Rebuilding a deploy pipeline nobody trusted

The problem

Meridian shipped twice a week and dreaded both. Fifteen-minute builds, failures nobody could reproduce, and a rollback plan that was a person who knew which commit to revert.

Fig. 1 — the problem

What we found

The pipeline wasn't slow, it was repetitive. Three of the fifteen minutes were work; the rest was the same dependency tree resolved four times across four jobs that each started from nothing.

Fig. 2 — what we found

The fix

One cached install shared across jobs, type checking moved off the critical path, a preview URL per pull request, and rollback that promotes a previous build instead of reverting source.

Fig. 3 — the fix

Results