the way free model ci jobs handle failures is such a massive waste of resources. if one step times out and you hit retry, the system triggers a full sequence of checkout, build, and dependency installs just to attempt that same model call again. it feels like
re-running the whole world for a tiny error. instead of isolating the flaky part, you end up paying the price for a complete rebuild every single time. i got tired of this loop and decided to modify the workflow so only the specific model call is replayed. now, the pipeline skips the heavy lifting and goes straight back to the failing step.
it basically turns a massive headache into a simple one-step retry . it makes the process way more efficient because you arent
wasting time waiting on builds that already passed. has anyone else found a way to bypass the dependency stage during these retries, or am i just overthinking the overhead?
found this here:
https://dev.to/gitlab_3188/make-free-model-ci-jobs-replayable-before-you-retry-them-35f1