Grafana Adaptive Metrics: Cost-First vs Coverage-First Auto-Apply

An AI agent that queries your metrics on your behalf can quietly drive up your observability bill, because every query for data that was aggregated away can trigger its restoration at full resolution. Grafana’s answer, now generally available in Adaptive Metrics, is a cost-first auto-apply mode that guarantees active series never rise above where each cycle started. For platform teams who own the Grafana Cloud bill, it changes auto-apply from a coverage tool into a budget control.

How It Works

Adaptive Metrics, part of Grafana Cloud’s Adaptive Telemetry suite, watches how metrics are actually used and recommends how much detail to keep for each one. Each recommendation carries a projected change in series count, up or down. As Grafana’s Felix von Platen explains in the announcement post, auto-apply launched last year with only one behavior, coverage-first; teams wanting anything else had to review recommendations by hand. With general availability, a second mode arrives, and both run continuously with no human intervention while dashboards, alerts and queries keep working unchanged.

Coverage-first applies every recommendation as soon as it is generated. If someone queries data that was aggregated away, the recommendation to restore full resolution applies immediately. Metrics stay in lockstep with usage, but a spike in queries against aggregated data becomes a spike in series count.

Cost-first works in three steps. Recommendations that reduce series always apply. Recommendations that increase series are ranked from least to most impactful. Adaptive Metrics then works down that list and stops when the next one would push projected series above the cycle’s starting total.

Grafana’s own simplified example shows the difference over one cycle:

Under cost-first, the series count stays flat or drops and never rises above where the cycle started.

What This Means for You

Grafana is explicit about where each mode belongs. Coverage-first remains the better fit for business-critical data owned by teams who know exactly what they query and why, and it is the mode Grafana Labs uses for its own critical telemetry. Cost-first suits data where predictability matters more than guaranteed coverage. Grafana’s examples are a PromptQL query that asks for more than its author realized, or an agent issuing queries with no visibility into downstream cost, a pattern it calls increasingly common as less experienced users get access to larger datasets.

Teams don’t have to choose one mode for everything. The segments feature partitions telemetry by a chosen dimension, such as team, environment or cluster, and each segment can run its own mode. Both modes are available now to existing Adaptive Metrics users with no extra setup; Grafana recommends creating exemptions first for anything you can’t afford to lose, and enabling auto-apply requires the Admin role. Cost discipline on telemetry is a recurring theme in this vertical, as Stack Archive covered in how Datadog saves $1 million a month on AI.

Source: Grafana Labs Blog

Content created with AI assistance and reviewed for accuracy.