Current state: Under Discussion
...
Keywords: Vector Index
Released: TBD
Summary
We got requirement for supporting ScaNN index in Milvus. This proposal is about how to integrate a new index into knowhere and expose it to client.
Document about how knowhere works in milvus v1.x: m1_chain.pdf
Docement about how knowhere works in milvus v2.0: TODO
Motivation
Google recently released a new algorithm to find Approximate Nearest Neighbors called ScaNN.
...
They show through benchmarks that it can perform significantly better than the existing solutions like Annoy, FAISS, hnsw. The goal is to add support for ScaNN algorithm to the list of Milvus supported indexes.
Public Interfaces
Follwing public interfaces would be updated in knowhere =>
Code Block | ||||
---|---|---|---|---|
| ||||
namespace IndexEnum {
extern const char* INDEX_SCANN;
} |
Code Block | ||||
---|---|---|---|---|
|
...
- Faiss has released a version that outperform Scann. facebookresearch/faiss#1399
...
| |
namespace IndexEnum {
const char* INDEX_SCANN = "SCANN";
} |
Design Details
As part of third party library, we will have to add a new library for ScaNN. This directory here: https://github.com/google-research/google-research/tree/master/scann/scann/scann_ops contains APIs for build/search etc. Also, ScaNN uses Bazel as it's dependency management while we would need to use CMake.
One the library has been added,
Compatibility, Deprecation, and Migration Plan
...