one thing that's particularly annoying is when an error occurs in `ci-gpu`. I can't run `ci-gpu` locally because I don't have nvidia-docker or a compatible GPU, and people on e.g. Windows will be in exactly the same boat. it's sometimes possible to quickly run the failing code locally in e.g. `ci-cpu` and reasonably expect it to fail in the same way. but you may or may not reproduce the error in doing that. in this case, you may wonder whether there is a discrepancy between Python deps that's responsible, but there's no good way to eliminate that variable.
I agree with your sentiment that strong consistency might not contribute to the overall health of the CI, but lack of consistency can make problems harder to debug when you don't have access to the full spectrum of CI variants. I think we should be judicious when we allow the CI to vary, and i'm not sure that e.g. allowing attrs to vary between containers buys us anything other than uncertainty. there are a couple of other problems that arise from a lack of standardized packages: - if you want to replicate the CI on some other linux system: which container's `pip freeze` should you use? right now, you have to be a CI expert to know the answer to that. it would be hard, as a newcomer to the project, to decipher which container tests which functionality, so you probably just guess and hope you picked right. - if we ever start relying on dev packages from `ci-gpu` which are also present in `ci-lint` (i.e. rerunning pylint for some reason), then discrepancies in those precise versions could cause late-stage CI failures - it makes it difficult to rebuild the CI containers if you want to change something other than the Python packages, since you can't avoid signing up to update them. --- [Visit Topic](https://discuss.tvm.apache.org/t/rfc-python-dependencies-in-tvm-ci-containers/9011/13) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.apache.org/email/unsubscribe/f9ed8b64b60e9e3efea973908448d1e7503b9aad003cad1b9978391ab186e9c7).
