Latency percentiles on Vercel
Vercel latency percentiles come down to one for functions, P75, and only with Observability Plus on a Pro or Enterprise team. It has no P95 or P99 view and no alert on a latency threshold. To see P50, P95 and P99 per project and be alerted when they climb, read function durations from a Vercel drain, which Pro and Enterprise teams can set up. For what each percentile means, see P50 vs P95 vs P99 latency explained.
What Vercel shows on each tier
| Observability | Observability Plus | |
|---|---|---|
| Function latency | No P75, no breakdown by path | P75, sort by P75, breakdown by path and route |
| External APIs | Request totals per hostname | P75 duration, sort and filter |
| Data kept | 12 hours on Hobby, 1 day on Pro, 3 days on Enterprise | 30 days |
| Price | Included | Billed per million events |
Why P75 is not enough
P75 tells you about the typical request, not the slow tail. A function whose slowest 5% hang on a cold start or a slow database can keep a flat P75 while users wait seconds. In the P95 latency worked example, P75 and P90 look fine while P95 is more than three times the median.
What Vercel alerts on
Vercel alerts need Observability Plus. They fire on usage anomalies and on 5xx error spikes well above the 24-hour average, by email, Slack or webhook. There is no alert for a duration threshold, such as P95 above 1 second.
Which percentile to alert on
Alert on P95 duration per project, held for several minutes, next to the function error rate. Use P99 to find cold starts and slow dependencies, not to page.