perf: use name in map_blocks to bypass tokenization

[ilan-gold]

• Closes ad.concat is slow on lazy data on account of tokenize #1989
• Tests added (benchmark)
• Release note added (or unnecessary)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.44%. Comparing base (52344db) to head (39cb3c3).
⚠️ Report is 3 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2121      +/-   ##
==========================================
- Coverage   84.58%   84.44%   -0.14%     
==========================================
  Files          46       46              
  Lines        7105     7106       +1     
==========================================
- Hits         6010     6001       -9     
- Misses       1095     1105      +10     

Files with missing lines	Coverage Δ
src/anndata/_core/merge.py	85.11% <100.00%> (+0.02%)	⬆️

... and 2 files with indirect coverage changes

[scverse-benchmark]

Benchmark changes

Change	Before [52344db]	After [39cb3c3]	Ratio	Benchmark (Parameter)
-	647±2ms	239±3ms	0.37	dataset2d.Dataset2D.time_concat('h5ad', (-1,), 'cat')
-	651±3ms	209±0.7ms	0.32	dataset2d.Dataset2D.time_concat('h5ad', None, 'cat')
-	988±100ms	499±10ms	0.51	dataset2d.Dataset2D.time_concat('zarr', (-1,), 'cat')
-	1.11±0.01s	566±20ms	0.51	dataset2d.Dataset2D.time_concat('zarr', None, 'cat')
-	2.21±0.1ms	1.96±0.01ms	0.89	dataset2d.Dataset2D.time_full_to_memory('h5ad', (-1,), 'cat')
-	4.46±0.04ms	3.43±0.2ms	0.77	dataset2d.Dataset2D.time_full_to_memory('h5ad', (-1,), 'numeric')
+	5.47±0.09ms	6.56±0.04ms	1.2	dataset2d.Dataset2D.time_getitem_bool_mask('h5ad', (-1,), 'string-array')
-	16.4±0.3ms	14.2±0.5ms	0.87	dataset2d.Dataset2D.time_getitem_slice('h5ad', None, 'numeric')
-	15.1±0.1ms	13.4±0.07ms	0.89	dataset2d.Dataset2D.time_read_lazy_default('h5ad', (-1,), 'numeric')
-	6.0	5.0	0.83	readwrite.H5ADWriteSuite.track_peakmem_write_compressed('pbmc3k')

Comparison: https://github.com/scverse/anndata/compare/52344dbb40037704f15d79bdd9329f31ed75074d..39cb3c3c7c76876342ff3c206f771f96e79a9987
Last changed: Tue, 28 Oct 2025 16:44:52 +0000

More details: https://github.com/scverse/anndata/pull/2121/checks?check_run_id=53884619945

[ilan-gold]

I considered use name for the sparse block mapping but it had no appreciable effect: https://github.com/scverse/anndata/runs/51718340360 is the result and

anndata/src/anndata/_io/specs/lazy_methods.py

Lines 172 to 178 in 59041d4

	da_mtx = da.map_blocks(
	make_chunk,
	dtype=dtype,
	chunks=chunk_layout,
	meta=memory_format((0, 0), dtype=dtype),
	name=f"{uuid.uuid4()}/{path_or_sparse_dataset}/{elem_name}-{dtype}",
	)

shows that the commit on which the benchmark was run contained the name parameter. I leave the benchmark in anyway

[ilan-gold]

Ok @flying-sheep I know you reviewed, but the PR is grossly simplified now and hopefully the becnhmark results make some sense i.e., the big change is concat