setitem with repeated indices is not guaranteed in practice

[dcherian]

Zarr version

v2.18.3, 3.0.2 and 3.0.3

Numcodecs version

?

Python Version

?

Operating System

linux & windows

Installation

either

Description

behaviour of setitem with repeated indices depends on platform. The example below succeeds on Mac and fails on ubuntu, windows with numpy 2.2. It succeeds in all CI environments with numpy 1.25.

I can reproduce on zarr v2.18.4, v3.0.2 on linux.

I say we disallow this

Steps to reproduce

import zarr
import numpy as np

array = zarr.array(data=np.zeros((4,)), chunks=(1,))
indexer = np.array([-1, -1, 0, 0])
array.oindex[(indexer,)] = [0, 1, 2, 3]
np.testing.assert_array_equal(array[:], np.array([3, 0, 0, 1]))

fails on windows & linux, numpy=2.2 in CI.

cc @LDeakin @ilan-gold

[d-v-b]

is numpy's behavior also platform-dependent? If not, then I think we should copy numpy here. On my OSX machine, numpy uses the last value assigned:

>>> x = np.arange(4)
>>> x[np.ones(10, dtype='int')] = np.arange(10)
>>> x
array([0, 9, 2, 3]) # value at index 1 is set to the last value of np.arange(10)

We probably need to transform the input such that it defines an indexing operation that only applies 1 change to each output index.

[LDeakin]

We probably need to transform the input such that it defines an indexing operation that only applies 1 change to each output index.

That'd be perfect

[dcherian]

numpy behaviour is always consistent and pretty well-defined since it's serial.

I don't understand what's going on here. transforming the input does seem sensible.