how to: "undo" 'dvc add' + other "how to undo mistakes" recipes

[gwerbin]

This issue is 1 of 3 that I was asked to post here, following a conversation on Discord with mroutis and ivan

I encountered this problem within my first 15 minutes of using DVC: "how do I undo dvc add?"

The answer turns out to be:

1. unprotect the file, just in case
2. Delete its corresponding .dvc file
3. Optionally run gc or look up the MD5 in data.dvc and delete it from .dvc/cache by hand.

It would be nice if this recipe were included somewhere in the docs.

Part of the reason this procedure was so non-obvious is that I never thought to look in https://dvc.org/doc/commands-reference/add. This turns out to have a complete technical explanation of linking, and what "protect" means in DVC. This document should be referenced early and often, maybe with a sentence like "For a technical explanation of the caching and linking process, see here".

In the meantime, hopefully my Stackoverflow self-answer will help other users, as well as get the DVC ball rolling on Stackoverflow.

DVC, like Git, admits the possibility of potentially-destructive beginner mistakes. I think it would be good practice to try and anticipate and provide recipes for undoing mistakes, and link to them prominently in the getting started documentation.

[dashohoxha]

@gwerbin There is an example of how to undo dvc add here: https://katacoda.com/dvc/courses/basics/data-management
However this is still unfinished work and not yet advertised on the website of DVC.

[dashohoxha]

In the meantime, hopefully my Stackoverflow self-answer will help other users, as well as get the DVC ball rolling on Stackoverflow.

There is also a stackexchange about DS: https://datascience.stackexchange.com

[Suor]

I feel that there should be an easier way)

[shcheklein]

@Suor do you meant that there is an easier way already, or rather that it's a good sign to implement one? like dvc reset or something?

[Suor]

@shcheklein I don't see one, but there probably should be. I am surprised dvc remove does something arcane instead of this.

[Suor]

BTW, this recipe is also dangerous considering how dvc gc operates.

[shcheklein]

@Suor agreed! Could you create a ticket for this on the DVC core repo? I also think that using current dvc gc is close to nuclear bombing. I would recommend to remove file manually for now.

[Suor]

Removing file from cache by hand is also unsafe, that one might be referenced somewhere else.