[v3] Missing array attributes: nbytes, nchunks, nchunks_initialized

[tomwhite]

Zarr version

3.0.0a1

Numcodecs version

0.12.1

Python Version

3.11.9

Operating System

Mac

Installation

pip

Description

The array attributes nbytes, nchunks, and nchunks_initialized are all missing in v3.

Steps to reproduce

>>> import zarr.v2 as zarr
>>> z = zarr.open(store='example-v2.zarr', mode='w', shape=(3, 2))
>>> z.nbytes, z.nchunks, z.nchunks_initialized
(48, 1, 0)
>>> import zarr
>>> z = zarr.open(store='example-v3.zarr', mode='w', shape=(3, 2))
>>> z.nbytes
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'Array' object has no attribute 'nbytes'
>>> z.nchunks
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'Array' object has no attribute 'nchunks'. Did you mean: 'chunks'?
>>> z.nchunks_initialized
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'Array' object has no attribute 'nchunks_initialized'

Additional output

No response

[d-v-b]

nbytes and nchunks should be easy, but i'm wondering if / how we should implement nchunks_initialized, since this triggers a lot of IO.

[d-v-b]

@tomwhite how essential is nchunks_initialized for your usage?

[tomwhite]

@tomwhite how essential is nchunks_initialized for your usage?

It's not as important as the other two, but we use it in Cubed to see if an array has been fully computed when resuming a computation after a restart. Currently we don't need to know how many chunks have been initialized, only if any haven't been - in which case we'll recompute them all. So an API like the one mentioned in zarr-developers/zarr-specs#300 (comment) for TensorStore would be sufficient.

[d-v-b]

so my personal view is that "what's going on with my chunks" is a very important question and zarr-python should provide basic tools with which people can answer that question, and making nchunks_initialized an array attribute gives the false impression that this is a cheap quantity to represent accurately, which it is not. So for me the easiest thing would be to spin nchunks_initialized off into a standalone function in array.py, with a docstring that explains the caveats and whatnot. Would that work for Cubed?

[tomwhite]

Would that work for Cubed?

Yes, thanks

[d-v-b]

while we are at it, chunks_initialized(array) -> tuple[chunk_identifier, ...] seems a bit better than nchunks_initialized(array) -> int. Something to think about when this PR gets opened.