Ensure extractTrainingVectors return a list of at most MAX_PQ_TRAINING_SET_SIZE#610
Merged
tlwillke merged 3 commits intodatastax:mainfrom Feb 12, 2026
Merged
Conversation
Adding this as its own commit since the canonical implementation is shown as 1-based and I want to make it clear how I've modified it.
michaeljmarshall
requested review from
MarkWolters,
jshook and
tlwillke
as code owners
There was a problem hiding this comment.
Pull request overview
This PR optimizes the training vector extraction process in ProductQuantization by implementing Floyd's random sampling algorithm to ensure at most MAX_PQ_TRAINING_SET_SIZE vectors are selected. This replaces the previous approach of filtering all vectors with a random probability check.
Changes:
- Replaced probabilistic filtering with deterministic sampling using Floyd's algorithm
- Added special handling for cases where total vectors ≤ MAX_PQ_TRAINING_SET_SIZE
- Changed from ThreadLocalRandom to SplittableRandom with a fixed seed for reproducibility
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
jvector-base/src/main/java/io/github/jbellis/jvector/quantization/ProductQuantization.java
Show resolved
Hide resolved
jvector-base/src/main/java/io/github/jbellis/jvector/quantization/ProductQuantization.java
Show resolved
Hide resolved
tlwillke
approved these changes
Feb 12, 2026
Collaborator
tlwillke
left a comment
tlwillke
left a comment
There was a problem hiding this comment.
Ok with the deterministic sampling. Performance for extracting the vectors checks out. LGTM.
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fixes #590
Uses floyd's random sampling algorithm to select random training vectors from the
RandomAccessVectorValues. The solution has two phases. The first is to selectMAX_PQ_TRAINING_SET_SIZErandom ordinals. Then, it maps those ordinals to vectors. Here is a reference to the algorithm: https://math.stackexchange.com/questions/178690/whats-the-proof-of-correctness-for-robert-floyds-algorithm-for-selecting-a-sin.The algorithm is essentially constant time, which is an improvement on what we currently had. We will now only generate
MAX_PQ_TRAINING_SET_SIZErandom numbers instead ofravv.size()random numbers. The slight increase cost is checking a hash set for containment.This change also handles the boundary case where the vector values object has at most
MAX_PQ_TRAINING_SET_SIZE.