[v3] Fixed-width unicode string support in zarr v3

[TomAugspurger]

Zarr version

v3

Numcodecs version

na

Python Version

na

Operating System

na

Installation

na

Description

Mentioned in #2323 (comment), right now we can't create a fixed-width string dtype in zarr v3.

In [1]: import zarr

In [2]: arr = zarr.create(shape=(3,), dtype="U3")

In [3]: arr[:] = ['a', 'bb', 'ccc']

In [4]: arr[:]
Out[4]: array(['a', 'bb', 'ccc'], dtype=StringDType())

We would want the NumPy dtype of that array to be U3, a fixed-width unicode string dtype. We'd want to support this in addition to the variable width strings being used currently. Some initial questions I don't know the answer to:

1. What data_type shows up in the metadata?
2. What codecs are needed?
3. How are the actual bytes stored? In parquet, fixed_len_byte_array is one of the primitive types.

Steps to reproduce

.

Additional output

No response

[TomAugspurger]

I'm not sure if I'll be able to work on this, but here's some notes on the V2 behavior, and some things:

>>> import numpy as np
>>> import zarr
>>> import json

>>> b = np.array([b'a', b'bb', b'ccc'])
>>> u = np.array(['a', 'bb', 'ccc'])
>>> store = {}

>>> zarr.array(b, store=store, path="bytes", compressor=None)
>>> zarr.array(u, store=store, path="unicode", compressor=None)

>>> print(json.loads(store['bytes/.zarray'])['dtype'])
# |S3


>>> print(json.loads(store['unicode/.zarray'])['dtype'])
# <U3

assert store['bytes/0'] == b.tobytes()
assert store['unicode/0'] == u.tobytes()

NumPy uses 32-bit UCS-4 codepoints for Unicode data ref. (I think that len(u.tobytes()) is something like 4 bytes per character * 3, since that's the fixed width). For bytes data, it uses the ASCII values padded with null bytes to the fixed width.

[dstansby]

Given it doesn't look like this funcitonality will get into 3.0, it looks like this breaking change is something else to add to #2596