Exactly, that is the real issue, vectorization might be good enough in terms of performance. It doesn't seem to be mentioned in the article at all.
> It’s worth noting that converting parts of / everything to vectorized numpy might be possible for this toy library, but will be nearly impossible for the real library while making the code much less readable and modifiable, and the gains are going to be limited (here’s a partially vertorized version, which is faster but far from the results we are going to achieve).
https://github.com/ohadravid/poly-match/blob/main/poly_match...
Expecting Python engineers unable to read defacto standard numpy code but meanwhile expect everyone can read Rust.....
Not to mention that the semi-vectorized code is still suboptimal. Too many for loops despite the author clearly know they can all be vectorized.
For example instead the author can just write something like:
np.argmin(
distances[distances<=threshold]
)
np.xxx( np.xxx, np.xxx + np.xxx)
Of couse you can prelocate memory for size=3000, and append stuff in a loop. But this saves only 10ms. Too insignificant.
For the original library we did all the numpy tricks we could think of, but we really needed to do this type of exhaustive search for some of the data.
If someone wants to open a PR with a "fully optimized" numpy code, that would be very cool just for comparison :)