Why average latency hides Friday afternoon failures

Average response time is a comforting number. It is also frequently wrong for decisions about a single endpoint during a known busy window.

In retail and membership systems we review across Australia, Friday afternoons often combine customer traffic with inventory, reporting, or partner sync jobs. An endpoint that serves checkout or renewals may spend most of the week inside an acceptable band, then queue behind a shared resource for ninety minutes.

What we ask for first

• Percentile views for the same endpoint across the suspect window, not only the weekly mean
• A list of scheduled jobs that touch the same data stores
• Confirmation of whether retries amplify load when the first call slows

A practical reading habit

Compare the endpoint’s quiet-hour baseline with the overlapping job window before proposing a rewrite. Many stabilisations are scheduling and pool limits, not new architectures. Application analytics that ignore time-of-week patterns will keep approving averages that users do not experience.