40 mindown to 6 minDeploy queue, Okapi Freight
“We cut our deploy queue from 40 minutes to 6. Nobody batches PRs on a Friday afternoon anymore, because there's nothing left to wait for.”
The flaky-test quarantine paid for the seats in the first month. We'd been retrying the same four tests since 2023.
I stopped writing YAML. That's the review.
Support answered a question about cache eviction at 11pm on a Sunday. With a diff.
−$7,400 / monthCI spend
We ran our own runner fleet: 40 spot instances and a cron job whose only purpose was killing zombie containers. Migrating took two afternoons. Remote cache hits went from 31% to 88% in the first week.
Our monorepo has 214 packages. Lockstep works out which nine actually changed, every time, and only builds those.
Rollbacks are one command and they're boring. Boring was the whole goal.
We rolled it out to 60 engineers without writing a migration doc. The defaults were right.
Preview environments on every PR changed how design review works here. Designers comment on the real build, not a screenshot from Tuesday.
18m 40s → 4m 12sp95 build time
Same hardware budget. I've stopped asking for more runners in quarterly planning.
The audit log is the first one our SOC 2 auditor didn't have follow-up questions about. Every deploy has an approver, a diff and a reason.
Our runbook used to have a section called 'babysit the release train'. I deleted it in March and nobody noticed.
Unedited, from teams that moved their pipelines over this year. We asked for numbers; most sent some.
Read all 212 case notes