v2: encoding for / in group and array names

[joshmoore]

In AnnData, / is frequently used in key names. This leads to subdirectories in Zarr. In HDF5 it appears to be disallowed:

A path name is a string of components separated by '/'. Each component is the name of a (hard or soft) link, or the special characters "." (meaning current group). Link names (components) can be any string of ASCII characters not containing '/' (except the string ".", which is reserved). However, users are advised to avoid the use of punctuation and non-printing characters, because they may create problems for other software. Figure 39 gives a BNF grammar for HDF5 path names.

(via stackoveflow)

• Should text be added to the spec to be clearer about this issue?
• Should a preferred encoding of / be added?

cc: @ivirshup

[ivirshup]

The case where it comes up most commonly is in user provided column names in dataframes, e.g. "CD4+/CD8-". Our dataframe spec/ implementation conveniently (?) works around this by storing the names of columns as an attribute, so things end up working as expected in many cases. E.g. df_group["CD4+/CD8-"] returns the right data. Would definitely prefer if this didn't just work by implementation detail.

Some more discussion on the anndata repo (mostly discussing hdf5): scverse/anndata#321 (comment)