1 min lesson
Taming heavy tails
Compare two rows from "Taming heavy tails", then say when each one fits.
Step 1 of 2
Taming heavy tailswhy the mean lies
- Technique
- Winsorize / cap
- What it does
- Clip extreme values to a high percentile before averaging
- The cost
- Hides genuine tail regressions if the cap is too aggressive
- Technique
- Log-transform
- What it does
- Compresses the tail so the mean is stable and tests are sensitive
- The cost
- Effect is now multiplicative; translate back for stakeholders
- Technique
- Percentile / quantile test
- What it does
- Test p95 latency directly, the number users actually feel
- The cost
- Quantile variance is higher; use bootstrap or quantile-regression SEs
| Technique | What it does | The cost |
|---|---|---|
| Winsorize / cap | Clip extreme values to a high percentile before averaging | Hides genuine tail regressions if the cap is too aggressive |
| Log-transform | Compresses the tail so the mean is stable and tests are sensitive | Effect is now multiplicative; translate back for stakeholders |
| Percentile / quantile test | Test p95 latency directly, the number users actually feel | Quantile variance is higher; use bootstrap or quantile-regression SEs |
Match the transform to the question: averages for cost, percentiles for felt experience.
Test the metric that matters, not the convenient one
A mean-latency test can read flat while p99 doubles, because the mean is anchored by millions of fast turns. If the reliability complaint is about the tail, power the test on the tail. Saying "I'd test p95 with bootstrapped CIs, not the mean" is the kind of specificity this role screens for.