Inconsistent `()` indexing/size behavior from v2->v3

[ilan-gold]

Zarr version

v3.0.0

Numcodecs version

0.14.1

Python Version

3.12

Operating System

Mac

Installation

uv

Description

Previously something like z[()] where z is a () shape array would return a scalar and now returns a one-elem array e.g., np.array(1) instead of np.float64(1.0)

Steps to reproduce

v3:

import numpy as np
import zarr

z = zarr.open_group('foo.zarr')
z.create_array("bar", dtype=float, shape=())
z["bar"][...] = np.array(1) # why doesn't [:] work?
z["bar"][()] # array

v2:

import numpy as np
import zarr

z = zarr.open_group('foo.zarr')
z.create_dataset("bar", dtype=float, shape=(), data=np.array(1))
z["bar"][()] # np.float

Additional output

No response

[jhamman]

Thanks @ilan-gold for the report. Looks like we missed the () indexer in #2583

[brokkoli71]

I'll take a look at it.

regarding your comment in:
z["bar"][...] = np.array(1) # why doesn't [:] work?

The semantics of [:] is "select everything in the first dimension". I think since we have 0 dimensions this is expected to fail because there is no first dimension.

[ilan-gold]

The semantics of [:] is "select everything in the first dimension". I think since we have 0 dimensions this is expected to fail because there is no first dimension

Makes sense!

[brokkoli71]

Do we want to return scalars when accessing zero-dimensional arrays? This would mean that the return value of __getitem__ is no longer NDArrayLike and may make typing more difficult

[dstansby]

Returning a scalar here would be consistent with numpy (2.2.1):

>>> import numpy as np
>>> a = np.array(1)
>>> a
array(1)
>>> a[()]
np.int64(1)