I tried running the benchmarks locally using this PR but couldn't actually finish running them. Can you make sure that these can complete in a reasonable amount of time? For reference, the current Unique benchmarks on main complete in about 10 seconds on my development machine. I was waiting at least two minutes before I killed the benchmark run on this PR.

You can run a "quick" version of the benchmarks you modified locally with this command:

spin bench --version --quick -t Unique

I trimmed the parameter grid (fewer nans ratios) in order to reduce the number of repetitions per case. With these changes the Unique benchmarks complete in about 6-7 minutes on my machine.

$ time spin bench --quick -t Unique
...
real    6m39.247s
user    6m16.166s
sys     0m10.816s

For large StringDType arrays, a single trial can take several seconds, so keeping the entire benchmark run to ~10 seconds while still collecting meaningful measurements is difficult.

This is run for every PR in CI, so keeping the time down is critical. The whole benchmark CI run now takes 20 minutes in this PR, where previously it took about 5-6 minutes.

@math-hiyoko I know you're more concerned with getting the performance improvement merged than this. I'll try to set aside some time to suggest how to trim down on the parameterization or data size to get a more reasonable benchmark.

Unfortunately benchmark is difficult and the numpy benchmarks are imperfect. We might have to err on the side of missing stuff just to have a benchmark suite that developers can run and iterate on in a reasonable amount of time.

If we want the ability to run more thorough benchmarks on some occasions and not others, one possible solution is to follow the model used by scikit-learn , or something similar to it. For example, they have an SKLBENCH_PROFILE environment variable that can be swapped between regular, fast, and large_scale.

I agree this would be a good direction, but I won’t implement it in this PR.
I’ll open a separate issue to track this idea.

We might have to err on the side of missing stuff just to have a benchmark suite that developers can run and iterate on in a reasonable amount of time.

I agree that prioritizing reasonable runtime is the right approach.

With the current state of the commits, the Unique benchmarks complete in under a minute on my local machine.
Do you all think this is still insufficient?

$ time spin bench --quick -t Unique
...

real    0m56.388s
user    0m53.458s
sys     0m2.464s

Benchmarking in CI is now 5m29s, as opposed to before this PR it was 3m10s - 3m40s.

Now it completes in 4m27s.
https://github.com/numpy/numpy/actions/runs/17467174382/job/49606517073?pr=29621