For some reason a dataset I've been working on produces BytesCodec instances where endian is None, which causes the translation from v3 to v2 metadata to fail.

are you able to share this dataset? If not, what dtype is this for? I'm wondering if it's a single byte dtype where the configuration should be empty. I'm reluctant to accept this change because an endian of None seems to be out-of-spec.

are you able to share this dataset?

Not sure, I'll have to find a public version of it.

what dtype is this for?

these are single-byte dtypes (in my case, int8, after a CF scale+offset compression). The encoding in the virtual dataset is indeed empty, but it still creates a BytesCodec instance with endian=None. The repr of v3_metadata is:

(BytesCodec(endian=None), Shuffle(codec_name='numcodecs.shuffle', codec_config={'elementsize': 1}), Zlib(codec_name='numcodecs.zlib', codec_config={'level': 1}))

which I assume is the default for this dtype?

from the v3 spec:

endian:

Required for data types for which endianness is applicable. For example, this includes multi-byte data types, such as uint16 and int32, but not single-byte data types, such as uint8 or bool. If present, the value MUST be a string equal to either "big" or "little".

but that's describing a JSON object, not an in-memory python object. in zarr python, our in-memory representation of the bytes codec uses endian=None to indicate that the JSON object should have endian key unset.

not exactly the same dataset, but this also fails with that error:

from obstore.store import from_url
from virtualizarr.parsers import HDFParser
from virtualizarr.registry import ObjectStoreRegistry

from virtualizarr import open_virtual_dataset

url = "https://data-cersat.ifremer.fr/data/sea-surface-temperature/odyssea/l4/glob/nrt/data/v2.1/2021/001/20210101000000-IFR-L4_GHRSST-SSTfnd-ODYSSEA-GLOB_010-v02.1-fv01.0.nc"
store = from_url(url)
registry = ObjectStoreRegistry({url: store})

vds = open_virtual_dataset(
    url=url,
    loadable_variables=[],
    parser=HDFParser(),
    registry=registry,
)
display(vds)
vds.vz.to_kerchunk("refs.parquet", format="parquet")

in digging through the zarr code, I noticed that

BytesCodec.from_dict({"name": "bytes", "configuration": {}})
BytesCodec.from_dict({"name": "bytes"})

both get the default system endian because that's the default for __init__, and it is only when explicitly passing None that that appears:

BytesCodec.from_dict({"name": "bytes", "configuration": {"endian": None}})

(but that would be against the spec, right?)

Also:

from zarr.codecs import BytesCodec

codec = BytesCodec(endian=None)
roundtripped = BytesCodec.from_dict(codec.to_dict())

assert codec.endian == roundtripped.endian

fails for the same reason.

@d-v-b, any idea on what the intended behavior should be?

the failing RTD build looks unrelated (it fails while creating the env)

We just ran into this and was about to open the same PR, thanks for working on this @keewis. It'll be great if this gets merged and we get a new release since this is a blocker for all earthaccess virtualizarr workflows. cc @maxrjones