feat: Move files to .trash folder if they are in use

[adament]

Description

Attempt to move files to .trash folder if they are in use, this should solve #578. The issue is that on Windows we cannot delete files that are in use. When using a global package cache and hardlinking files into environments this means that if a given DLL or exe is in use in any environment with the same package the deletion files. But we can rename the hardlink. Conda and mamba have similar fallback behaviour.

The solution in this PR is slightly different than the ones used in conda and mamba as far as I am aware. This PR:

• Creates a folder in .trash in the root of the environment rather than having a file like mamba_trash.txt to keep track of all the files which could not be deleted. Then a separate empty_trash procedure should simply attempt to delete all files in the .trash folder and if successful, it should remove the .trash folder.
• When moving a file to .trash it is renamed to uuid.rattler_trash. Mamba and conda kept the original filename and appended .mamba_trash, but then they had to worry about filename collisions and having some fallback procedure for coming up with new names on a collision. It thus seemed simpler to me to just give a unique name to each file in the trash folder. Though in this way we lose information about the origin of the file.

Please tell me if you think the approach or the implementation is wrong.

[wolfv]

Awesome! I am wondering - did you check that "moving" to a different folder works when the file is in use? If yes, then this is a great solution.

I also thought that we could try to have filename.<uuid>.trash as a filename but it might get longer than the allowed length.

I think going for UUID is fine with me. We would probably also be on the safe side with, idk, filename + 4 random characters or so.

There are also ways to add metadata to files that are NTFS specific (similar to xattrs on Unix) but not sure if that would be worth it to investigate.

[adament]

I have attempted to add a test to the test suite which should test that the trash procedure works when a program is running, and I just checked manually in explorer that I can move python.exe when it is in use. My current workaround for pixi not having trash support is to just move the entire .pixi folder somewhere else and then run pixi install. So I feel reasonably confident that it should work.

I am fine with any sort of filename template. However when I looked at it rand is not currently a dependency of rattler, so generating a uuid seemed the simplest way to get a sufficient amount of randomness to not have to worry about collisions. I like the filename.<uuid>.trash template but as you write I am also a bit worried about exceeding path length if the filename is very long. Do you want me to make the change to include the filename?

[baszalmstra]

Lets add the filename indeed to make it slightly easier to debug. If filename becomes to long we can worry about it then.

[adament]

I have changed it to the filename.<uuid>.trash pattern and attempted to make the test more robust.

[baszalmstra]

Thank you so much for this PR and your patience reviewing our comments! Highly appreciated! :)

[adament]

Thank you for taking the time to review and comment.