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

❌ Patch coverage is 90.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 85.61%. Comparing base (9da064f) to head (81004c5).
⚠️ Report is 2 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/anndata/_io/specs/lazy_methods.py	75.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2121      +/-   ##
==========================================
+ Coverage   85.42%   85.61%   +0.18%     
==========================================
  Files          46       46              
  Lines        7081     7083       +2     
==========================================
+ Hits         6049     6064      +15     
+ Misses       1032     1019      -13     

Files with missing lines	Coverage Δ
src/anndata/_core/aligned_df.py	96.66% <100.00%> (+0.17%)	⬆️
src/anndata/_core/merge.py	85.13% <100.00%> (+0.02%)	⬆️
src/anndata/experimental/backed/_lazy_arrays.py	91.59% <100.00%> (ø)
src/anndata/_io/specs/lazy_methods.py	95.45% <75.00%> (-0.70%)	⬇️

... and 4 files with indirect coverage changes

[scverse-benchmark]

Benchmark changes

Change	Before [e88a6c2]	After [7a35269]	Ratio	Benchmark (Parameter)
-	3.49±0.01s	629±4ms	0.18	dataset2d.Dataset2D.time_concat(<function Dataset2D.> (0), (-1,))
-	3.50±0.02s	631±0.9ms	0.18	dataset2d.Dataset2D.time_concat(<function Dataset2D.> (0), None)
-	4.12±0.2s	1.52±0.01s	0.37	dataset2d.Dataset2D.time_concat(<function Dataset2D.> (1), (-1,))
-	5.10±0.02s	1.55±0.01s	0.3	dataset2d.Dataset2D.time_concat(<function Dataset2D.> (1), None)
+	9.98±0.1ms	12.6±0.2ms	1.26	dataset2d.Dataset2D.time_getitem_slice(<function Dataset2D.> (0), (-1,))

Comparison: https://github.com/scverse/anndata/compare/e88a6c2397ccc199eb8265e075e914fc53e8abb1..7a35269b955a90f3d4407c1be3877b3a201c3d16
Last changed: Thu, 2 Oct 2025 14:24:41 +0000

More details: https://github.com/scverse/anndata/pull/2121/checks?check_run_id=51800151862

[flying-sheep]

Can you explain the idea behind the uuid? Why is this the right thing to do for this use case and not other uses of map_blocks? Is there anything in the dask docs that recommends this pattern for certain use cases?

[ilan-gold]

Can you explain the idea behind the uuid?

I just want to be 100% sure the name I create is unique. From the docs:

The key name to use for the output array. Note that this fully specifies the output key name, and must be unique. If not provided, will be determined by a hash of the arguments.

Why is this the right thing to do for this use case

I want to avoid the tokenization of the input function, which appears to be expensive in certain cases. I had originally added name everywhere we do map_blocks but this was overkill it seems and does not seem to have affected performance.

[flying-sheep]

I see! Weird that adding it everywhere didn’t affect performance if adding it here does. Maybe we can form an intuition on where it’s expensive so we don’t have to guess/measure as much?

[ilan-gold]

Requested review based on the last non-merge commit @flying-sheep because I added an optimization for something I noticed in #2156 , although I don't know why it fails for Selman (it worked for me but was slow, hence the additional commit after your approval)