Metric release guard
Compare Launch rollout variants against Observer error rate, latency, monitors and anomalies.
How it works
- Opt in per flag to compare treatment with control for 30 minutes after a ramp.
- Error rate needs 200 exposures on both sides and at least five treatment errors. It must be 1.5 times control and at least one percentage point worse.
- Latency p95 needs 200 spans on both sides. It must be 1.3 times control and at least 100 ms worse.
- Small samples do not hold or kill a rollout. They read as not enough data yet.
- A finding holds an active rollout. Automatic kill still requires the separate per-flag opt-in.
- Errors are counted across the whole Observer project, so use a separate project per environment if staging and production share variant names.
- Monitor down and anomaly checks are not tied to a flag: any monitor going down in the workspace, or any anomaly alert in the project, during the window counts against the rollout.
Read the Launch overview or the SDK guides.