(feat): custom reopen with read_elem_as_dask for remote h5ad

[ilan-gold]

Not sure about the typing on reopen (with Generator[...] as before, it was complaining, but a normal Callable with return felt wrong); however, the following seems to work:

import dask.distributed as dd
import remfile
import h5py
from anndata.experimental import read_elem_as_dask

# optionally log all requests to see roughly that this is lazy https://stackoverflow.com/a/24588289/8060591
import logging
import contextlib
from http.client import HTTPConnection

def debug_all_requests():

    HTTPConnection.debuglevel = 1
    logging.basicConfig()
    logging.getLogger().setLevel(logging.DEBUG)
    requests_log = logging.getLogger("requests.packages.urllib3")
    requests_log.setLevel(logging.DEBUG)
    requests_log.propagate = True

def debug_no_requests():

    HTTPConnection.debuglevel = 0
    root_logger = logging.getLogger()
    root_logger.setLevel(logging.WARNING)
    root_logger.handlers = []
    requests_log = logging.getLogger("requests.packages.urllib3")
    requests_log.setLevel(logging.WARNING)
    requests_log.propagate = False

ADATA_URI = "https://allen-brain-cell-atlas.s3.us-west-2.amazonaws.com/expression_matrices/WMB-10Xv2/20230630/WMB-10Xv2-TH-log2.h5ad"

cluster = dd.LocalCluster(n_workers=1, threads_per_worker=1)
client = dd.Client(cluster)

file_h5 = h5py.File(remfile.File(ADATA_URI), "r")
def reopen():
   yield h5py.File(remfile.File(ADATA_URI), "r")["X"]

elem = read_elem_as_dask(file_h5["X"], reopen=reopen)
elem[:20, :].compute()

The question is whether or not we'd want to internalize this - my guess is "no." I think we should only promise support for core zarr/hdf5. The flip side then is how do we test this....good questions to grapple with!

• Towards HDF5 Cloud Storage #634 (comment) with dask
• Tests added
• Release note added (or unnecessary)

[ilan-gold]

I should note that I could not figure out how to extract the remfile.RemFile.RemFile object out of the h5py.File object. If I could, maybe no need for a reopen param...

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 84.40%. Comparing base (d7643e9) to head (abc13bd).
Report is 25 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1665      +/-   ##
==========================================
- Coverage   86.87%   84.40%   -2.47%     
==========================================
  Files          39       39              
  Lines        6033     6033              
==========================================
- Hits         5241     5092     -149     
- Misses        792      941     +149     

Files with missing lines	Coverage Δ
src/anndata/_io/specs/lazy_methods.py	100.00% <100.00%> (ø)
src/anndata/_io/specs/registry.py	95.45% <100.00%> (-0.57%)	⬇️

... and 7 files with indirect coverage changes

[flying-sheep]

The previous type was correct, the new one isn’t.

The way decorators interact with typing is that you normally type the decorated function (e.g. if it contains yield, the return type is Generator). The decorator than transforms the function from whatever it is to whatever the decorator wants.

I.e.

@contextmanager
def foo(*args: Unpack[Args]) -> Generator[Ret, None, None]: ...

is the same as

def _foo(*args: Unpack[Args]) -> Generator[Ret, None, None]: ...

foo: Callable[Args, AbstractContextManager[Ret]] = contextmanager(_foo)

(I’m not 100% sure I got the “unpack” syntax right, but you know what I mean)

[flying-sheep]

so the idea is “reopen is a callable that can be transformed into a context manager using contextlib.contextmanager.

Why not just “reopen is a callable that returns an contextlib.AbstractContextManager[StorageType]?

[ivirshup]

Do you know what the difference is between what remfile and fsspec are doing?

[ilan-gold]

@ivirshup My understanding is that fsspec + kerchunk requires writing a separate JSON index file: https://fsspec.github.io/kerchunk/reference.html#kerchunk.hdf.SingleHdf5ToZarr. remfile reads directly.

[ilan-gold]

@ivirshup any thoughts on this for 0.12?

[scverse-benchmark]

Benchmark changes

Change	Before [ccfb6e3]	After [a1be0be]	Ratio	Benchmark (Parameter)
+	79.6±0.6ms	91.0±1ms	1.14	readwrite.H5ADWriteSuite.time_write_full('pbmc3k')

Comparison: https://github.com/scverse/anndata/compare/ccfb6e32806626765725ea9d1cec6eeede3176e3..a1be0be92a843312c3e7e500031ca5dfdfd1c218
Last changed: Tue, 1 Apr 2025 14:24:20 +0000

More details: https://github.com/scverse/anndata/pull/1665/checks?check_run_id=39778000357

[ilan-gold]

As an aside here: even if we use fsspec + hfpy i.e., fsspec's file object, this sort of thing is still necessary. So the need for the PR remains