dirfs info not working with ftp implementation when path is root, appeared in 2025.3.0 release

[gmaze]

Hi,
My CI upstream tests at argopy started to fail after the 2025.3.0 fsspec release and it seems to arise from changes added by #1798 in the last 2025.3.0 release.

Here is a reproducible example.

The following works ok:

fs = fsspec.filesystem("dir", fs=fsspec.filesystem('ftp', host='ftp.ifremer.fr'), path='ifremer/argo')
fs.info('dac/aoml/13857/13857_meta.nc')

but in this scenario:

fs = fsspec.filesystem("dir", fs=fsspec.filesystem('ftp', host='ftp.ifremer.fr'), path='/')
fs.info('ifremer/argo/dac/aoml/13857/13857_meta.nc')

we get the error:

---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
Cell In[2], line 2
      1 fs = fsspec.filesystem("dir", fs=fsspec.filesystem('ftp', host='ftp.ifremer.fr'), path='/')
----> 2 fs.info('ifremer/argo/dac/aoml/13857/13857_meta.nc')

File ~/miniconda3/envs/argopy-py310-all-free/lib/python3.10/site-packages/fsspec/implementations/dirfs.py:242, in DirFileSystem.info(self, path, **kwargs)
    240 info = self.fs.info(self._join(path), **kwargs)
    241 info = info.copy()
--> 242 info["name"] = self._relpath(info["name"])
    243 return info

File ~/miniconda3/envs/argopy-py310-all-free/lib/python3.10/site-packages/fsspec/implementations/dirfs.py:74, in DirFileSystem._relpath(self, path)
     72     if self.path.startswith(self.fs.sep) and not path.startswith(self.fs.sep):
     73         prefix = prefix[1:]
---> 74     assert path.startswith(prefix)
     75     return path[len(prefix) :]
     76 return [self._relpath(_path) for _path in path]

AssertionError:

If we run the same test on a s3 implementation, it works ok:

fs = fsspec.filesystem("dir", fs=fsspec.filesystem('s3'), path='argo-gdac-sandbox/pub')
fs.info('dac/aoml/13857/13857_meta.nc')

provides similar results to:

fs = fsspec.filesystem("dir", fs=fsspec.filesystem('s3'), path='/')
fs.info('argo-gdac-sandbox/pub/dac/aoml/13857/13857_meta.nc')

[gmaze]

Note that I noticed that:

fs = fsspec.filesystem("dir", fs=fsspec.filesystem('ftp', host='ftp.ifremer.fr'), path='/.')
fs.info('ifremer/argo/dac/aoml/13857/13857_meta.nc')

and

fs = fsspec.filesystem("dir", fs=fsspec.filesystem('ftp', host='ftp.ifremer.fr'), path='.')
fs.info('ifremer/argo/dac/aoml/13857/13857_meta.nc')

works fine

[martindurant]

Of course we should fix it, but naturally a prefix of "/" should essentially be a no-op for dirFS, not useful.

[gmaze]

actually I quite don't understand why @martindurant
(sorry to bother, I'm trying to make sense for all this in order to fix my library, I'm afraid to write non-necessary code downstream because I don't get how fsspec works)

To me, the root level "/" refer to the top level of the file system hierarchy,
which also makes sense in the case of a dirFS that simply shift the root level to a specific different level

this is probably also because I don't understand why it was necessary to make ls and info attribute name to return the same value in #1798 (the breaking change for me between 2025.2 and 2025.2):
info returned file information, whatever the file system (dirFS or not), hence an absolute path,
while ls return file information for a given file system (hence relative or absolute, with dirFS or not). It was even complementary with each others.

[martindurant]

I convention has been, that ls(path)[0] == info(path) where the path is for a file (as opposed to a directory). In fact, that's how it's coded in the base class fallback implementation.

The "name" key should always be relative the filesystem in which it is being used. It would be reasonable to include extra information in other keys, something like "original name" for this case. But dirFS is supposed to work something like posix mount, that you can't necessarily even know that your files live within some other root directory.

[gmaze]

thanks for your time in explaining this !
closing