Set a Feature Freshness SLO
Define a practical freshness SLO for a feature used in online inference.
The ETA model uses a driver location feature whose stale values create bad arrival estimates while the model API remains technically available. Freshness SLO flow: decision harm -> SLI -> target and window -> alert threshold -> runbook response The common trap is monitoring only model endpoint uptime and assuming a successful feature read means the evidence is good enough. Before Dashboard: feature server 99.95 percent available. No explicit freshness target for driver locations. After Freshness SLO: 99 percent of active-trip location feature reads use events less than 120 seconds old over a rolling 30-minute window; page on rapid burn…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in