Field note · 3 May 2026

Compile cache hit rates that lie politely

A cache that claims success while still recompiling the expensive units is worse than an honest miss.

A high cache hit rate feels like proof of virtue. It is only proof that something was reused. It does not prove that the expensive compile units were among the reused set.

A quick honesty check

Pick ten builds on a busy branch. For each, note:

  1. Reported cache hit percentage
  2. Wall-clock time in the heaviest compile stage
  3. Whether that stage’s inputs actually changed

If hit rate stays high while the heavy stage barely moves, your metric is congratulating the wrong work.

What to record instead

Pair hit rate with heavy-stage duration and a short note on input change. The pair is harder to game and easier to defend in a release meeting.

Workshop use

Our release metric workshop often retires standalone hit-rate slides in favour of this paired reading. Participants grumble; cut week gets quieter.

← Back to field notes