Add ability to skip "dirty" commits during `dvc gc`

[courentin]

Following #3885, I'd like to add the ability to skip commits that generate errors during dvc gc.

To sum-up, running dvc gc in a codebase that contains a commit where dvc is in a dirty state (eg. the dvc.yaml is badly formatted, one of the output is dvcignored, etc.) make it failing.

Thus, if feel assuming that every commits are clean and let dvc in a usable state is a bit too optimistic. Except if you have a very rigorous process with dvc, it is almost impossible.

Thus, I'd suggest 2 things.
First adding a log to explain where (which branch/commit) the error has been raised.
Second, adding an option to gc: dvc gc --skip-failing-collect. Not sure of the naming but the behaviour would be to skip the collection of branches/commits that produce DvcException.

What do think?
I've started an implementation that I'll link to this issue.

[skshetry]

@courentin, I think --ignore-revs/--ignore-commits are more reliable than --skip-failing-collect (being a conditional flag) as it is clear on what your intentions are, considering gc is a dangerous command. What do you think?

[courentin]

Yes, explicit is better than implicit indeed!

[amka66]

I think this is not a feature request but rather a bug.

I would like to reduce the cache as it is eating up my hard drive, but i cannot do it due to one or two broken dvc.yaml files in my entire git commit history.
@efiop or anyone, please help, what can i do?

Unless there is a reasonable solution, I am stuck we a full hard drive. This requires a quick resolution. Thanks!

[WilliamHarvey97]

I had this issue. Nothing to add here, but this looks quiet serious to me because it makes dvc gc -A such a dangerous command.

[dberenbaum]

@courentin, I think --ignore-revs/--ignore-commits are more reliable than --skip-failing-collect (being a conditional flag) as it is clear on what your intentions are, considering gc is a dangerous command. What do you think?

In addition, we recently introduced -n to various dvc exp commands to specify the last n commits. The same option might be useful for gc to keep some limited history.

[daavoo]

In addition, we recently introduced -n to various dvc exp commands to specify the last n commits. The same option might be useful for gc to keep some limited history.

#7695