Read/write errors when / is in var names

[LucaMarconato]

Please make sure these conditions are met

• I have checked that this issue has not already been reported.
• I have confirmed this bug exists on the latest version of anndata.
• (optional) I have confirmed this bug exists on the master branch of anndata.

Report

Related to #321.

When a var name contains a slash, for instance a/b, this is written on disk as a folder a, and then b a subfolder of a. See screenshot. This creates problems, such as in the code below. Previously in spatialdata we observed also this OS-dependent problem (appearing on Windows for some users, but that we couldn't reproduce) #1447.

Code:

import anndata as ad
import pandas as pd

adata0 = ad.AnnData(var=pd.DataFrame([[1, 2], [3, 4]], columns=['a/b', 'a']))
adata1 = ad.AnnData(var=pd.DataFrame([[1, 2], [3, 4]], columns=['a', 'a/b']))
adata2 = ad.AnnData(var=pd.DataFrame([1, 2], columns=['a/b']))

# ok
adata0.write_h5ad('temp/adata.h5ad')
adata0.write_zarr('temp/adata.zarr')

# error
_ = ad.read_h5ad('temp/adata.h5ad')

# error
_ = ad.read_zarr('temp/adata.zarr')

# error
adata1.write_h5ad('temp/adata.h5ad')

# error
adata1.write_zarr('temp/adata.zarr')

# ok
adata2.write_h5ad('temp/adata.h5ad')
_ = ad.read_h5ad('temp/adata.h5ad')
adata2.write_zarr('temp/adata.zarr')
_ = ad.read_zarr('temp/adata.zarr')

The reasons of the errors is that on disk a/b is saved as a folder a with b as a subfolder.

In the first 2 errors, b is not present in a since the var name a overwrites the previous a folder.

In the latest 2 error the a/b folder cannot be written because a was written first.

In the last part of the script (no error), you can reproduce having b written as a subfolder of a.

Proposed solution

I propose to disallow the character / for column names of obs, var, obsm, varm, obsp, varp, uns.
See also #321 (comment)

Versions

/opt/miniconda3/envs/ome311/lib/python3.11/site-packages/session_info2/__init__.py:124: UserWarning: The '__version__' attribute is deprecated and will be removed in MarkupSafe 3.1. Use feature detection, or `importlib.metadata.version("markupsafe")`, instead.
  and (v := getattr(pkg, "__version__", None))
anndata	0.11.3
----	----
scipy	1.15.0
session-info2	0.1.2
typing_extensions	4.12.2
MarkupSafe	3.0.2
Jinja2	3.1.4 (3.1.5)
tblib	3.0.0
natsort	8.4.0
zarr	2.18.3
PyYAML	6.0.2
zipp	3.21.0
sphinxcontrib-qthelp	2.0.0
sphinxcontrib-jsmath	1.0.1
asciitree	0.3.3
charset-normalizer	3.4.1
packaging	24.2
torch	2.5.1
numcodecs	0.14.1
importlib_metadata	8.5.0
pyarrow	18.1.0
sphinxcontrib-bibtex	2.6.3
tqdm	4.67.1
h5py	3.12.1
sphinxcontrib-applehelp	2.0.0
psutil	6.1.0
dask	2024.11.2
pytz	2024.1
msgpack	1.1.0
numpy	1.26.4 (2.0.0)
sphinxcontrib-serializinghtml	2.0.0
six	1.16.0 (1.17.0)
python-dateutil	2.9.0.post0
sphinxcontrib-htmlhelp	2.1.0
pandas	2.2.3
cloudpickle	3.1.0
toolz	1.0.0
setuptools	75.6.0
sphinxcontrib-devhelp	2.0.0
----	----
Python	3.11.0 | packaged by conda-forge | (main, Jan 14 2023, 12:26:40) [Clang 14.0.6 ]
OS	macOS-14.6.1-arm64-arm-64bit
Updated	2025-01-20 11:18

[flying-sheep]

strange, why is that only a problem on windows?

[LucaMarconato]

The bug that I am reporting now occurs on macOS. The linked bug was a problem on windows because on macOS, if I write only a/b and not also a, the read write works, but when moving such anndata Zarr store (on disk) to a Windows machine, I guess something gets messed up due to the difference between / being a path separator on macOS, but not on Windows.