← W31D1 hub · Alert Designer

Alert Designer

Pick a failure scenario, set your thresholds, see which rules fire. The goal: catch real incidents without paging on noise.

Pick a scenario

Your alert thresholds

Latency p95 threshold (ms): 500 Error-rate threshold (%): 2.0 Accuracy floor (%): 85 Drift PSI threshold: 0.20

Adjust thresholds to see how alerts behave.

Live readings

How to think about thresholds

→ Too tight = noisy alerts at 3am for normal variance. Get muted, miss the real one.
→ Too loose = silent failures while users churn. Worst of both worlds.
→ Add dwell time (e.g. "for 10 minutes") so a single spike doesn't page anyone.
→ Threshold per system, not per industry. A fraud model's 2% error rate is an emergency; a recsys's 2% is fine.