Currently our model deployment pipelines are composed of two stages:
- run-test
- run-production
The tests are simple formatting/syntax checks and the production stage builds the production image and if successful, pushes it to the WMF docker registry. One issue with this is that for each patchset in a CR, the image is pushed to the registry, which is confusing and wastes space/resources because after the CR is merged, the image gets built and published again.
We need to split the image publishing section to it's own stage that only runs during the gate-and-submit jobs.