Summary
A team that finds its delivery slow usually asks for faster pipelines. The article maps a whole CI/CD chain instead, from the code change to the deployment on each environment, with the three measures of a value stream map: the steps, the process time of each step and the wait time between steps, in average minutes. The integration pipeline took 14 minutes; a deployment to one environment took 802, 70 of them work.
On the CI side, polling, the nightly build and the webhook ran the same 14-minute pipeline. Their lead times, 34, 1,454 and 14 minutes, differed only by the wait before the first step: 20 minutes, 1,440 and none. The target keeps a single webhook and one pipeline per event: 12 minutes for a merge request, 14 for a push on develop, 5 for a feature branch, with no step made faster. On the CD side, six manual activities, repeated on every environment, made a lead time of 802 minutes, 732 of them spent in two six-hour queues and four three-minute waits.
Waiting dominates because every hand-off goes through a queue whose length depends on the availability of whoever empties it, not on the work, and because the queues sit between teams, where nobody owns them. The target gives three figures: 18 minutes on the development environment, 783 for a chain change, 378 per iteration. From 802 to 378, processing fell by 52 minutes and waiting by 372. The remaining 378 are still more than 95% waiting: the queue before the installation request, which automation did not touch because it is not an automation problem.
Key ideas
- A value stream map records three things: the steps, the process time of each and the wait between them, in average minutes; the lead time is the sum of the two times.
- Polling, nightly and webhook triggers ran the same 14-minute pipeline: their lead times of 34, 1,454 and 14 minutes came from the wait before the first step, not from any step.
- Of the 802 minutes needed to deploy to one environment, 732 were spent waiting, mostly in two six-hour queues, one at the qualification team and one at operations.
- The target cuts the lead time to 378 minutes per iteration mostly by removing a queue: processing fell by 52 minutes, waiting by 372.
- Queues sit between teams, where nobody owns them, so the map must be read by those who decide the organisation, not only by those who write the pipelines.