What API latency metrics actually matter

Averaging your API latency is the fastest way to miss the problem, because the average is dominated by the requests that were fine. Here is what to measure instead.

Percentiles, not averages

A p50 tells you the typical request is healthy, which is rarely the question. Watch p95 and p99, where customers actually feel the pain, and keep an eye on p99.9 for the worst case that support will hear about first.

Segment before you aggregate

• Per endpoint, so one slow route does not hide behind nine fast ones
• Per region, so a regional collapse is not averaged away globally
• Per client type, because mobile and web fail differently
• Per dependency, because a slow provider call is the usual cause

Measure success, not just speed

Latency without success rate is half a picture. A fast 200 wrapping a failed payment is still a failure, so track what the response means, not only how quickly it arrived.

Thresholds worth setting

• Alert when p95 on an endpoint breaches its budget for several minutes, not on a single spike
• Alert on a success-rate drop per endpoint
• Alert on a climbing retry rate, often the earliest sign a dependency is degrading