useful reading/links
- https://zarr-specs.readthedocs.io/en/latest/v3/core/v3.0.html#id21
- zarr-developers/zarr-python#556
- https://medium.com/pangeo/cloud-performant-netcdf4-hdf5-with-zarr-fsspec-and-intake-3d3a3e7cb935
- https://github.com/zarr-developers/geozarr-spec
- https://github.com/ome/ome-zarr-py
- https://xarray-datatree.readthedocs.io/en/latest/
- napari generative zarr: link to relevant part of PR and the yt in napari generative zarr experiment
- https://github.com/manzt/napari-dzi-zarr zarr container for DZI for reading in napari
- https://janeliascicomp.github.io/pydantic-ome-ngff/ pydantic models of ome-ngff
- zarr-developers/zarr-specs#50 zarr specs discussion about storing image pyramids
- https://github.com/cloudnativegeo/cloud-optimized-geospatial-formats-guide/blob/feat/determine-chunking/zarr/determine-chunk-shape.ipynb chunk size optimization
- https://github.com/zarr-developers/VirtualiZarr
- napari
- napari/napari#513
- neuroglancer? https://github.com/google/neuroglancer https://www.brainimagelibrary.org/visual.html
- sph frontends: should be fairly straightforward to implement a frontend for an sph code with a zarr-hdf backend.
- grid-based: uniform or stretched grids with no refinement also should be simple-ish
- grid-based, refined: refined grids are a little trickier... could implement a custom zarr store that fills at a given refinement level (similar to a YTArbitraryGrid). this would kinda look like napari's generative zarr