Grafana puts an error budget on an agent's pass rate

The button the docs actually describe

Grafana's cloud documentation says that when the Grafana SLO app is installed, a team can track an evaluation's pass rate as a service level objective. The control appears on an evaluation that produces a pass or fail verdict, which means at least one evaluator has a pass condition. The person needs permission to create SLOs. They set a target pass rate. If the evaluation has at least 50 scores in the last 28 days, the product shows the current pass rate and prefills a recommended target. Otherwise the target is typed in. A burn-rate alert is optional. The software desk isSoftware.

What the budget is

The September 24 blog, which InfoQ picked up on October 16, uses a familiar reliability idea. A target of 95 percent fulfilled over a window leaves 5 percent that is not automatically a failure. That remainder is the budget: room for a bad day, an experiment, or a regression that has not yet exhausted the allowance. The blog's window in the example is 30 days. The docs' window for the provisioned SLO is 28 days. This page will not silently swap those numbers. The paper isFutureweb.

What InfoQ is, here

InfoQ's October 16 item is a news account that Grafana is applying SLOs and error budgets to agent behavior. The mechanism this brief is willing to describe comes from Grafana's docs and from Grafana's own blog. A news item can point at the feature. It does not replace the field names in the docs. The distributed-cloud desk is Asked directly, When does the product suggest a target? The docs say that if an evaluation has at least 50 scores in the last 28 days, Agent Observability shows the current pass rate and prefills a recommended target. Otherwise the target is entered by hand.Distributed Cloud.

The object the product creates

The docs say Agent Observability creates a provisioned SLO in the Grafana SLO app, measured over a rolling 28-day window, and labels it so it can be told apart from other SLOs. Deleting the evaluation deletes that SLO. A generated dashboard and the SLO definition can be opened from the evaluation. Provisioned SLOs can also be edited or deleted from the SLO app. Alerts follow Grafana notification policies. Those are product facts. They are not a claim that a score of pass means the answer was true. The artificial-intelligence desk isArtificial Intelligence.

What a pass rate does not settle

An evaluation passes or fails against a condition the team configured. The SLO then asks whether that pass rate stays above a target. A model can pass a rubric and still be wrong about the world. The budget tells an on-call engineer how much failure the team agreed to tolerate. It does not grade the rubric. An evaluation a second lab can run, which is a different question from an internal pass rate, isan evaluation a second lab can run.

Other behaviors, same app

The docs say SLOs for other agent behaviors, such as request success or latency, are created from Analytics, either by describing the objective or by starting from a template and building it with Grafana Assistant. That path is wider than the pass-rate button. This brief's center is the pass-rate path, because that is the one tied to an evaluation id and deleted with the evaluation. The security desk is Asked directly, Where do the alerts go? Through Grafana notification policies. The docs say the SLO is labeled so a policy can match it, and that extra labels can be added.Security.

Who is in scope

The docs describe a Grafana Cloud feature behind the SLO app and a permission check. A reader without that app does not get the button. The blog is aimed at someone who already has evaluations running and wants a target. InfoQ's reader is anyone who saw the October 16 headline. The three audiences are not the same, and the brief is about the documented control. Hardware is Asked directly, What gets deleted with the evaluation? The docs say deleting the evaluation also deletes its SLO.Hardware.

What this page will keep separate

A 95 percent illustration is not a recommended target for every agent. The recommended target, when the docs offer one, is prefilled from the evaluation's own recent scores, and only after 50 of them. A team that copies 95 percent out of a blog has left the document. Networking is The short answer this page will stand on is this: If the Grafana SLO app is installed, an evaluation that already yields pass or fail can get a target pass rate. The docs say the resulting SLO is measured on a rolling 28-day window. The gap under the target is the budget.Networking.

Where the dates sit

The argument was posted September 24. The wire item is October 16. The docs page this brief used does not carry a matching news date, so the behavior above is attributed to the documentation as fetched, not to a changelog line. If the button moves, the docs are the page to re-read. Asked directly, How long is the window in the docs? A rolling 28 days. The September blog's example of 95 percent over 30 days is an illustration, not the window the cloud docs print.