> another approach would be to add an extra tag in the build that is not pushed anywhere, and use that for the smoketest run.
Yeah. However, it's good to have `set -e` anyway; in fact I'm surprised Azure Pipelines doesn't set that for you -- GitHub Actions, CircleCI, etc. all run something like `set -euo pipefail` automatically in bash steps to prevent silently ignoring errors.
> Rerun the failed build centos8_3.3.9, I'll ignore the nvidia ones and check once there is a new tag on hub.docker.com/r/nvidia/cuda/tags
`centos_3.3.9` failure appears to be a network hiccup: exit code 28 is `CURLE_OPERATION_TIMEDOUT`. `nvidia/cuda` have been updated again and according to [this comment](https://gitlab.com/nvidia/container-images/cuda/-/issues/71#note_364799365) it has been fixed for real this time.
I don't think I have the permissions to trigger a rerun though? You might have to do that yourself, or I can push again, but then everything is rebuilt.