test: add Dask chunk grid benchmark scaffold

[ehsanestaji]

Summary

This adds an exploratory benchmark scaffold for #2036 so we can compare virtual Dask chunk choices against HDF5/Zarr on-disk chunk layouts before changing AnnData defaults.

The benchmark runner:

• creates dense X arrays with controlled HDF5/Zarr chunks and optional Zarr v3 shards
• reads X lazily through anndata.experimental.read_elem_lazy
• varies virtual Dask chunks, worker counts, and thread/process settings
• records runtime/package metadata, store size, task count, elapsed time, result shape/size, and coarse process/worker RSS readings
• includes array-level workloads and a Scanpy-style scanpy_normalize_log1p workload

This also adds a small notebook for summarizing the generated CSV and README instructions for smoke/larger-grid runs. Generated benchmark outputs are ignored under benchmarks/results.

Local signal

A modest local grid (3000x800, HDF5/Zarr, on-disk chunks 250x800 and 1000x800, default vs 1000x-1, 1/2 workers, sum_axis0 and scanpy_normalize_log1p) produced 32 rows. For small on-disk chunks (250x800), 1000x-1 reduced task counts and improved timings in the 1-worker Scanpy-style case by about 1.16x for HDF5 and 1.26x for Zarr.

These numbers are only an initial local smoke signal; the intent is to make the benchmark/review path available before proposing default behavior changes.

Checks

• ruff check benchmarks/scripts/dask_chunk_grid.py tests/test_dask_chunk_grid_script.py
• .venv/bin/python -m pytest tests/test_dask_chunk_grid_script.py -q
• python3 -m json.tool benchmarks/notebooks/dask_chunk_grid_analysis.ipynb
• git diff --check
• tiny end-to-end benchmark smoke run wrote 8 CSV result rows

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.60%. Comparing base (829abb6) to head (95bf7a1).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2465   +/-   ##
=======================================
  Coverage   85.60%   85.60%           
=======================================
  Files          49       49           
  Lines        7671     7671           
=======================================
  Hits         6567     6567           
  Misses       1104     1104           

[ehsanestaji]

I updated the title to a semantic PR title. I don’t seem to have permission to add labels on this repo, but the current validation failures look like triage metadata rather than code failures. Could a maintainer please add the appropriate labels, likely no milestone and skip-gpu-ci? benchmark, type: dask array, and performance 🐌 may also fit this benchmark-only PR.

[ilan-gold]

These numbers are only an initial local smoke signal; the intent is to make the benchmark/review path available before proposing default behavior changes.

So to this end, it might make sense for you @ehsanestaji to create a separate repo with benchmarks that can be run. This repo would produce a graph/table at the end and then propose changes/reasons for those changes. Does that sound reasonable? I don't really see why anndata itself should carry this code around.

Then there would be a PR to update the defaults, link to the benchmarking effort, and perhaps write a small documentation note/page explaining things here.