feat: add cast_value and scale_offset codecs

[d-v-b]

Defines two new codecs that together provide a v3-native replacement for the existing numcodecs.fixedscaleoffset codec.

The cast_value codec requires an optional dependency on the cast-value-rs package. The reason for bringing in rust here is because it gives us better average speed and much better memory usage than a numpy implementation. Happy to show this with benchmarks if needed.

fwiw, I would rather not define these codecs in zarr-python. In fact I already defined these codec classes in another package. But it's not possible to make zarr-python depend on an externally-defined codec without creating a circular dependency, so here we are. See discussion in the related issue.

related, and possibly closes #3867

edit: if depending on a new optional package for cast_value is too much, then we should revisit #3774. The conclusion there was that defining the cast_value codec in zarr-python itself was too much. And if the conclusion is that both defining the codec externally is too much, and defining it internally is too much, then we should do something like deprecate numcodecs.fixedscaleoffset and direct people to the python packages that implement its replacement.

[d-v-b]

cc @emmanuelmathot

[codecov]

Codecov Report

❌ Patch coverage is 82.79070% with 37 lines in your changes missing coverage. Please review.
✅ Project coverage is 92.88%. Comparing base (0ea15fd) to head (2e8d644).

Files with missing lines	Patch %	Lines
src/zarr/codecs/cast_value.py	75.18%	34 Missing ⚠️
src/zarr/codecs/scale_offset.py	95.94%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3874      +/-   ##
==========================================
- Coverage   93.07%   92.88%   -0.20%     
==========================================
  Files          85       87       +2     
  Lines       11228    11443     +215     
==========================================
+ Hits        10451    10629     +178     
- Misses        777      814      +37     

Files with missing lines	Coverage Δ
src/zarr/codecs/__init__.py	100.00% <100.00%> (ø)
src/zarr/codecs/scale_offset.py	95.94% <95.94%> (ø)
src/zarr/codecs/cast_value.py	75.18% <75.18%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[d-v-b]

the coverage gaps are spurious because some of these routines only run when the optional cast-value-rs package is available. this is ready for review.

[vincentsarago]

👋 This is great.

With this codec we could reduce the size of our stored dataset 🙏

[maxrjones]

Nice, thanks @d-v-b! A few comments:

protect against data corruption

I think you need to protect against integer overflow/underflow. Here's a couple tests that demonstrate corrupted data:

def test_encode_uint8_underflow_should_error():
    """uint8(0) - offset(1) wraps to 255, then * 2 = 254. Decode gives 128, not 0."""
    arr = zarr.create_array(
        store={},
        shape=(3,),
        dtype="uint8",
        chunks=(3,),
        filters=[ScaleOffset(offset=1, scale=2)],
        compressors=None,
        fill_value=1,
    )
    data = np.array([0, 1, 2], dtype="uint8")
    with pytest.raises((ValueError, OverflowError)):
        arr[:] = data


def test_encode_int8_overflow_should_error():
    """int8(50) * scale(3) = 150 wraps to -106. Decode gives -36, not 50."""
    arr = zarr.create_array(
        store={},
        shape=(1,),
        dtype="int8",
        chunks=(1,),
        filters=[ScaleOffset(offset=0, scale=3)],
        compressors=None,
        fill_value=0,
    )
    data = np.array([50], dtype="int8")
    with pytest.raises((ValueError, OverflowError)):
        arr[:] = data

user-guide documentation

I think it'd be worth having some documentation page for users coming from other data formats, since they could easily mix up the scale offset parameters. Here's a summary:

The main ones across scientific/geospatial data:

CF conventions (NetCDF): unpacked = packed * scale_factor + add_offset

• Scale means "multiply when unpacking"

GDAL/GeoTIFF: value = raw * scale + offset

• Same formula as CF. GDAL uses this for raster band transformations.

HDF-EOS5: value = scale_factor * (stored - offset)

• Subtract offset before scaling — different from CF which adds offset after scaling. This is a common source of confusion between HDF-EOS and CF.

GRIB (WMO): value = reference_value + (raw * 2^binary_scale_factor) / 10^decimal_scale_factor

• Two scale factors (binary and decimal) plus a reference value. Much more complex.

numcodecs FixedScaleOffset (zarr v2): encode: round((x - offset) * scale), decode: x / scale + offset

• Same formula as the new zarr spec, plus explicit rounding on encode.

Zarr v3 scale_offset: encode: (x - offset) * scale, decode: x / scale + offset

• Same as numcodecs but without built-in rounding (that's now delegated to cast_value).

ML quantization (e.g., TensorFlow Lite): real = (quantized - zero_point) * scale

• Scale means "multiply when unpacking" (like CF). zero_point is the integer that maps to real zero.

The key axes of variation:

	Scale means	Order	Rounding
CF / GDAL / ML quant	multiply to unpack	scale then offset	implicit
HDF-EOS5	multiply to unpack	offset then scale	implicit
Zarr v3 / numcodecs	multiply to pack	offset then scale	explicit (cast_value)
GRIB	two scale factors	its own thing	N/A

The biggest gotcha is that CF's scale_factor=0.1 means the same thing as zarr's scale=10. Users migrating CF data need to invert the scale value.

sustainability

Will you add me or the zarr team as a maintainer on PyPI for cast-value-rs if this is merged?

[maxrjones]

per the spec, should this use / and error if the result isn't representable in the data type?

[maxrjones]

these should be tested in the CI