move label_to_label.yml to an issue?

[zou3519]

Today there are rules for adding labels if a label exists already. These are specified in label_to_label.yml. Would it be possible to read this from an issue instead of a checked-in yaml? Reading from the issue has the following benefits:

• it lets us modify the file without waiting for code review. This is useful to test label configurations (if some of them generate too much noise, then I want to turn them off).
• it is more discoverable

cc @ZainRizvi @kit1980 @huydhn @clee2000

[ZainRizvi]

Hey @zou3519, thinking about this feature idea I see the following pros and cons:

Pros:

• Speed of adding new labels (which has been needed once in the past month?)

Cons:

• Higher risk of fat fingering
• Security concerns (who has access to edit the issue unilaterally?)

Ambiguous:

• I'm not sure how issues would be more discoverable vs a yaml file. Both would need some pointer to discover, except yaml files might get discovered via vscode or github's global code search.

Overall, I'm not sure if this is a good idea to support especially since this is something that seems like it would rarely be updated.

Personally, I prefer the idea of using issues for changes where:

• Speed is critical (e.g. every minute without a change causes CI failures for someone)
• CI job update and read the issue (basically a poor man's database)

Admittedly, historically we may have been a bit too liberal with using issues as per the above ideas

I'm open to different opinions though

[zou3519]

I need to update the labels. It is likely that I need to modify them 5-10 times, because if some labels result in too noisy notifications I will remove them from the list. I haven't updated the labels I'm waiting for this issue to be resolved :P.

More seriously -- #24422 gets edited 6 times a month. What's the frequency for which we'd make a cutoff for features like this?

[malfet]

Nothing at PyTorch is someone elses problem, so @zou3519 if it's really blocking you, just submit a PR against test infra and ask for a review.

I wasn't even aware we have this label-to-label workflow, which probably causes system to sometimes drop labels, as bot and workflow might attempt to do it concurrently.

IMO we should create an umbrella issue/wiki somewhere, that outlines how we want to approach triage automation, as right now we have 4 distinct approaches:

• auto-label bot
• https://github.com/pytorch/pytorch/blob/main/.github/label_to_label.yml
• https://github.com/pytorch/pytorch/blob/main/.github/labeler.yml
• GitHub project automation