Latency, Throughput, Bandwidth, and Percentiles

Your website's average response time is 200 milliseconds. The dashboard looks healthy, but some customers wait several seconds. You upgrade the network connection, yet the complaints continue. Which measurement tells you what is actually slow?

Latency measures delay across a defined boundary. For a web request, you might measure from sending the request to receiving the complete response. Network round-trip time measures a different boundary; it does not include all application work. Throughput measures how much work completes per second, such as 500 requests per second. Bandwidth describes data-carrying capacity, such as a 1 Gbit/s connection. Actual transfer throughput can be lower. More bandwidth will not remove a slow database query. Percentiles show the distribution of measured delays. p50 is the median. If p95 is 800 milliseconds, approximately 95% of the measured requests completed within that time; the slowest 5% took longer. p99 examines further into that slow tail, but is not the maximum. Always state the operation, measurement boundary, and time window. These observations describe the measured requests, not a guarantee for the next customer.