perf: fast path for unchunked minor axis dask sparse reindexing when concating

[ilan-gold]

Someone tried annbatch shuffling-ing the whole cellxgene census which raised this bug where our outer-join, while correct, was effectively a densification operation for CSR matrices. Why? The previous iteration basically did this for every block

import scipy.sparse as sp

mat = sp.random(10, 20, format="csr")
mat[0, 15] = 0
assert 0 in mat.data

which means 0's are actually stored. After fixing this, it took sub-7 hours, which is basically what we'd expect given our benchmark in our paper. The speedup / memory gains are confirmed on the benchmark (see below for results).

• Closes #
• Tests added

• Release note not necessary because:

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.47%. Comparing base (3aefe2e) to head (7bef9e9).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2395      +/-   ##
==========================================
- Coverage   87.43%   85.47%   -1.96%     
==========================================
  Files          49       49              
  Lines        7720     7725       +5     
==========================================
- Hits         6750     6603     -147     
- Misses        970     1122     +152     

Files with missing lines	Coverage Δ
src/anndata/_core/merge.py	84.78% <100.00%> (-8.81%)	⬇️

... and 7 files with indirect coverage changes

[scverse-benchmark]

Benchmark changes

Change	Before [30f20c8]	After [ece963b]	Ratio	Benchmark (Parameter)
-	22.1±0.06ms	20.0±1ms	0.91	dataset2d.Dataset2D.time_getitem_bool_mask('zarr', (-1,), 'cat')
-	22.0±0.2ms	19.9±1ms	0.9	dataset2d.Dataset2D.time_getitem_bool_mask('zarr', None, 'cat')
+	17.2±0.6ms	21.3±0.5ms	1.24	sparse_dataset.SparseCSRContiguousSlice.time_getitem('array', False)
+	512±50μs	576±10μs	1.13	sparse_dataset.SparseCSRContiguousSlice.time_getitem_adata('alternating', False)
-	358±0.8μs	314±2μs	0.88	sparse_dataset.SparseCSRContiguousSlice.time_getitem_adata('array', False)
-	7.09G	1.08G	0.15	sparse_dataset.SparseCSRDaskConcat.peakmem_concat_with_mem('outer')
-	536±20ms	402±5ms	0.75	sparse_dataset.SparseCSRDaskConcat.time_concat('outer')
-	6.01±0.02s	1.46±0.02s	0.24	sparse_dataset.SparseCSRDaskConcat.time_concat_with_mem('outer')

Comparison: https://github.com/scverse/anndata/compare/30f20c8588259827926df75b458d541e6b0bb434..ece963b4de3184890ad5f5747e3017cfbb6855e9
Last changed: Mon, 20 Apr 2026 13:52:10 +0000

More details: https://github.com/scverse/anndata/pull/2395/checks?check_run_id=72116855662

[flying-sheep]

Looks great! One nitpick, otherwise let’s go!