Where the shortening came from
Previously the pipelines took 200-500 seconds, because every build started from scratch. Today a typical PHP code change goes through the pipeline in about 60-90 seconds as the target. Three decisions made the difference: layer caching of the image, the order of the layers and splitting the tests into two parts.
Dependency installation is tied to the lock files, and the source code goes in as the last layer. Hosts therefore download only the changed layers instead of the whole image of about 357 MB. The tests were split into two parts, because the speed of the runners differed by roughly 17 times.
What the shortening does not change
A faster pipeline does not mean a faster, less controlled deployment. Deployment to the host is a task started manually and cannot be interrupted. A new commit cancels an old, unfinished analysis pipeline, but it never stops a deployment that has already started halfway.
Once a day the first build on a runner runs without cache, with the base images refreshed. The shortening therefore applies to a typical change, and it does not replace the regular check of the build path from scratch.
Pipeline time before and after (midpoints of ranges)
Resource cost and the limits of speeding up
Parallel builds on a machine with 8 GB of memory led in September 2026 to the database process being killed by the OOM mechanism. The fix had three steps: cancelling outdated pipelines, running one build at a time and 5 GB of swap memory.
The conclusion for costs: speeding up does not mean adding parallel jobs to a weak machine. The resource cost grows more slowly when the pipeline uses the cache and cancels what is no longer needed.
Sample calculation of engineering time
We assume 400 changes in a year, an average pipeline time before the change of 350 seconds (midpoint of the 200-500 range) and after the change of 75 seconds (midpoint of the 60-90 range). The difference is 275 seconds per change, which is 110,000 seconds a year, or about 30.6 hours.
Assuming that an engineer waits for the result in half of the cases and returns to the task after switching context, an hourly cost of 150 PLN can be used. The direct saving is then about 4,590 PLN a year. A much bigger effect comes from a shorter reaction time to a fix, which we do not price in this calculation.
What it means for the client
For the client a shorter pipeline means faster delivery of a fix to pre-production and a shorter wait before the deployment, but the date on production is still set by the manual start. The plant chooses the moment, not the vendor's schedule.
When discussing the cost of changes, it is worth asking how much time passes from approving a fix to checking it on pre-production. A pipeline of 60-90 seconds removes this part of the delay.
Pipeline time before and after
| Parameter | Value |
|---|---|
| Changes per year | 400 changes |
| Pipeline before (midpoint of 200-500) | 350 s |
| Pipeline after (midpoint of 60-90) | 75 s |
| Cost of one engineer hour | 150 PLN |
Result: 275 s x 400 = 110,000 s, which is about 30.6 hours a year, about 4,590 PLN.
Note: Sample calculation based on the assumptions above; it does not include the cost of switching context.
Video
Frequently asked questions
Does a shorter pipeline shorten the time to deploy to production?
It shortens the time to pre-production. The deployment to production remains a manually started task, so the plant chooses the moment.
Where does the difference from 200-500 down to 60-90 seconds come from?
From layer caching of the image, the right order of the layers and splitting the tests into two parts.
Are parallel builds safe?
Not on a weak machine. The incident of September 2026 showed that with 8 GB of memory the database process can be the victim.
Is the calculation an offer?
No. It is a sample calculation with explicit assumptions. The price for a given plant is set individually.
Let's talk about the installation in your server room
We will present the requirements and the split of responsibilities for your plant based on the installation documentation. Book a conversation with the InnerWeb team.
Let's talk