Readers and forms
Which reader opens which published form, by codec
Every form of a dataset holds the same byte references. Whether a reader can use a form depends on the codecs the source chunks need, so the matrix is per source family:
- Bluelink (BRAN netCDF): int16 chunks, HDF5 shuffle + zlib. In Zarr v3 (Icechunk) the shuffle is
numcodecs.shuffle; in Zarr v2 (kerchunk) it is theshufflefilter. - GHRSST MUR COG: int16 tiles, ZSTD with TIFF horizontal predictor (
PREDICTOR=2), written as theimagecodecs_deltafilter.
ok read and checked against GDAL on the sources; partial read with a caveat; no the reader rejects the form; not run no run recorded yet.
| Reader | Form | Bluelink | GHRSST COG |
|---|---|---|---|
| xarray + icechunk (zarr-python) | Icechunk | ok at zarr async.concurrency 4 (THREDDS resets at 10) |
ok with imagecodecs registered |
| xarray + fsspec (zarr-python) | kerchunk | ok with engine="pyarrow" refs (see below) |
ok with imagecodecs registered |
GDAL multidim (osgeo.gdal, gdalxarray, gdal mdim) |
mdim VRT | ok | ok one HEAD + 2 GET per source; parallel by process |
GDAL Zarr driver (ZARR:"/vsicurl/.../kerchunk/") |
kerchunk | ok same values as the VRT | no no imagecodecs_delta |
GDAL /vsiicechunk/{...} |
Icechunk | no no numcodecs.shuffle in Zarr v3 |
no no imagecodecs_delta |
terra (rast(vrt, md = TRUE)) |
mdim VRT | ok same values as gdalxarray; layers Time fastest, then depth | ok with CPL_VSIL_CURL_CACHE_SIZE raised |
| wxy-refs | flat refs | partial synthetic BRAN only | no predictor decode pending |
| gdalraster multidim | mdim VRT | not run | not run |
Notes. fsspec reads kerchunk refs with fastparquet by default, which drops chunk paths in some refs files published before 2026-11-03; pass engine="pyarrow". xarray decodes packed int16 to float64 from the Zarr forms, because the JSON attributes do not keep the float32 scale_factor. The “no” cells are GDAL codec gaps, tracked upstream.