Skip to content

perf(search): share a thread-local RNG for HNSW and drop seed (#2398) #3562

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
advisedy wants to merge 8 commits into
base: unstable
Choose a base branch
from
+11 −13
Open

perf(search): share a thread-local RNG for HNSW and drop seed (#2398) #3562

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
3e3b1aa
perf(search): lazily initialize HNSW random generator
advisedy Jul 20, 2026
5cfd455
fix(search): only call random_device when seed is not set
advisedy Jul 20, 2026
6c3ad7d
Merge branch 'unstable' into #2398
advisedy Jul 20, 2026
d53751b
Merge branch 'unstable' into #2398
jihuayu Jul 21, 2026
3b7a688
perf(search): use a static thread-local RNG for HNSW indexes
advisedy Jul 23, 2026
ff60ae2
perf(search): drop seed, make the RNG a class member
advisedy Jul 23, 2026
859bbaf
Merge branch 'unstable' into #2398
advisedy Jul 23, 2026
03c941f
fix(search): mark RandomizeLayer and InsertVectorEntry as const
advisedy Jul 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 7 additions & 3 deletions src/search/hnsw_indexer.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,16 +173,20 @@ StatusOr<double> ComputeSimilarity(const VectorItem& left, const VectorItem& rig
}

HnswIndex::HnswIndex(const SearchKey& search_key, HnswVectorFieldMetadata* vector, engine::Storage* storage,
std::random_device::result_type seed)
std::optional<std::random_device::result_type> seed)
: search_key(search_key),
metadata(vector),
storage(storage),
generator(std::mt19937(seed)),
seed(seed),
m_level_normalization_factor(1.0 / std::log(metadata->m)) {}

uint16_t HnswIndex::RandomizeLayer() {
if (!generator) {
const auto actual_seed = seed ? *seed : std::random_device()();
generator = std::make_unique<std::mt19937>(actual_seed);

@git-hulk git-hulk Jul 23, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

From the original issue, it should expect to keep only one random generator for all HNSW indexes? @PragmaTwice

@PragmaTwice PragmaTwice Jul 23, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yup. maybe thread local to avoid race.

@advisedy advisedy Jul 23, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

switched to a single static thread_local generator as you suggested. please take a look.

}
std::uniform_real_distribution<double> level_dist(0.0, 1.0);
double r = level_dist(generator);
double r = level_dist(*generator);
double log_val = -std::log(r);
double layer_val = log_val * m_level_normalization_factor;
return static_cast<uint16_t>(std::floor(layer_val));
Expand Down
7 changes: 5 additions & 2 deletions src/search/hnsw_indexer.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,8 @@

#pragma once

#include <memory>
#include <optional>
#include <random>
#include <string>
#include <vector>
Expand Down Expand Up @@ -89,11 +91,12 @@ struct HnswIndex {
HnswVectorFieldMetadata* metadata;
engine::Storage* storage = nullptr;

std::mt19937 generator;
std::optional<std::random_device::result_type> seed;
std::unique_ptr<std::mt19937> generator;
double m_level_normalization_factor;

HnswIndex(const SearchKey& search_key, HnswVectorFieldMetadata* vector, engine::Storage* storage,
std::random_device::result_type seed = std::random_device()());
std::optional<std::random_device::result_type> seed = std::nullopt);

static StatusOr<std::vector<VectorItem>> DecodeNodesToVectorItems(engine::Context& ctx,
const std::vector<NodeKey>& node_key,
Expand Down
6 changes: 6 additions & 0 deletions tests/cppunit/hnsw_index_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -118,6 +118,12 @@ TEST_F(HnswIndexTest, ComputeSimilarity) {
hnsw_index->metadata->distance_metric = redis::DistanceMetric::L2;
}

TEST_F(HnswIndexTest, RandomGeneratorIsInitializedLazily) {
EXPECT_FALSE(hnsw_index->generator);
hnsw_index->RandomizeLayer();
EXPECT_TRUE(hnsw_index->generator);
}

TEST_F(HnswIndexTest, RandomizeLayer) {
constexpr size_t kSampleSize = 50000;

Expand Down
Loading