Space Operation

http://master_server is the master service, $db_name is the name of the created database, $space_name is the name of the created tablespace.

Create Space

  1. curl -XPUT -H "content-type: application/json" -d'
  2. {
  3. "name": "space1",
  4. "partition_num": 1,
  5. "replica_num": 1,
  6. "engine": {
  7. "index_size": 70000,
  8. "id_type": "String",
  9. "retrieval_type": "IVFPQ",
  10. "retrieval_param": {
  11. "ncentroids": 256,
  12. "nsubvector": 32
  13. }
  14. },
  15. "properties": {
  16. "field1": {
  17. "type": "keyword"
  18. },
  19. "field2": {
  20. "type": "integer"
  21. },
  22. "field3": {
  23. "type": "float",
  24. "index": true
  25. },
  26. "field4": {
  27. "type": "string",
  28. "array": true,
  29. "index": true
  30. },
  31. "field5": {
  32. "type": "integer",
  33. "index": true
  34. },
  35. "field6": {
  36. "type": "vector",
  37. "dimension": 128
  38. },
  39. "field7": {
  40. "type": "vector",
  41. "dimension": 256,
  42. "format": "normalization"
  43. }
  44. }
  45. }
  46. ' http://master_server/space/$db_name/_create

Parameter description:

field name

field description

field type

must

remarks

name

space name

string

true

partition_num

partition number

int

true

replica_num

replica number

int

true

engine

engine config

json

true

properties

schema config

json

true

define space field

1、Space name not be empty, do not start with numbers or underscores, try not to use special characters, etc.

2、partition_num: Specify the number of tablespace data fragments. Different fragments can be distributed on different machines to avoid the resource limitation of a single machine.

3、replica_num: The number of copies is recommended to be set to 3, which means that each piece of data has two backups to ensure high availability of data.

engine config:

field name

field description

field type

must

remarks

index_size

slice index threshold

int

false

id_type

Unique primary key type

string

false

retrieval_type

search model

string

true

retrieval_param

model config

json

false

  1. index_size: Specify the number of records in each partition to start index creation. If not specified, no index will be created.

  2. id_type Specify primary key type, can be string or long.

  3. retrieval_type search model, now support IVFPQ,HNSW,GPU,IVFFLAT,BINARYIVF,FLAT.

IVFPQ:

field name

field description

field type

must

remarks

metric_type

computer type

string

true

L2 orInnerProduct

ncentroids

number of buckets for indexing

int

false

default 2048

nsubvector

PQ disassembler vector size

int

false

default 64

bucket_init_size

bucket init size

int

false

default 1000

bucket_max_size

max size for each bucket

int

false

default 1280000

  1. "retrieval_type": "IVFPQ",
  2. "retrieval_param": {
  3. "metric_type": "InnerProduct",
  4. "ncentroids": 2048,
  5. "nsubvector": 64
  6. }

set ivfpq with hnsw:

  1. "index_size": 2600000,
  2. "id_type": "string",
  3. "retrieval_type": "IVFPQ",
  4. "retrieval_param": {
  5. "metric_type": "InnerProduct",
  6. "ncentroids": 65536,
  7. "nsubvector": 64,
  8. "hnsw" : {
  9. "nlinks": 32,
  10. "efConstruction": 200,
  11. "efSearch": 64
  12. }
  13. }

HNSW:

field name

field description

field type

must

remarks

metric_type

computer type

string

true

L2 orInnerProduct

nlinks

Number of node neighbors

int

false

default 32

efConstruction

Composition traversal depth

int

false

default 40

  1. "retrieval_type": "HNSW",
  2. "retrieval_param": {
  3. "metric_type": "L2",
  4. "nlinks": 32,
  5. "efConstruction": 40
  6. }
  7. Note: 1. Vector storage only supports MemoryOnly
  8. 2. No training is required to create an index, and the index_size value can be greater than 0

GPU (Compiled version for GPU):

field name

field description

field type

must

remarks

metric_type

computer type

string

true

L2 orInnerProduct

ncentroids

number of buckets for indexing

int

false

default 2048

nsubvector

PQ disassembler vector size

int

false

default 64, must be a multiple of 4

  1. "retrieval_type": "GPU",
  2. "retrieval_param": {
  3. "metric_type": "InnerProduct",
  4. "ncentroids": 2048,
  5. "nsubvector": 64
  6. }

IVFFLAT:

field name

field description

field type

must

remarks

metric_type

computer type

string

true

L2 orInnerProduct

ncentroids

number of buckets for indexing

int

default

default 256

  1. "retrieval_type": "IVFFLAT",
  2. "retrieval_param": {
  3. "metric_type": "InnerProduct",
  4. "ncentroids": 256
  5. }
  6. Note: 1. The vector storage method only supports RocksDB

BINARYIVF:

field name

field description

field type

must

remarks

ncentroids

number of buckets for indexing

int

default

default 256

  1. "retrieval_type": "BINARYIVF",
  2. "retrieval_param": {
  3. "ncentroids": 256
  4. }
  5. Note: 1. The vector length is a multiple of 8

FLAT:

field name

field description

field type

must

remarks

metric_type

computer type

string

true

L2 orInnerProduct

  1. "retrieval_type": "FLAT",
  2. "retrieval_param": {
  3. "metric_type": "InnerProduct"
  4. }
  5. Note: 1. The vector storage method only supports MemoryOnly

properties config:

  1. There are four types (that is, the value of type) supported by the field defined by the table space structure: keyword, integer, float, vector (keyword is equivalent to string).

  2. The keyword type fields support index and array attributes. Index defines whether to create an index, and array specifies whether to allow multiple values.

  3. Integer, float type fields support the index attribute, and the fields with index set to true support the use of numeric range filtering queries.

  4. Vector type fields are feature fields. Multiple feature fields are supported in a table space. The attributes supported by vector type fields are as follows:

field name

field description

field type

must

remarks

dimension

feature dimension

int

true

Value is an integral multiple of the above nsubvector value

store_type

feature storage type

string

false

support MemoryOnly and RocksDB

store_param

storage parameter settings

json

false

set the memory size of data

model_id

feature plug-in model

string

false

Specify when using the feature plug-in service

  1. dimension: define that type is the field of vector, and specify the dimension size of the feature.

  2. store_type: raw vector storage type, there are the following options

“MemoryOnly”: Vectors are stored in the memory, and the amount of stored vectors is limited by the memory. It is suitable for scenarios where the amount of vectors on a single machine is not large (10 millions) and high performance requirements

“RocksDB”: Vectors are stored in RockDB (disk), and the amount of stored vectors is limited by the size of the disk. It is suitable for scenarios where the amount of vectors on a single machine is huge (above 100 millions) and performance requirements are not high.

  1. store_param: storage parameters of different store_type, it contains the following two sub-parameters

cache_size: interge type, the unit is M bytes, the default is 1024. When store_type=”RocksDB”, it indicates the read buffer size of RocksDB. The larger the value, the better the performance of reading vector. Generally set 1024, 2048, 4096 and 6144; store_type =”MemoryOnly”, cache_size is not in effect.

Scalar Index Gamma engine supports scalar index, provides the filtering function for scalar data, the opening method refers to the 2nd and 3rd in the “properties config”, and the retrieval method refers to the “filter json structure elucidation” in the “Search”

View Space

  1. curl -XGET http://master_server/space/$db_name/$space_name

View Space document number

  1. curl -XGET http://master_server/_cluster/health?db=$db_name&space=$space_name

check the doc_num field

Delete Space

  1. curl -XDELETE http://master_server/space/$db_name/$space_name

Modify cache size

  1. curl -H "content-type: application/json" -XPOST -d'
  2. {
  3. "cache_models": [
  4. {
  5. "name": "table",
  6. "cache_size": 1024,
  7. },
  8. {
  9. "name": "string",
  10. "cache_size": 1024,
  11. },
  12. {
  13. "name": "field7",
  14. "cache_size": 1024,
  15. }
  16. ]
  17. }
  18. ' http://master_server/config/$db_name/$space_name
  1. table cache size: Represents the cache size of all fixed-length scalar fields (integer, long, float, double). The default value is 512M.

  2. string cache size: Represents the cache size of all variable-length scalar fields (string). The default value is 512M.

  3. store_type is the vector field of Mmap that can modify the cache size.

Get cache size

  1. curl -XGET http://master_server/config/$db_name/$space_name
  1. store_type is the vector field of Mmap to view the cache size. Other storage methods for vector fields do not support viewing the cache size.