A repeatable test
A benchmark might measure how many images a service processes each second or how long a database takes to answer a question. The original measurements form a baseline: a starting point for comparison. Running the same test after a change shows whether performance improved or got worse.
A microbenchmark tests one small task. A broader benchmark tests a complete process, closer to what customers experience.
What the numbers leave out
Results can change with the device, amount of data, network, settings, or whether the system has recently handled similar work. A fast result for an artificial task may not mean a faster product. Useful benchmarks use realistic tasks, record the test conditions, and report more than just the fastest run.