Comparisons
Two tools, two questions. Neither column wins — what decides is which question you actually have in front of you.
Metrics vs TracesTracing vs ProfilingUtilization vs Saturationp50 (median) vs p99 (tail)Add capacity vs Fix the bottleneckLoad test vs Production telemetryCause-based alerts vs Symptom-based alertsMicrobenchmark vs End-to-end benchmark
p50 (median) vs p99 (tail)
The median describes the typical request; p99 describes the experience that generates support tickets. At scale, the tail is not an edge case — it is thousands of users an hour.
p50 (median)Open lesson →
Best for
Judging whether the common path is healthy and whether work got cheaper
Blind to
Everything that makes users angry
Costs
Stable and cheap; moves only when something systemic changes
Answers
"Did the work itself get faster?"
p99 (tail)Open lesson →
Best for
Judging whether the system is under stress and whether users are suffering
Blind to
Whether the typical request improved at all
Costs
Noisy at low traffic; needs volume and careful aggregation to mean anything
Answers
"Is anything queueing, contending or occasionally catastrophic?"
| Dimension | p50 (median) | p99 (tail) |
|---|---|---|
| Moves when | Service time changes | Waiting, contention or variance appears |
| At 10 req/s | Meaningful | Mostly noise — a handful of requests |
| At 10k req/s | Meaningful | Hundreds of users per second |
| Averaging across instances | Roughly acceptable | Mathematically invalid |