feat(search): Add HNSW encoding index & insertion/deletion algorithm

[Beihao-Zhou]

Implement Proposal at #2316

Encoding

HNSW vector field metadata encoding:

ns | FIELD_META | index name | field name -> field flag | vector_type | dimension | distance_metric | initial_cap | m | ef_construction | ef_runtime | epsilon | num_levels

HNSW node index encoding:

ns | FIELD | index name | field name | level | NODE | key -> num_neighbours | vector dimension | [vector...]

HNSW edge index encoding:

ns | FIELD | index name | field name | level | EDGE |  key1 | key2 -> (nil)

Reference for other index encoding: #2329

Future steps

• Add the plan operator and the corresponding executor
• Add expression node (i.e. SQL/RediSearch parsers) for vector search
• Modify some passes (eg. index_selection) to convert the expression node to plan operator
• Improve HnswIndex construction (Avoid HnswIndex heavy construction because of mt19937 #2398)

[Beihao-Zhou]

Hi @PragmaTwice , this PR is ready for review! :)

[PragmaTwice]

The rest looks good to me. Thank you!

Also there are still some clang-tidy issues that need to be fixed.

[PragmaTwice]

The code looks fine to me.

But there are some issues in CI:

• one unit test case failed in macOS arm64,
• some memory issues (likely use-after-free) reported by ASan/TSan.

Could you try to investigate them?

[Yangsx-1]

I wonder if there is any heuristic logic to avoid isolated cluster?

[Beihao-Zhou]

I wonder if there is any heuristic logic to avoid isolated cluster?

@Yangsx-1 Good question but not yet, the plan for the PR is just to implement the hnsw construction first so didn't think ahead that much. Also I saw KQIR cannot be enabled on cluster mode yet[1], so also didn't take this into the scope of this PR.

[1] KQIR: a query engine for Apache Kvrocks that supports both SQL and RediSearch queries

[Beihao-Zhou]

The code looks fine to me.

But there are some issues in CI:

• one unit test case failed in macOS arm64,
• some memory issues (likely use-after-free) reported by ASan/TSan.

Could you try to investigate them?

@git-hulk @PragmaTwice
The issue was caused by the ComputeSimilarity calculates the distance between VectorItem based on the HnswVectorMetadata. In the unit test, I initialized VectorItem where its vector size less than metadata->dim, so looping through the vector causes memory leak.

I changed the code with one VectorItem::Create to do this validation early. Let me know if this still looks good to you <3

CR: 224141f
Successful workflow: https://github.com/Beihao-Zhou/kvrocks/actions/runs/9914854795

[sonarqubecloud]

Quality Gate failed

Failed conditions
2 Security Hotspots
D Reliability Rating on New Code (required ≥ A)

See analysis details on SonarCloud

Catch issues before they fail your Quality Gate with our IDE extension SonarLint

[PragmaTwice]

Awesome. Thank you for your contribution!

[PragmaTwice]

Hi @Beihao-Zhou , could you also open a tracking issue to track all issues and PRs for vector search in Kvrocks?

[Beihao-Zhou]

Hi @Beihao-Zhou , could you also open a tracking issue to track all issues and PRs for vector search in Kvrocks?

Sure, will do that later today <3