alerts use names like DatasourceNoData and DatasourceError, so your team may not recognize them or connect them back to your rule.

How to handle evaluation issues

You can choose other options to keep your rule out of the No Data and Error states. This means no separate alert is created, and the response stays tied to your rule’s own name and routing.

Alert state	Use this state when	What happens
Alerting	No data means something is broken, like a heartbeat check where silence equals an outage	Makes your own rule fire
Normal	Gaps are expected and harmless, like a metric that only reports during business hours	Keeps your rule quiet and healthy
Keep Last State	Gaps are brief and self-healing, and you want to avoid the rule flapping on a missed scrape	Holds whatever state your rule had before the gap
No Data/Error (default)	You want a separate signal that the data pipeline broke, routed apart from your rule	Leaves your rule quiet and raises a separate alert named `DatasourceNoData` (for No Data) or `DatasourceError` (for Error)

Common mistake

The Configure no data and error handling options in Grafana are powerful, but they require close attention. Don’t set the Alert state if no data or all values are null option to Normal unless you’re certain that data gaps are safe. Otherwise your rule stays silent during a real outage, and you hear nothing when the underlying data source stops reporting.