Hierarchical centroid storage for DiskBBQ #132010
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
elasticsearchmachine
added
the
Team:Search Relevance
|
Pinging @elastic/es-search-relevance (Team:Search Relevance) |
|
woah! 💯 |
afoucret
pushed a commit
to afoucret/elasticsearch
that referenced
this pull request
Jul 31, 2025
This commit presents a hierarchical layer on top of the DiskBBQ centroids to reduce the number of centroids scored at search time.
smalyshev
pushed a commit
to smalyshev/elasticsearch
that referenced
this pull request
Jul 31, 2025
This commit presents a hierarchical layer on top of the DiskBBQ centroids to reduce the number of centroids scored at search time.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This commit presents a hierarchical layer on top of the DiskBBQ centroids. The idea is to run the k-means algorithm on the generated centroids to get a centroid parent layer that can be used at search time to prevent having to score every single centroid in a neighbour queue. This implementation uses a fix of 16 centroids per parent cluster.
At search time, we will score all parents. In order to keep the same recall as of now, this PR uses a fix length queue of children centroids. The size is currently define by a percentage of the number of centroids, which is set to 10%. Therefore, we will process the parent centroid queue until we fill the children queue.
Once the children queue is full. we start visiting the posting lists in order. Note that whenever e remove a centroid from the children posting list, we add a new centroid from the parent list, so we always have a fix number of centroids scored.
We observe with this approach that we are doing 75% - 80% less centroid operations while keeping the recall.