Design robust primary and guardrail metrics

For a newly launched feature, define a primary metric that best maps to long-term value and at least three guardrails. Specify exact formulas, units, numerator/denominator definitions, event deduplication rules, windowing (e.g., 7-day vs same-day), and handling of outliers (e.g., winsorization) and bots. Discuss ratio-of-sums vs sum-of-ratios trade-offs, sensitivity to backfilled/late events, timezone alignment, and how you would detect and correct a silent logging change mid-experiment. Compute power and MDE for a 14-day test given historical variance; explain sequential monitoring strategy and alpha spending. Propose canary thresholds for guardrails (e.g., crash rate, p95 latency, complaint rate), plus what to do when the primary improves but a guardrail slightly regresses.