dvc: checkout: consistency in handling files that are missing version info

dvc checkout(and dvc pull since it uses it internally) will error-out if version-info is missing, while dvc push will just warn us, which creates a confusing inconsistency. For example, it makes CML guys use || true for dvc pull https://github.com/DavidGOrtega/cml-dvc-test/runs/2438634469 CC @DavidGOrtega , as they don’t have dvc.lock on initial run.

Current dvc push behaviour with warnings is also quite annoying and doesn’t scale well, as it might get way to noisy for a big pipeline. So we should probably agree on one common behaviour for all such cases that will make sense to our users. CC @dberenbaum

About this issue

Original URL
State: closed
Created 3 years ago
Comments: 21 (20 by maintainers)

Most upvoted comments

related #4746

pmrowla on Jul 6, 2021

@karajan1001 This is the scenario as far as I remember:

dvc add data and commit .dvc file.
dvc stage add and commit dvc.yaml (the stage does not get run).
git push and dvc push, which warns the user that version info is missing for the outputs of the dvc.yaml stage.
On remote server in CML, git pull and dvc pull, which throws an error because version info is missing for the outputs of the dvc.yaml stage.

EDIT: A much simpler summary is that no dvc.lock file exists.

dberenbaum on May 19, 2021

I don’t have a good idea on why it deletes the file on missing version info. DVC behaves like this in a lot of places, which I’d like to fix (eg: dvc add removing stage files on failure, etc). Pinging @efiop.

skshetry on Jun 23, 2021

Current dvc push behaviour with warnings is also quite annoying and doesn’t scale well, as it might get way to noisy for a big pipeline.

Isn’t this what --quiet is for? It would be great to have some nice summary option that would be a compromise between no output and overly verbose output, but I think that’s separate from the issue here, which is about an inconsistency between push and pull. If I have done dvc stage add ... && dvc push, then dvc pull right after should not fail.

dberenbaum on May 3, 2021