On disk versioning feedback

[ivirshup]

@giovp

Two main points of feedback on how versioning is done for the on-disk format. One minor one.

Version the elements separately

I think it makes a lot of sense to version the elements separately. For example, I can see wanting to change how points are encoded, but only that. I don't think we need to bump the version number for tables, shapes, images, just because points changed.

Use something other than @type

I think @type is a weird option for reserved keyword since it's not how either anndata or OME specifies the type of an element. How we can do "metadata conventions" in a consistent way has been discussed around zarr quite a few times (e.g. ZEP 4). I'm not completley settled on what should be done, but can suggest some alternatives.

I'll try and figure out which option I like the most ASAP.

Use anndata's encoding_type, encoding_version

We could just use anndata's encoding_type, encoding_version.

Advantages

• consistency with anndata
• Could eventually hook into the anndata IO dispatch system

Main disadvantage

The spatial data table is a superset of an AnnData. So if we're now using the same key, what's the type? If it's anndata, the attributes won't be part of the spec. If it's spatialdata-table or ngff:regions_table it's overwriting that it's also an anndata.

That said, the anndata library could just register ngff:regions_table to be read in as an AnnData. This library could even do that registration with:

from anndata._io.specs import _REGISTRY, IOSpec
from anndata._io.specs.methods import read_anndata
import zarr

_REGISTRY.register_read(zarr.Group, IOSpec("ngff:region_table", "0.5"))(read_anndata)

But that's only if you DON'T want to do something custom at read time, e.g. anything with the other attributes.

Use some separate field

This could be something like spatialdata-encoding-type. I personally would like this over @type, since @type is a kinda taken convention, which nothing else here is

Use some separate name-spaced field

I think long term it would make sense to namespace these things. For example:

.attrs = {
	"encoding": {
		"spatialdata": {"type": "spatialdata-table", "0.1"},
		"anndata": {"type": "anndata", "version": "0.1.0"},
		"ngff": {"type": "table", "version": "0.5"}
	}
	...
}

Where spatialdata knows that it's a superset of "anndata", and to check for attrs["encoding"]["anndata"].

While this would eventually require coordination, one could start doing it here. Would probably require more thought as I think this will actually be an issue for arrow tables.

Minor point: also write the software version into the attributes

Something that mudata does (like this), which could be quite helpful for debugging, is to write the library and version into metadata.

It would also be nice for devs to see that one weird file is weird because it was written by a development version.

[LucaMarconato]

Thanks for initiating the discussion.

I agree on the versioning of the elements separately and on having the software version in the attributes. I am not familiar with the rest so I won't comment it.