Julien Jerphanion

> @jjerphan to help understand the nature of the failures found by `assert_argkmin_results_quasi_equality` it would help to report the original distances (before rounding) for the nighbor indices in the 2...

Thank you for the heads-up, @cmarmo. Yes, this can be closed: the proposed changes have been implemented. I should probably spend some time looking at the oldest issues related to...

@Micky774: Thanks for splitting this out. As of now, I think it's better that we focus on this PR to get it merged and that we then work on the...

Thanks for the benchmarks results, @Micky774. I would like to understand what's wrong. Can you provide the script you used to produce this plot? Also, generally keeping the scope of...

With the current heuristic, we get using the `time_PairwiseDistancesArgKmin` benchmark the following results (:tada:): ``` before after ratio [06e733e1] [fec7843c] - 109±20ms 82.2±1ms 0.75 pairwise_distances_reductions.PairwiseDistancesReductionsBenchmark.time_PairwiseDistancesArgKmin(1000, 1000, 100, 'manhattan', 'auto') -...

> Do you confirm that this fixes the scalability degradation observed in the plot of the description of https://github.com/scikit-learn/scikit-learn/pull/23865 when the number of CPU threads is large compared to the...

I've accepted @ogrisel's suggestion which code-wise revert this branch to fec7843c, the revision for which benchmark results' (see https://github.com/scikit-learn/scikit-learn/pull/24043#issuecomment-1206797092) were reported. I think this can be further finer-grained refined (e.g....

I would say that maintainers have not discussed and agreed in modularizing those implementations yet. Ideally, it should be modular, but I do not think that's on the agenda on...

Hi @adam2392, Thank you for this long comment. I will add your contributions to discussion to [the next core-developer meeting](https://calendar.google.com/calendar/event?eid=MjFhbDVhdDNoZ3BodXV0aDdpdGkwYnV0NWUgc29jaWFsLnNjaWtpdGxlYXJuQG0&ctz=UTC). You should be able to join it freely.

Thanks, @Micky774. I hope this will help you with future experimentations.