Latency and percentiles

How to read it

With 5 parallel calls, the probability that at least one lands on the p99 is 4.9% — your effective p99 behaves like a p95. And a 3-call serial chain starts at ~360 ms median, but its realistic worst case already carries one full p99: ~1,040 ms.

Assumptions and limits

• Assumes independent latencies across calls: real correlation (same saturated host) makes it worse.
• The serial approximation uses (K−1)·p50 + p99 as a realistic worst case, not the absolute worst case.
• Does not model queues or retries: combine it with the retry budget.

The article behind it

P50, P95 and P99: percentiles without self-deception · Latency in distributed systems