Flaky test / buggy handling of large datetimes

[TomAugspurger]

Zarr version

v3

Numcodecs version

n/a

Python Version

n/a

Operating System

n/a

Installation

n/a

Description

Curious failure at https://github.com/zarr-developers/zarr-python/actions/runs/10549325420/job/29224231309?pr=2118 in #2118.

Steps to reproduce

I can reproduce locally, but I'm not too sure what's going on. With this diff:

diff --git a/tests/v3/test_properties.py b/tests/v3/test_properties.py
index 9204cdc5..5d2404b6 100644
--- a/tests/v3/test_properties.py
+++ b/tests/v3/test_properties.py
@@ -6,11 +6,12 @@ pytest.importorskip("hypothesis")
 
 import hypothesis.extra.numpy as npst  # noqa
 import hypothesis.strategies as st  # noqa
-from hypothesis import given, settings  # noqa
+from hypothesis import given, reproduce_failure, settings  # noqa
 from zarr.testing.strategies import arrays, np_arrays, basic_indices  # noqa
 
 
 @given(st.data())
+@reproduce_failure('6.111.2', b'AXicY2VgYmSAAhiDEcQAAADwAAw=')
 def test_roundtrip(data):
     nparray = data.draw(np_arrays)
     zarray = data.draw(arrays(arrays=st.just(nparray)))

I can get

(Pdb) pp nparray
array(['770652312998-01-12'], dtype='datetime64[D]')
(Pdb) pp zarray[:]
array(['1970-01-01'], dtype='datetime64[D]')

It seems like for large enough values, NumPy's datetime64[D] dtype is behaving strangely:

(Pdb) pp np.array(281474976710656, dtype='datetime64[D]').item()
281474976710656
(Pdb) pp np.array(0, dtype='datetime64[D]').item()
datetime.date(1970, 1, 1)

Additional output

Ahh, I think this is numpy/numpy#7619 (again). Python's datatime.date has limited precision, and so np.array(3000000, dtype='datetime64[D]') can't be represented as a date time. I'm pretty confused about what that test is doing, but presumably something inside Zarr isn't handling that edge case.