User attributes not being saved to file

[dstansby]

Zarr version

2.18.2

Numcodecs version

0.12.1

Python Version

3.11

Operating System

macOS

Installation

conda

Description

I'm trying to save user attributes to file, but they don't seem to be saved when calling zarr.save_array.

Steps to reproduce

import numpy as np
import zarr

arr = zarr.ones((512, 512, 512), dtype=np.uint8, chunks=256)
arr.attrs["attr"] = "value"
zarr.save_array("test_attr_arr.zarr", arr)

The resulting zarr array on disk does not have a .zattrs file as I would expect.

Additional output

No response

[jhamman]

I think this is actually the expected behavior. Without save_array the attrs are saved.

import numpy as np
import zarr

arr = zarr.ones((512, 512, 512), dtype=np.uint8, chunks=256)
arr.attrs["attr"] = "value"
print(arr.store[".zattrs"])
# b'{\n    "attr": "value"\n}'

But save_array creates a new array where the arr argument is a numpy-like thing and does not look for attrs. So your last line is creating a new array in a new store.

[dstansby]

🤔 so how do I save the attributes to disk?

[jhamman]

When you create an array, you need to associate a store (the default, as you have shown, uses a memory store). This works 👇

import numpy as np
import zarr

arr = zarr.ones((512, 512, 512), dtype=np.uint8, chunks=256, store="test_attr_arr.zarr")
arr.attrs["attr"] = "value"

[jhamman]

Going to take the ❤️ as permission to close this. Feel free to reopen if you run into more gotchas!

(btw, you'll be able to do this all in one call in v3:

arr = zarr.ones(
    shape=(512, 512, 512),
    dtype=np.uint8,
    chunks=256,
    store="test_attr_arr.zarr",
    attributes={"attr": "value"}
)

[dstansby]

What I really want to do is work with an array in memory (including adding some attributes), and then save it in one go to disk. So I think it's worth at least documenting how to do that workflow somewhere?

[d-v-b]

I think this would be equivalent to copying the memory store to a directory store. Is there not a copy_store routine in v2? With the v2 mutable mapping api, you might be able to to do local_store.update(**memory_store), although users should see something a bit more familiar

[dstansby]

Yep, there is a copy_store function.

Clearly my mental model of how zarr (or at least save_array) works wasn't very good, so I might suggest some improvements to the save_array docstring explaining the difference between save_array and copy_store.

[d-v-b]

a few thoughts about this flow:

• it's good to make it dead-simple for users to save numpy arrays to zarr. If zarr v2 allowed specifying attrs at array creation time, then your first attempt would have worked @dstansby, because save_array forwards **kwargs to _create_array under the hood, and that's a route attrs could take, but only if _create_array took attrs, which it does not. A million years ago I had a PR to fix this, but I think with the v3 api joe showed we don't have this problem any more.
• there are other arrays to consider. in particular, dask arrays come with chunks, and xarray.DataArrays come with optional chunks AND attrs. I think we should have a single entry point for fluently converting generic chunked-array-with-attrs-like objects into Zarr arrays. Over in pydantic-zarr, I implemented a from_array function that takes an array-like input (i.e., shape and dtype are required) and checks if that input has an attrs attribute, or a chunks attribute, or a filters attribute, and so on for all the zarr array attributes, to create the resulting (model) zarr array (and this attribute inference can be overridden with a concrete value). I think we should do something similar in zarr-python. This gets a bit more complicated for array-like objects that have different "syntax" for the same semantics, e.g. the dask array chunks attribute is an explicit list of chunk sizes. We might need to define this variation via protocols, and dispatch on the shape of incoming array-like objects. Work to be done for sure, but I think this is something that a lot of users would appreciate.

[d-v-b]

This issue made me think about array creation routines, I wrote up some ideas here: https://github.com/zarr-developers/zarr-python/issues/2083