Euclidean metric is much slower than Angular?

[ronxin]

I am running an experiment with 1.2 million nodes, each represented by a 500-dim vector. On the same machine, when using Angular distance, it takes about 1 minute to build a tree; whereas using Euclidean, it takes about 1 hour. Is this difference caused by Euclidean distance having to calculate an extra offset value, or other things?

[erikbern]

I never used Euclidean distance much, so the code isn't very optimized. There's some hacky stuff in create_split that's really slow for Euclidean, and it's probably not super hard to optimize if you have some time to spend on it: https://github.com/spotify/annoy/blob/master/src/annoylib.h#L193

[erikbern]

See #65 – this makes Euclidean indexes substantially faster