provide guideline or mechanism to resolve conflicts in DVC files on merge

[shcheklein]

When we change something (let's say a stage command) in one branch (experiment) and want to propagate this change to the derived branch it's easy to get a merge conflict in a DVC file.

Usually merge strategy theirs (git merge -X theirs) do the job.

We need to think if it's always the case. We should specify a guideline how to resolve these conflicts. If there are some ideas how to avoid them, let's discuss this. Right now it's frustrating and unclear how to manage these conflicts.

[dashohoxha]

When we change something (let's say a stage command) in one branch (experiment) and want to propagate this change to the derived branch it's easy to get a merge conflict in a DVC file.

For small changes, I think that cherry-pick would be better and safer than a merge.

From what I have seen (and I have zero ML experience, I have just read some DVC tutorials), we merge a branch A onto a branch B when we want to completely overwrite B with A, for example when we merge the branch with the best metrics onto branch master. In this case it is safe to use the merge strategy theirs. Then, after making as well a dvc checkout, all the data and pipelines will be up to date (or at least most of them).

[jorgeorpinel]

I think this is done! See https://dvc.org/doc/user-guide/merge-conflicts

Closing for now.