Remove all experiments in the workspace

[dberenbaum]

Have a dvc exp remove --workspace command or similar to drop all experiments in the workspace. Should be equivalent to dvc exp list --names-only | xargs dvc exp remove. This allows users to reset their experiments workspace if they have made mistakes or were playing around with some changes they ultimately want to roll back.

Some background:

• https://discord.com/channels/485586884165107732/485596304961962003/823901428652965898
• exp run: reset when the lock file is in Git #5553 (comment)
• Dropping experiments #5437 (comment)

[gdiepen]

This indeed would be a great functionality!

One possible (very likely) outcome of running experiment is that you find this path is not worth pursuing. In that case, would like to purge all of the experiments to be able to continue with the next idea

Also, the xargs solutions don't work that easy for me under windows ;)

[dberenbaum]

dvc exp remove -A/--all-commits might also be useful.

@pmrowla What do you think about moving up the priority of this one? The workaround is not difficult, but it's not that easy for people to find in this thread, and it seems like it's a straightforward feature to implement.

[pmrowla]

I guess we just need to be clear that by workspace here we mean all experiments based on the current HEAD right?

Regarding priority, I think it's up to you. This is an issue that anyone should be able to fix - codebase-wise, it won't actually require touching anything on the experiments internals side of things.

[dberenbaum]

I guess we just need to be clear that by workspace here we mean all experiments based on the current HEAD right?

Right, it should match dvc exp show, which is described in https://dvc.org/doc/command-reference/exp/show#description:

Only the experiments derived from the Git HEAD are shown by default

This is an issue that anyone should be able to fix - codebase-wise, it won't actually require touching anything on the experiments internals side of things.

Yes, this could be a good opportunity for either one of the other devs or contributors.

[casperdcl]

I find the dvc exp gc flags a little confusing. They're mostly "remove all except" (exclusive) rather than "remove only" (inclusive). Ideally we should have both. Related #6006.

I'd prefer a new flag which inverts the meaning:

  -w, --workspace     Keep experiments derived from the current workspace.
...
  --queued            Keep queued experiments (experiments run queue will be
                      cleared by default).
  -R, --remove/--no-keep   Inverts meaning of the above options
                           (remove rather than keep).

[iesahin]

Could we support --keep and --remove options, that can take

workspace
all-branches
all-tags
all-commits
<specific-experiment-id>

as parameters.

So dvc exp gc --keep workspace will be identical with the current dvc exp gc --workspace and dvc exp gc --remove exp-abc123 will remove that.

We can also use dvc exp remove for this. gc can just be a flag to remove, that basically says "run dvc gc after removing these to clean up the cache" So ınstead of dvc exp gc --workspace && dvc gc --workspace, the user can run dvc exp remove --keep workspace --gc.

@efiop @dberenbaum @pmrowla @karajan1001

[dberenbaum]

I'd prefer a new flag which inverts the meaning:

Is this the same functionality as the proposed dvc exp remove -w? Is the point that you would prefer dvc exp gc -w -R to dvc exp remove -w? I agree that gc can be confusing, which is why I suggested to add this option to remove instead.

I'd say that introducing gc -R or gc --keep/--remove is a separate feature request since it does more than what's requested here, and gc applies to more than experiments. For example, would dvc exp gc -R/--remove ever make sense for all branches or tags? I think the ideas make sense, but I don't want them to block the simpler feature request here.

If the --workspace option seems awkward in dvc exp remove, we can simply implement dvc exp remove --all and save the distinctions between --workspace and similar options to another feature request.

[iesahin]

Is this the same functionality as the proposed dvc exp remove -w? Is the point that you would prefer dvc exp gc -w -R to dvc exp remove -w? I agree that gc can be confusing, which is why I suggested to add this option to remove instead.

IMHO, the flags should be explicit. Having a --reverse flag is confusing. The user shouldn't have to look up the help text about the meaning. What are we reversing to what?

I'd say that introducing gc -R or gc --keep/--remove is a separate feature request since it does more than what's requested here, and gc applies to more than experiments. For example, would dvc exp gc -R/--remove ever make sense for all branches or tags? I think the ideas make sense, but I don't want them to block the simpler feature request here.

Implementation-wise, these are not that different. Basically, the command will iterate over all experiments and decide whether an experiment is within the group specified for deletion. If some of the groups are empty, so be it.

If the --workspace option seems awkward in dvc exp remove, we can simply implement dvc exp remove --all and save the distinctions between --workspace and similar options to another feature request.

--all is better than --workspace. --workspace is a smaller set than --all.

BTW if dvc exp remove will be the primary removal command, having --include and --exclude options may be better semantically. dvc exp remove --include all --exclude exp-123abc says "delete all experiments except exp-123abc", but this is another feature as you say.

@dberenbaum