Moved from weekly release trains to twelve promotions a day without adding headcount.

Everything shipped on Thursday. Around thirty changes from six teams were batched into a weekly release train, preceded by a coordination meeting nobody enjoyed. When the train broke, the batch made it genuinely hard to say which change was responsible, so the response was to roll back all of it and rebuild the batch for the following week. A one-line fix waited six days for the next departure.

Rollout became progressive and per-change: each promotion goes to a canary slice first and advances only while the Prometheus error-rate and latency queries stay inside their thresholds, rolling back on its own and paging through PagerDuty when they do not. Because a bad change reverts itself in minutes, batching stopped being the safety mechanism and the train had nothing left to do.

Helio promotes about twelve times a day, lead time from merge to production fell 84%, and the Thursday meeting no longer exists. The platform team is still three engineers — the cadence changed without headcount following it.