An empty numpy array that defines strides may not provide stable tokenization upon pickle roundtrips

import numpy as np
arr = np.array([])

Note: This only works when using the array constructor, not the ndarray since ndarray will initialize strides to ()

The above numpy array is empty but is defining a strides attribute

>>> arr.strides
(0,)

which means that the normalize_token will return a four tuple where the last element are the strides.

>>> from dask.base import normalize_token, tokenize
>>> normalize_token(arr)
('da39a3ee5e6b4b0d3255bfef95601890afd80709', dtype('int64'), (0,), (0,))

A pickle roundtrip of this empty array now sets the strides to a non-zero element

>>> import pickle
>>> pickle.loads(pickle.dumps(arr)).strides
(8,)

# This should be true
>>> assert tokenize(arr) != tokenize(roundtrip)

which is causing our deterministic tokenization to break. This is a problem for graphs that are only generated on the scheduler since the original keys on client side may differ from those on scheduler side.

numpy                     1.26.2          py310h30ee222_0    conda-forge
pandas                    2.1.4           py310h6e3cc31_0    conda-forge
python                    3.10.13         h2469fbe_0_cpython    conda-forge