`hdf5` dataset returns only 0's when indexed

[ilan-gold]

With the same data from #664 I was able to open a "shallow" group successfully but then the data was just 0's!

import virtualizarr

store = virtualizarr.parsers.HDFParser(group="X")("/Users/ilangold/sc_dataset.h5ad", object_store=obstore.store.LocalStore('./')) # needs fullly qualified path now?
zarr.open_group(store, mode="r", zarr_format=3) # evidently I need to specify v3 otherwise this crashes without finding an group?
assert sum(z["indices"][...]) == 0

but I can see that:

import h5py

with h5py.File("./sc_dataset.h5ad") as f:
    assert sum(f["X"]["indices"][...]) == 0

fails, as it should because there is more than just 0's located in this store!

[maxrjones]

I'm having trouble reproducing this issue. For example, the following works for me. It's possible it was incidentally fixed by a recent PR. Would you be willing to check if the issue persists for you after the latest commits to develop?

import h5py
import numpy as np
import obstore
import zarr

import virtualizarr

file_url = "file:///Users/max/Documents/Code/zarr-developers/VirtualiZarr/.vscode/sc_dataset.h5ad"
store = obstore.store.LocalStore(prefix="/Users/max/Documents/Code/zarr-developers/VirtualiZarr/.vscode/")
ms = virtualizarr.parsers.HDFParser(group="X")(file_url, object_store=store) # needs fullly qualified path now?
z = zarr.open_group(ms, mode="r", zarr_format=3) # evidently I need to specify v3 otherwise this crashes without finding an group?
assert not sum(z["indices"][...]) == 0
arr = z["indices"]

with h5py.File("/Users/max/Documents/Code/zarr-developers/VirtualiZarr/.vscode/sc_dataset.h5ad") as f:
    np.testing.assert_allclose(arr[...], f["X"]["indices"][...])

[ilan-gold]

I still get an assertion error!

Here's my env:

asttokens==3.0.0
crc32c==2.7.1
decorator==5.2.1
donfig==0.8.1.post1
executing==2.2.0
fsspec==2025.5.1
h5py==3.14.0
hdf5plugin==5.1.0
imagecodecs==2025.3.30
imagecodecs-numcodecs==2024.6.1
ipython==9.4.0
ipython-pygments-lexers==1.1.1
jedi==0.19.2
matplotlib-inline==0.1.7
numcodecs==0.16.1
numpy==2.3.1
obstore==0.7.0
packaging==25.0
pandas==2.3.1
parso==0.8.4
pexpect==4.9.0
prompt-toolkit==3.0.51
ptyprocess==0.7.0
pure-eval==0.2.3
pygments==2.19.2
python-dateutil==2.9.0.post0
pytz==2025.2
pyyaml==6.0.2
six==1.17.0
stack-data==0.6.3
traitlets==5.14.3
typing-extensions==4.14.1
tzdata==2025.2
ujson==5.10.0
universal-pathlib==0.2.6
-e file:///Users/ilangold/Projects/Theis/VirtualiZarr
wcwidth==0.2.13
xarray==2025.6.1
zarr==3.0.10

I'm running on MacOS Sequoia 15.1. Very Strange...

[maxrjones]

#683 might help here. Previously the chunk would be filled by the fill_value if a get request failed due to a FileNotFoundError, which matched the behavior in most other Zarr-Python store implementations that support remote data access (due to the write_empty_chunks=False option). #683 PR removes those allowed exceptions, since they are not appropriate for virtual stores.

[ilan-gold]

Ah rad @maxrjones things work now, even on develop. Maybe I didn't pull or something but it seems to be resolved now!