How does `osknnqstate` file work for reducing memory when knn searching?

Plugins k-NN
, , ,
1

Versions (relevant - OpenSearch/Dashboard/Server OS/Browser):
OpenSearch 2.17.1

Describe the issue:
I’ve tested the advantages of quantization with two groups of indices:

green  open   quantized.comment_2021 U7GNCo1qStmGOf7PZksTAg   3   1     132124           60      1.8gb        975.9mb
green  open   quantized.comment_2022 Dhz8w2PpTBaAMUT_BzeG9w   3   1     101407            0      1.4gb        734.9mb
green  open   quantized.comment_2023 qfWsD3zuS1Wz16QA4G-xsw   3   1     105771            0      1.7gb        895.3mb
green  open   comment_2021           D7KJLcnNRa-7UHx52LGyEA   3   1     132124         7243      2.4gb          1.2gb
green  open   comment_2022           zCBcgeArRuyD-3zCCTlCMw   3   1     101716         1301      1.7gb        894.1mb
green  open   comment_2023           Ytf7vWPKT6OzR5LtEHoyDw   3   1     105771        21041      2.3gb          1.2gb

The only difference between them is index mapping, whether knn.mode is on_disk(for quantization) or in_memory(default).

"COMMENT": {
  "type": "nested",
  "properties": {
    "knn": {
      "mode": "on_disk",
      "space_type": "innerproduct",
      "data_type": "float",
      "dimension": 768,
      "type": "knn_vector"
    }
  }
}

"COMMENT": {
  "type": "nested",
  "properties": {
    "knn": {
      "method": {
        "engine": "faiss",
        "space_type": "l2",
        "name": "hnsw",
        "parameters": {
          "ef_construction": 128,
          "m": 16
        }
      },
      "dimension": 768,
      "type": "knn_vector"
    }
  }
}

At first, I expected that COMMENT.knn fields would be different but they aren’t. The only significant difference in stored indices,(/usr/share/opensearch/data/nodes/0/indices/D7KJLcnNRa-7UHx52LGyEA/2/index) is the existence of NativeEngines990KnnVectorsFormat_0.osknnqstate files.


  1. With these files, do OpenSearch’s KNN990QuantizationStateWriter and KNN990QuantizationStateReader help data nodes to index/search documents lighter and faster?

  2. According to Disk-based vector search - OpenSearch Documentation, Disk-based vector search uses binary quantization, compressing vectors and thereby reducing the memory requirements. But disk reduction is only 25%, just same as byte vector(8bit integer quantization).