Negative zero not preserved

[tomwhite]

Zarr version

v3.0.8

Numcodecs version

v0.15.1

Python Version

3.11

Operating System

Mac

Installation

pip in a venv

Description

Negative zero (-0.0) is not distinguished from zero when reading back values from a Zarr array. This differs from the expected behaviour (and Zarr Python v2).

Steps to reproduce

# /// script
# requires-python = ">=3.11"
# dependencies = [
#   "zarr@git+https://github.com/zarr-developers/zarr-python.git@main",
# ]
# ///
#
# This script automatically imports the development branch of zarr to check for issues

import numpy as np
import zarr
print(zarr.array(np.array([0.0, -0.0]))[:])
#zarr.print_debug_info()

This prints

[0. 0.]

But with Zarr 2.18.7 it correctly prints

[ 0. -0.]

Additional output

No response

[d-v-b]

thanks for this report, I did a quick check and this bug was present prior to the recent dtype changes, so that should help narrow the search down a bit.

[tomwhite]

thanks for this report, I did a quick check and this bug was present prior to the recent dtype changes, so that should help narrow the search down a bit.

Yes, sorry I forgot to say it is in 3.0.8 as well as main.

[d-v-b]

# /// script
# requires-python = ">=3.11"
# dependencies = [
#   "zarr==3.0.8",
# ]
# ///
#
# This script automatically imports the development branch of zarr to check for issues

import numpy as np
import zarr
print(zarr.array(np.array([0.0, -0.0]), fill_value=1)[:])

Adding a non-0 fill value preserves the signed 0 as expected. I also checked that the original test case doesn't write any chunks. So the error comes from zarr mistakenly identifying the chunk as empty during array creation, thenfailing to write it, which results in zarr using the fill value instead on indexing.

[tomwhite]

Aha - which explains why it works on v2 - the default for write_empty_chunks is True (it's False in v3). On v2 the following fails too:

>>> print(zarr.array(np.array([0.0, -0.0]), write_empty_chunks=False)[:])
[0. 0.]

[d-v-b]

I suspect we are relying on this behavior:

>>> np.array([0.0]) == np.array([-0.0])
array([ True])

We probably need to define a stricter type of equality.