Field note · 12 Jun 2026

When average build time hides a bad afternoon

Why a smooth average can conceal the compile spike that actually missed your cut.

Release packs love a single average. It fits on a slide. It rarely survives the afternoon when three large compile jobs land on the same agents.

What the average erases

An average build duration mixes queue wait, clean compiles, warm-cache hits, and the one branch that rebuilds half the world. If your cut decision happens on a Thursday afternoon, you need the distribution for Thursday afternoons—not the month’s mean.

A practical substitute

Keep the average if finance wants a headline. For the release checklist, add:

  • Median duration for the last twenty builds on the release branch
  • 90th percentile for the same window
  • Share of builds that crossed your “too late to cut” threshold

Those three numbers still fit on one page of a ledger.

How we use this in audits

During a build & compile analytics audit, we ask for the raw samples before we accept any averaged slide. Teams that already track percentiles move faster; teams that only track means spend the first day reconstructing them.

← Back to field notes