cml ci: --fetch-depth

Backtracking on not doing this earlier - cml ci --fetch-depth=<int> is a good idea

depth=0 could be an alias for --unshallow
could thus deprecate (hide) --unshallow
things like depth=2 is useful for e.g. dvc metrics diff HEAD~1 CC @shcheklein @dberenbaum @pmrowla
cml ci --fetch-depth=2 is more trustworthy than manually running git fetch --depth=2 because the latter may not work as expected (?) if the repo was checked out weirdly (vis actions/checkout@v3 breaking changes)
cml ci --fetch-depth=2 is cross-CI compatible unlike actions/checkout.with.fetch-depth: 2

Apr 18 '22 16:04 casperdcl

btw just to confirm @dberenbaum does git fetch --depth=2 also obtain DVC exp's hidden reflog?

Apr 19 '22 10:04 casperdcl

Hm, good question. I think not since those refs will likely not be part of the linear history of the branch being fetched (plus it seems a bit DVC-specific). cc @pmrowla

Apr 19 '22 11:04 dberenbaum

btw just to confirm @dberenbaum does git fetch --depth=2 also obtain DVC exp's hidden reflog?

No, this just sets the depth of whatever refs have already been cloned/fetched to 2. So by default it will only include git branches and tags (and if actions/checkout uses --single-branch it would only include the PR branch)

Apr 19 '22 11:04 pmrowla

ok so I assume only git fetch --unshallow will work? Is there no other way? What if the repo is massive and a full unshallow is undesirable?

Apr 21 '22 14:04 casperdcl

The experiment refs aren't needed, so usually fetch depth of 2 or so on the cloned branch will be enough. They are needed to merge the head and the experiment changed afaik.

Apr 21 '22 16:04 dberenbaum

ok so I assume only git fetch --unshallow will work? Is there no other way? What if the repo is massive and a full unshallow is undesirable?

This won't fetch exp refs either, git commands only fetch branches/tags by default. You have to either be specifying the full exp refs with git fetch <options> refs/exps/... or using the dvc exp pull commands

Apr 22 '22 02:04 pmrowla

To be clear, fetching exp refs is unnecessary for many exp operations, like exp run. Having fetch depth is still useful for exp run.

Apr 22 '22 12:04 dberenbaum

Hello! I've been curious about CML for some time, and Hacktoberfest seems like a good opportunity to dive in. Can I take a stab at this issue?

Based on a quick look, I think this would involve modifying similar code as in https://github.com/iterative/cml/pull/957/files. @casperdcl Just to confirm, based on your description, should the fix deprecate --unshallow at the same time?

Oct 03 '22 14:10 deepyaman

thanks @deepyaman! Yes feel free to have a go. --unshallow can be soft-deprecated using .hidden :)

Oct 03 '22 14:10 casperdcl

Sorry for the delay. This is still very much on my list of things to do; got busy, but I hope to have a PR by later this week.

Oct 12 '22 07:10 deepyaman

no problem, looking forward to a PR @deepyaman :)

Oct 14 '22 15:10 casperdcl

Are there contribution guidelines? I came across https://cml.dev/doc/contributing/core, but I'd already done npm install and tried npm test, but it gives a lot of issues around tokens not being found, etc. I also assume the docs should require pip install tensorboard (I see it in the CI config, and am getting some related errors), but haven't gotten a chance to set that up yet.

~Alternatively, I think I can result any errors if the CI is allowed to run.~ I see it ran.

Oct 19 '22 15:10 deepyaman

Ah don't worry about testing locally - a lot of functionality needs actual CI to test effectively.

Don't worry if PR tests also fail, we can fix them :)

Oct 24 '22 23:10 casperdcl