build: make zarr dependency optional

[tlambert03]

proof of principle to make the zarr dependency optional. (plenty of space to discuss the exact code patterns, and how the hatch matrix was done)

[dstansby]

Thanks for this, and the PR over at ome-zarr models! I'm very pro having an optional dependency on zarr-python in both places, if it's possible.

My immediate question is how are you envisaging using this library without zarr-python, with both to_zarr() and from_zarr() not working? All my personal use cases need one or other of those methods, but it's very possible I'm missing a way to use these libraries!

[d-v-b]

pydantic-zarr essentially defines a bidirectional mapping from zarr hierarchies to JSON documents, and zarr-python is just one of several possible ways to handle the translation.

eventually, an API like this would be great:

# use zarr python
GroupSpec.to_zarr(engine='zarr-python')
# use tensorstore
GroupSpec.to_zarr(engine='tensorstore')
# use a custom zarr engine
GroupSpec.to_zarr(engine=MyCustomZarrEngine)

But the return type of to_zarr starts to get complicated here!

[tlambert03]

how are you envisaging using this library without zarr-python, with both to_zarr() and from_zarr() not working

yeah, the mapping to/from json documents that @d-v-b mentioned is what I’m after. And the use case is https://github.com/pymmcore-plus/ome-writers … which is a library for writing both ome-tiff and ome-zarr as data comes off the microscope. We have alternative backends: for zarr it’s either tensorstore or acquire-zarr (high performance libs that are both much faster than zarr-python).

Having a model that simplifies the process of constructing valid, spec-conforming json documents, without stipulating exactly how those documents will be written to disk, is what I’m after.

[dstansby]

So the idea is you would manually create one of the models, and then serialise it to JSON using the methods available through pydantic (.model_dump_json()) and 'manually' write it to the correct Zarr metadata file? That sounds good to me - just want to make sure, partly so the workflow in this case can be documented.

[tlambert03]

Yes exactly.👍

Edit: for what it’s worth, if this ome-models-zarr were no more than pydantic, I would also likely use it as a public object for declaring intention prior to creating a stream: I.e “I want to acquire a well plate with these wells, etc”

[tlambert03]

@d-v-b, let me know if you have a better idea for doing the hatch matrix here... or if you're happy to switch this repo over to using uv and do it similarly to ome-zarr-models/ome-zarr-models-py#280