01 / MEASUREMENT

There is more than one kind of delay.

A number needs a boundary. A round-trip measurement describes a packet travelling to a destination and back. A completed HTTPS request includes other work: name resolution, connection establishment, a security handshake, server processing and response transfer.

Name the interval.

Record what starts the clock and what stops it. “Time to first response byte” and “time to download the whole file” answer different questions. Neither should quietly replace the other in a comparison.

Keep the quiet failures.

A median describes the middle of successful measurements. It says nothing about requests that did not finish. Record failures and deadlines beside the timing distribution. A low median with frequent timeouts is not a reliable service.

Compare like with like.

Use the same client, network and destination. Alternate the old and new conditions where practical instead of taking all measurements for one condition first. A busy wireless channel or a background download can otherwise look like a server regression.

The useful result is often a narrower question: not “the network is slow”, but “new connections are slow while reused connections are steady”. That is a much better place to begin.