p95

How it works

Evaluates tail distribution by isolating the slowest 5% of requests. This ensures that high-volume fast requests do not mask significant latency issues affecting a minority of users.

When to use

• Monitoring user-facing latency where worst-case experience is the priority.
• Detecting performance issues affecting specific request subsets.
• Enforcing performance SLAs.

When not to use

• Very small sample sizes where percentiles are mathematically unstable.
• Non-latency metrics where averages are more representative of total cost.

State explanations