create_array w/ overwrite=True doesn't clean up an existing array

[d-v-b]

MRE

# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "zarr == 3.0.0",
# ]
# ///

from zarr import create_array
import numpy as np
shapes = (2,1), (2,)
for shape in shapes:
    data = np.arange(np.prod(shape) , dtype='uint8').reshape(shape)
    arr = create_array(
        'data/v2_array_example.zarr',
        shape=data.shape,
        dtype=data.dtype,
        zarr_format=2,
        compressors='auto',
        overwrite=True,
        chunk_key_encoding={'name': 'v2', 'configuration': {'separator': '/'}},
        )

    arr[:] = data # fill the array with values
    print(arr.info_complete())

Traceback (the latter bit):

...
  File "/Users/bennettd/Library/Caches/uv/archive-v0/Dt_Vb9Xzyh8CGtlKxZyUO/lib/python3.11/site-packages/zarr/storage/_local.py", line 62, in _put
    with path.open(mode=mode) as f:
         ^^^^^^^^^^^^^^^^^^^^
  File "/Users/bennettd/.pyenv/versions/3.11.9/lib/python3.11/pathlib.py", line 1044, in open
    return io.open(self, mode, buffering, encoding, errors, newline)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
IsADirectoryError: [Errno 21] Is a directory: 'data/v2_array_example.zarr/0'

This is a simple bug, but it also reveals the meaning of "overwrite" is not clear.

• I would expect creating an array, then creating an array at the same path withoverwrite=True to completely remove the pre-existing array.
• I'm not sure I would expect overwrite=True to erase absolutely everything prefixed underneath the array. So if I had saved cat.jpgin the same directory as my array, I might not expect that to get erased.

It might be reasonable to define "create_x(overwrite=True)" to mean "ensure that all keys necessary to create object x are available, but no other keys will be touched". But issuing a huge volume of delete_object requests against metered cloud storage could be much more expensive than a simple "delete_prefix". I don't know much about this, someone good at s3 billing should help here.

Curious to hear what other people think about what "overwrite" should mean here.

[jhamman]

I feel that we should be clearing chunks whenever overwrite is used. Calling Store.delete_dir whenever overwite is called seems like the prudent action.