The average response may hide the request everybody is waiting for.

Many quick operations can coexist with a few slow ones. Whether that matters depends on the work those slow operations carry.

A report can wait for its final dependency. A person can experience a single delayed transaction rather than the average across thousands. An intermittent failure can interrupt a consequential sequence despite reassuring routine behavior.

Keep the distribution in view.

Follow an important request.

What happens when reuse misses? When several dependencies slow together? When renewal overlaps with demand? When a returned result needs additional examination?

Those are questions to investigate in the actual system. They are not explanations of a failure merely because they are plausible.

The historical distributed-grid case involved intermittent degradation and eventual crashes. The surviving account establishes the nature of the challenge and Howard's contribution; it does not establish a particular root cause for readers to copy into another system.

Pair a number with its population.

An average, percentile, or maximum observation needs its measurement window, request population, and definition of the measured interval. What happened before or after the interval can matter to the usable result.

A system may improve its measured operation while leaving the user waiting on validation, retries, or another dependency. See the critical path.

Make the observation actionable.

The purpose of examining the tail is to understand which conditions affect useful work. A graph is valuable when someone can connect it with an operating question, a design choice, or a reproducible investigation.

Read a useful baseline, operating responsibilities, and diagnostic experience.

The request that matters to your next finish deserves more than an average impression.