The pipeline is the slowest part of the day
A twenty-minute build turns review into batching, batching turns into bigger changes, and bigger changes turn into riskier deploys. Meanwhile the deploy itself is a button nobody fully understands, and the answer to "what is in production right now" involves opening three tabs.
What the work involves
- Measure first: where the time actually goes, per stage, before touching anything. an illustrative pipeline, measured stage by stage →
- Caching that survives a cold runner, dependency layers that do not invalidate on every commit, and test parallelism that is worth its coordination cost.
- Deployments that can be reversed in one step, and that record what was deployed, by whom and from which commit.
- GitOps delivery with Argo CD or Flux where it fits — so the repository remains an accurate description of production rather than an aspiration.
- Pipeline cost as a first-class number: runner minutes are a line on the bill and are usually the cheapest thing on it to halve.
How the work goes
Measure
A week of pipeline telemetry: stage durations, cache hit rates, queue time, failure reasons. Most teams are surprised by which stage is actually the problem.
Cut
The changes with the best ratio of minutes saved to risk introduced, applied in order, each one measured against the same baseline.
Make it reversible
Rollback rehearsed, not assumed. A deploy you have never reversed is a deploy you cannot reverse under pressure.
Is this the right call?
A good fit when
- Teams whose build time has crept past the point of patience
- Anyone who deploys by hand because the automation is not trusted
Not us when
- Replacing your CI vendor as an end in itself — we will usually make the current one faster instead