Relax Path Safety on Parse

[tustvold]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.

Currently PathPart::parse and by extension Path::parse disallow paths containing "unsafe" characters.

This feels unnecessarily puritanical given:

• Most stores and filesystems do actually support UTF-8
• We don't exclude characters that some filesystems don't support, e.g. : on Mac OSX
• We don't exclude filenames that are forbidden on some systems, e.g. COM1 on Windows
• We exclude characters that are fairly common, e.g. ~
• We actually add % characters to paths, which the stores recommend against including in object names

Describe the solution you'd like

I think we should preserve the existing percent encoding behaviour when using the infallible methods, this gives people a safe default behaviour.

However, I would propose we change PathPart::parse to only reject segments that

• Are "" or "." or ".."
• Contain ASCII control characters
• Contain a delimiter

This broadly conforms to what is supported by most filesystems - https://en.wikipedia.org/wiki/Filename

Describe alternatives you've considered

Additional context

apache/datafusion#8014

[crepererum]

Do file-based object stores count as a "sandbox"? If so, we should probably double-check that there aren't some weird ways to escape the root directory via some bizarre unknown features.

Slightly related: do we follow symlinks?

[tustvold]

Do file-based object stores count as a "sandbox"

I don't think we have ever billed it as such, I certainly wouldn't rely on it, but that doesn't mean people aren't using it as such.

do we follow symlinks

Yes and they can allow escaping the "sandbox", this is documented here - https://docs.rs/object_store/latest/object_store/local/struct.LocalFileSystem.html#symlinks

escape the root directory via some bizarre unknown features

We prevent relative paths, but there might be other ways I'm not aware of, especially on Windows.

[tustvold]

label_issue.py automatically added labels {'object-store'} from #5020