-
Notifications
You must be signed in to change notification settings - Fork 81
introduce the roaring multiop in milli #581
base: main
Are you sure you want to change the base?
Conversation
Current benchmark is pretty bad actually. From what I understand, we’re missing two important functions. But it becomes worse when we need to merge roaring bitmaps lazily because there is too much of it, and we don’t want to collect everything in place.
|
8e4920b
to
1c93186
Compare
9c46d74
to
36e27e2
Compare
With a new batch of optimisations, the indexing part is now faster than on main (by ~5-10%);
That's cool. The search on wiki (thus, when there are a lot of words?) also improved;
Whatever, the real issue is that the search on songs lost a lot of performances on a lot of benchmarks (~20% loss with sometimes up to twice slower);
That's not acceptable, the next step is to profile milli to understand where it comes from. Here are flamegraph for the search "7000 Danses / Le Baiser / je me trompe de mots" which is two times slower on this branch. |
Thinking out loud: Could we integrate typical indexing datasets and workflows into roaring-rs performance benchmarks? IMO Preventing surprise perf regressions like this would be beneficial to both projects. |
@irevoire is it still relevant? |
Fix meilisearch/meilisearch#3363