vectordb-bench 0.0.8__tar.gz → 0.0.10__tar.gz

{vectordb_bench-0.0.8 → vectordb_bench-0.0.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: vectordb-bench
-Version: 0.0.8
+Version: 0.0.10
 Summary: VectorDBBench is not just an offering of benchmark results for mainstream vector databases and cloud services, it's your go-to tool for the ultimate performance and cost-effectiveness comparison. Designed with ease-of-use in mind, VectorDBBench is devised to help users, even non-professionals, reproduce results or test new systems, making the hunt for the optimal choice amongst a plethora of cloud services and open-source vector databases a breeze.
 Author-email: XuanYang-cn <xuan.yang@zilliz.com>
 Project-URL: repository, https://github.com/zilliztech/VectorDBBench
@@ -12,7 +12,7 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: pytz
 Requires-Dist: streamlit-autorefresh
-Requires-Dist: streamlit>=1.23.0
+Requires-Dist: streamlit!=1.34.0
 Requires-Dist: streamlit_extras
 Requires-Dist: tqdm
 Requires-Dist: s3fs
@@ -39,6 +39,7 @@ Requires-Dist: sqlalchemy; extra == "all"
 Requires-Dist: redis; extra == "all"
 Requires-Dist: chromadb; extra == "all"
 Requires-Dist: psycopg2; extra == "all"
+Requires-Dist: psycopg; extra == "all"
 Provides-Extra: qdrant
 Requires-Dist: qdrant-client; extra == "qdrant"
 Provides-Extra: pinecone
@@ -49,7 +50,7 @@ Provides-Extra: elastic
 Requires-Dist: elasticsearch; extra == "elastic"
 Provides-Extra: pgvector
 Requires-Dist: pgvector; extra == "pgvector"
-Requires-Dist: psycopg2; extra == "pgvector"
+Requires-Dist: psycopg; extra == "pgvector"
 Provides-Extra: pgvecto-rs
 Requires-Dist: psycopg2; extra == "pgvecto-rs"
 Provides-Extra: redis

{vectordb_bench-0.0.8 → vectordb_bench-0.0.10}/pyproject.toml

@@ -26,7 +26,7 @@ classifiers = [
 dependencies = [
     "pytz",
     "streamlit-autorefresh",
-    "streamlit>=1.23.0",
+    "streamlit!=1.34.0",
     "streamlit_extras",
     "tqdm",
     "s3fs",
@@ -59,13 +59,14 @@ all = [
     "redis",
     "chromadb",
     "psycopg2",
+    "psycopg",
 ]
 
 qdrant = [ "qdrant-client" ]
 pinecone = [ "pinecone-client" ]
 weaviate = [ "weaviate-client" ]
 elastic = [ "elasticsearch" ]
-pgvector = [ "pgvector", "psycopg2" ]
+pgvector = [ "pgvector", "psycopg" ]
 pgvecto_rs = [ "psycopg2" ]
 redis = [ "redis" ]
 chromadb = [ "chromadb" ]

{vectordb_bench-0.0.8 → vectordb_bench-0.0.10}/vectordb_bench/__init__.py

@@ -32,10 +32,10 @@ class config:
     LOAD_TIMEOUT_1536D_500K     = 2.5 * 3600 # 2.5h
     LOAD_TIMEOUT_1536D_5M       =  25 * 3600 # 25h
 
-    OPTIMIZE_TIMEOUT_DEFAULT    = 15 * 60   # 15min
-    OPTIMIZE_TIMEOUT_768D_1M    =  15 * 60   # 15min
-    OPTIMIZE_TIMEOUT_768D_10M   = 2.5 * 3600 # 2.5h
-    OPTIMIZE_TIMEOUT_768D_100M  =  25 * 3600 # 1.04d
+    OPTIMIZE_TIMEOUT_DEFAULT    = 30 * 60   # 30min
+    OPTIMIZE_TIMEOUT_768D_1M    =  30 * 60   # 30min
+    OPTIMIZE_TIMEOUT_768D_10M   = 5 * 3600 # 5h
+    OPTIMIZE_TIMEOUT_768D_100M  =  50 * 3600 # 50h
 
 
     OPTIMIZE_TIMEOUT_1536D_500K =  15 * 60   # 15min

{vectordb_bench-0.0.8 → vectordb_bench-0.0.10}/vectordb_bench/backend/clients/api.py

@@ -20,6 +20,7 @@ class IndexType(str, Enum):
     Flat = "FLAT"
     AUTOINDEX = "AUTOINDEX"
     ES_HNSW = "hnsw"
+    ES_IVFFlat = "ivfflat"
     GPU_IVF_FLAT = "GPU_IVF_FLAT"
     GPU_IVF_PQ = "GPU_IVF_PQ"
     GPU_CAGRA = "GPU_CAGRA"

{vectordb_bench-0.0.8 → vectordb_bench-0.0.10}/vectordb_bench/backend/clients/milvus/config.py

@@ -14,13 +14,17 @@ class MilvusIndexConfig(BaseModel):
 
     index: IndexType
     metric_type: MetricType | None = None
+    
+    @property
+    def is_gpu_index(self) -> bool:
+        return self.index in [IndexType.GPU_CAGRA, IndexType.GPU_IVF_FLAT, IndexType.GPU_IVF_PQ]
 
     def parse_metric(self) -> str:
         if not self.metric_type:
             return ""
 
-        # if self.metric_type == MetricType.COSINE:
-        #     return MetricType.L2.value
+        if self.is_gpu_index and self.metric_type == MetricType.COSINE:
+            return MetricType.L2.value
         return self.metric_type.value
 
 

{vectordb_bench-0.0.8 → vectordb_bench-0.0.10}/vectordb_bench/backend/clients/milvus/milvus.py

@@ -8,7 +8,7 @@ from typing import Iterable
 from pymilvus import Collection, utility
 from pymilvus import CollectionSchema, DataType, FieldSchema, MilvusException
 
-from ..api import VectorDB, IndexType
+from ..api import VectorDB
 from .config import MilvusIndexConfig
 
 
@@ -89,6 +89,7 @@ class Milvus(VectorDB):
         connections.disconnect("default")
 
     def _optimize(self):
+        self._post_insert()
         log.info(f"{self.name} optimizing before search")
         try:
             self.col.load()
@@ -116,9 +117,9 @@ class Milvus(VectorDB):
                     time.sleep(5)
 
             wait_index()
-            
+
             # Skip compaction if use GPU indexType
-            if self.case_config.index in [IndexType.GPU_CAGRA, IndexType.GPU_IVF_FLAT, IndexType.GPU_IVF_PQ]:
+            if self.case_config.is_gpu_index:
                 log.debug("skip compaction for gpu index type.")
             else :
                 self.col.compact()
@@ -156,6 +157,10 @@ class Milvus(VectorDB):
 
     def need_normalize_cosine(self) -> bool:
         """Wheather this database need to normalize dataset to support COSINE"""
+        if self.case_config.is_gpu_index:
+            log.info(f"current gpu_index only supports IP / L2, cosine dataset need normalize.")
+            return True
+        
         return False
 
     def insert_embeddings(
@@ -179,8 +184,6 @@ class Milvus(VectorDB):
                 ]
                 res = self.col.insert(insert_data)
                 insert_count += len(res.primary_keys)
-            if kwargs.get("last_batch"):
-                self._post_insert()
         except MilvusException as e:
             log.info(f"Failed to insert data: {e}")
             return (insert_count, e)

vectordb_bench-0.0.10/vectordb_bench/backend/clients/pgvector/config.py

@@ -0,0 +1,215 @@
+from abc import abstractmethod
+from typing import Any, Mapping, Optional, Sequence, TypedDict
+from pydantic import BaseModel, SecretStr
+from typing_extensions import LiteralString
+from ..api import DBCaseConfig, DBConfig, IndexType, MetricType
+
+POSTGRE_URL_PLACEHOLDER = "postgresql://%s:%s@%s/%s"
+
+
+class PgVectorConfigDict(TypedDict):
+    """These keys will be directly used as kwargs in psycopg connection string,
+        so the names must match exactly psycopg API"""
+
+    user: str
+    password: str
+    host: str
+    port: int
+    dbname: str
+
+
+class PgVectorConfig(DBConfig):
+    user_name: SecretStr = SecretStr("postgres")
+    password: SecretStr
+    host: str = "localhost"
+    port: int = 5432
+    db_name: str
+
+    def to_dict(self) -> PgVectorConfigDict:
+        user_str = self.user_name.get_secret_value()
+        pwd_str = self.password.get_secret_value()
+        return {
+            "host": self.host,
+            "port": self.port,
+            "dbname": self.db_name,
+            "user": user_str,
+            "password": pwd_str,
+        }
+
+
+class PgVectorIndexParam(TypedDict):
+    metric: str
+    index_type: str
+    index_creation_with_options: Sequence[dict[str, Any]]
+    maintenance_work_mem: Optional[str]
+    max_parallel_workers: Optional[int]
+
+
+class PgVectorSearchParam(TypedDict):
+    metric_fun_op: LiteralString
+
+
+class PgVectorSessionCommands(TypedDict):
+    session_options: Sequence[dict[str, Any]]
+
+
+class PgVectorIndexConfig(BaseModel, DBCaseConfig):
+    metric_type: MetricType | None = None
+    create_index_before_load: bool = False
+    create_index_after_load: bool = True
+
+    def parse_metric(self) -> str:
+        if self.metric_type == MetricType.L2:
+            return "vector_l2_ops"
+        elif self.metric_type == MetricType.IP:
+            return "vector_ip_ops"
+        return "vector_cosine_ops"
+
+    def parse_metric_fun_op(self) -> LiteralString:
+        if self.metric_type == MetricType.L2:
+            return "<->"
+        elif self.metric_type == MetricType.IP:
+            return "<#>"
+        return "<=>"
+
+    def parse_metric_fun_str(self) -> str:
+        if self.metric_type == MetricType.L2:
+            return "l2_distance"
+        elif self.metric_type == MetricType.IP:
+            return "max_inner_product"
+        return "cosine_distance"
+
+    @abstractmethod
+    def index_param(self) -> PgVectorIndexParam:
+        ...
+
+    @abstractmethod
+    def search_param(self) -> PgVectorSearchParam:
+        ...
+
+    @abstractmethod
+    def session_param(self) -> PgVectorSessionCommands:
+        ...
+
+    @staticmethod
+    def _optionally_build_with_options(with_options: Mapping[str, Any]) -> Sequence[dict[str, Any]]:
+        """Walk through mappings, creating a List of {key1 = value} pairs. That will be used to build a where clause"""
+        options = []
+        for option_name, value in with_options.items():
+            if value is not None:
+                options.append(
+                    {
+                        "option_name": option_name,
+                        "val": str(value),
+                    }
+                )
+        return options
+
+    @staticmethod
+    def _optionally_build_set_options(
+        set_mapping: Mapping[str, Any]
+    ) -> Sequence[dict[str, Any]]:
+        """Walk through options, creating 'SET 'key1 = "value1";' commands"""
+        session_options = []
+        for setting_name, value in set_mapping.items():
+            if value:
+                session_options.append(
+                    {"parameter": {
+                            "setting_name": setting_name,
+                            "val": str(value),
+                        },
+                    }
+                )
+        return session_options
+
+
+class PgVectorIVFFlatConfig(PgVectorIndexConfig):
+    """
+    An IVFFlat index divides vectors into lists, and then searches a subset of those lists that are
+    closest to the query vector. It has faster build times and uses less memory than HNSW,
+    but has lower query performance (in terms of speed-recall tradeoff).
+
+    Three keys to achieving good recall are:
+
+    Create the index after the table has some data
+    Choose an appropriate number of lists - a good place to start is rows / 1000 for up to 1M rows and sqrt(rows) for
+    over 1M rows.
+    When querying, specify an appropriate number of probes (higher is better for recall, lower is better for speed) -
+    a good place to start is sqrt(lists)
+    """
+
+    lists: int | None
+    probes: int | None
+    index: IndexType = IndexType.ES_IVFFlat
+    maintenance_work_mem: Optional[str] = None
+    max_parallel_workers: Optional[int] = None
+
+    def index_param(self) -> PgVectorIndexParam:
+        index_parameters = {"lists": self.lists}
+        return {
+            "metric": self.parse_metric(),
+            "index_type": self.index.value,
+            "index_creation_with_options": self._optionally_build_with_options(
+                index_parameters
+            ),
+            "maintenance_work_mem": self.maintenance_work_mem,
+            "max_parallel_workers": self.max_parallel_workers,
+        }
+
+    def search_param(self) -> PgVectorSearchParam:
+        return {
+            "metric_fun_op": self.parse_metric_fun_op(),
+        }
+
+    def session_param(self) -> PgVectorSessionCommands:
+        session_parameters = {"ivfflat.probes": self.probes}
+        return {
+            "session_options": self._optionally_build_set_options(session_parameters)
+        }
+
+
+class PgVectorHNSWConfig(PgVectorIndexConfig):
+    """
+    An HNSW index creates a multilayer graph. It has better query performance than IVFFlat (in terms of
+    speed-recall tradeoff), but has slower build times and uses more memory. Also, an index can be
+    created without any data in the table since there isn't a training step like IVFFlat.
+    """
+
+    m: int | None  # DETAIL:  Valid values are between "2" and "100".
+    ef_construction: (
+        int | None
+    )  # ef_construction must be greater than or equal to 2 * m
+    ef_search: int | None
+    index: IndexType = IndexType.ES_HNSW
+    maintenance_work_mem: Optional[str] = None
+    max_parallel_workers: Optional[int] = None
+
+    def index_param(self) -> PgVectorIndexParam:
+        index_parameters = {"m": self.m, "ef_construction": self.ef_construction}
+        return {
+            "metric": self.parse_metric(),
+            "index_type": self.index.value,
+            "index_creation_with_options": self._optionally_build_with_options(
+                index_parameters
+            ),
+            "maintenance_work_mem": self.maintenance_work_mem,
+            "max_parallel_workers": self.max_parallel_workers,
+        }
+
+    def search_param(self) -> PgVectorSearchParam:
+        return {
+            "metric_fun_op": self.parse_metric_fun_op(),
+        }
+
+    def session_param(self) -> PgVectorSessionCommands:
+        session_parameters = {"hnsw.ef_search": self.ef_search}
+        return {
+            "session_options": self._optionally_build_set_options(session_parameters)
+        }
+
+
+_pgvector_case_config = {
+        IndexType.HNSW: PgVectorHNSWConfig,
+        IndexType.ES_HNSW: PgVectorHNSWConfig,
+        IndexType.IVFFlat: PgVectorIVFFlatConfig,
+}