vectordb-bench 0.0.8__tar.gz → 0.0.9__tar.gz

{vectordb_bench-0.0.8 → vectordb_bench-0.0.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: vectordb-bench
-Version: 0.0.8
+Version: 0.0.9
 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-0.0.9}/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-0.0.9}/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-0.0.9}/vectordb_bench/backend/clients/milvus/milvus.py

@@ -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()
@@ -116,7 +117,7 @@ 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]:
                 log.debug("skip compaction for gpu index type.")
@@ -179,8 +180,6 @@ 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-0.0.9/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,
+}

vectordb_bench-0.0.9/vectordb_bench/backend/clients/pgvector/pgvector.py

@@ -0,0 +1,347 @@
+"""Wrapper around the Pgvector vector database over VectorDB"""
+
+import logging
+import pprint
+from contextlib import contextmanager
+from typing import Any, Generator, Optional, Tuple, Sequence
+
+import numpy as np
+import psycopg
+from pgvector.psycopg import register_vector
+from psycopg import Connection, Cursor, sql
+
+from ..api import VectorDB
+from .config import PgVectorConfigDict, PgVectorIndexConfig
+
+log = logging.getLogger(__name__)
+
+
+class PgVector(VectorDB):
+    """Use psycopg instructions"""
+
+    conn: psycopg.Connection[Any] | None = None
+    cursor: psycopg.Cursor[Any] | None = None
+
+    # TODO add filters support
+    _unfiltered_search: sql.Composed
+
+    def __init__(
+        self,
+        dim: int,
+        db_config: PgVectorConfigDict,
+        db_case_config: PgVectorIndexConfig,
+        collection_name: str = "pg_vector_collection",
+        drop_old: bool = False,
+        **kwargs,
+    ):
+        self.name = "PgVector"
+        self.db_config = db_config
+        self.case_config = db_case_config
+        self.table_name = collection_name
+        self.dim = dim
+
+        self._index_name = "pgvector_index"
+        self._primary_field = "id"
+        self._vector_field = "embedding"
+
+        # construct basic units
+        self.conn, self.cursor = self._create_connection(**self.db_config)
+
+        # create vector extension
+        self.cursor.execute("CREATE EXTENSION IF NOT EXISTS vector")
+        self.conn.commit()
+
+        log.info(f"{self.name} config values: {self.db_config}\n{self.case_config}")
+        if not any(
+            (
+                self.case_config.create_index_before_load,
+                self.case_config.create_index_after_load,
+            )
+        ):
+            err = f"{self.name} config must create an index using create_index_before_load and/or create_index_after_load"
+            log.error(err)
+            raise RuntimeError(
+                f"{err}\n{pprint.pformat(self.db_config)}\n{pprint.pformat(self.case_config)}"
+            )
+
+        if drop_old:
+            # self.pg_table.drop(pg_engine, checkfirst=True)
+            self._drop_index()
+            self._drop_table()
+            self._create_table(dim)
+            if self.case_config.create_index_before_load:
+                self._create_index()
+
+        self.cursor.close()
+        self.conn.close()
+        self.cursor = None
+        self.conn = None
+
+    @staticmethod
+    def _create_connection(**kwargs) -> Tuple[Connection, Cursor]:
+        conn = psycopg.connect(**kwargs)
+        register_vector(conn)
+        conn.autocommit = False
+        cursor = conn.cursor()
+
+        assert conn is not None, "Connection is not initialized"
+        assert cursor is not None, "Cursor is not initialized"
+
+        return conn, cursor
+
+    @contextmanager
+    def init(self) -> Generator[None, None, None]:
+        """
+        Examples:
+            >>> with self.init():
+            >>>     self.insert_embeddings()
+            >>>     self.search_embedding()
+        """
+
+        self.conn, self.cursor = self._create_connection(**self.db_config)
+
+        # index configuration may have commands defined that we should set during each client session
+        session_options: Sequence[dict[str, Any]] = self.case_config.session_param()["session_options"]
+
+        if len(session_options) > 0:
+            for setting in session_options:
+                command = sql.SQL("SET {setting_name} " + "= {val};").format(
+                    setting_name=sql.Identifier(setting['parameter']['setting_name']),
+                    val=sql.Identifier(str(setting['parameter']['val'])),
+                )
+                log.debug(command.as_string(self.cursor))
+                self.cursor.execute(command)
+            self.conn.commit()
+
+        self._unfiltered_search = sql.Composed(
+            [
+                sql.SQL("SELECT id FROM public.{} ORDER BY embedding ").format(
+                    sql.Identifier(self.table_name)
+                ),
+                sql.SQL(self.case_config.search_param()["metric_fun_op"]),
+                sql.SQL(" %s::vector LIMIT %s::int"),
+            ]
+        )
+
+        try:
+            yield
+        finally:
+            self.cursor.close()
+            self.conn.close()
+            self.cursor = None
+            self.conn = None
+
+    def _drop_table(self):
+        assert self.conn is not None, "Connection is not initialized"
+        assert self.cursor is not None, "Cursor is not initialized"
+        log.info(f"{self.name} client drop table : {self.table_name}")
+
+        self.cursor.execute(
+            sql.SQL("DROP TABLE IF EXISTS public.{table_name}").format(
+                table_name=sql.Identifier(self.table_name)
+            )
+        )
+        self.conn.commit()
+
+    def ready_to_load(self):
+        pass
+
+    def optimize(self):
+        self._post_insert()
+
+    def _post_insert(self):
+        log.info(f"{self.name} post insert before optimize")
+        if self.case_config.create_index_after_load:
+            self._drop_index()
+            self._create_index()
+
+    def _drop_index(self):
+        assert self.conn is not None, "Connection is not initialized"
+        assert self.cursor is not None, "Cursor is not initialized"
+        log.info(f"{self.name} client drop index : {self._index_name}")
+
+        drop_index_sql = sql.SQL("DROP INDEX IF EXISTS {index_name}").format(
+            index_name=sql.Identifier(self._index_name)
+        )
+        log.debug(drop_index_sql.as_string(self.cursor))
+        self.cursor.execute(drop_index_sql)
+        self.conn.commit()
+
+    def _set_parallel_index_build_param(self):
+        assert self.conn is not None, "Connection is not initialized"
+        assert self.cursor is not None, "Cursor is not initialized"
+
+        index_param = self.case_config.index_param()
+
+        if index_param["maintenance_work_mem"] is not None:
+            self.cursor.execute(
+                sql.SQL("SET maintenance_work_mem TO {};").format(
+                    index_param["maintenance_work_mem"]
+                )
+            )
+            self.cursor.execute(
+                sql.SQL("ALTER USER {} SET maintenance_work_mem TO {};").format(
+                    sql.Identifier(self.db_config["user"]),
+                    index_param["maintenance_work_mem"],
+                )
+            )
+            self.conn.commit()
+
+        if index_param["max_parallel_workers"] is not None:
+            self.cursor.execute(
+                sql.SQL("SET max_parallel_maintenance_workers TO '{}';").format(
+                    index_param["max_parallel_workers"]
+                )
+            )
+            self.cursor.execute(
+                sql.SQL(
+                    "ALTER USER {} SET max_parallel_maintenance_workers TO '{}';"
+                ).format(
+                    sql.Identifier(self.db_config["user"]),
+                    index_param["max_parallel_workers"],
+                )
+            )
+            self.cursor.execute(
+                sql.SQL("SET max_parallel_workers TO '{}';").format(
+                    index_param["max_parallel_workers"]
+                )
+            )
+            self.cursor.execute(
+                sql.SQL(
+                    "ALTER USER {} SET max_parallel_workers TO '{}';"
+                ).format(
+                    sql.Identifier(self.db_config["user"]),
+                    index_param["max_parallel_workers"],
+                )
+            )
+            self.cursor.execute(
+                sql.SQL(
+                    "ALTER TABLE {} SET (parallel_workers = {});"
+                ).format(
+                    sql.Identifier(self.table_name),
+                    index_param["max_parallel_workers"],
+                )
+            )
+            self.conn.commit()
+
+        results = self.cursor.execute(
+            sql.SQL("SHOW max_parallel_maintenance_workers;")
+        ).fetchall()
+        results.extend(
+            self.cursor.execute(sql.SQL("SHOW max_parallel_workers;")).fetchall()
+        )
+        results.extend(
+            self.cursor.execute(sql.SQL("SHOW maintenance_work_mem;")).fetchall()
+        )
+        log.info(f"{self.name} parallel index creation parameters: {results}")
+
+    def _create_index(self):
+        assert self.conn is not None, "Connection is not initialized"
+        assert self.cursor is not None, "Cursor is not initialized"
+        log.info(f"{self.name} client create index : {self._index_name}")
+
+        index_param = self.case_config.index_param()
+        self._set_parallel_index_build_param()
+        options = []
+        for option in index_param["index_creation_with_options"]:
+            if option['val'] is not None:
+                options.append(
+                    sql.SQL("{option_name} = {val}").format(
+                        option_name=sql.Identifier(option['option_name']),
+                        val=sql.Identifier(str(option['val'])),
+                    )
+                )
+        if any(options):
+            with_clause = sql.SQL("WITH ({});").format(sql.SQL(", ").join(options))
+        else:
+            with_clause = sql.Composed(())
+
+        index_create_sql = sql.SQL(
+            "CREATE INDEX IF NOT EXISTS {index_name} ON public.{table_name} USING {index_type} (embedding {embedding_metric})"
+        ).format(
+            index_name=sql.Identifier(self._index_name),
+            table_name=sql.Identifier(self.table_name),
+            index_type=sql.Identifier(index_param["index_type"]),
+            embedding_metric=sql.Identifier(index_param["metric"]),
+        )
+        index_create_sql_with_with_clause = (
+            index_create_sql + with_clause
+        ).join(" ")
+        log.debug(index_create_sql_with_with_clause.as_string(self.cursor))
+        self.cursor.execute(index_create_sql_with_with_clause)
+        self.conn.commit()
+
+    def _create_table(self, dim: int):
+        assert self.conn is not None, "Connection is not initialized"
+        assert self.cursor is not None, "Cursor is not initialized"
+
+        try:
+            log.info(f"{self.name} client create table : {self.table_name}")
+
+            # create table
+            self.cursor.execute(
+                sql.SQL(
+                    "CREATE TABLE IF NOT EXISTS public.{table_name} (id BIGINT PRIMARY KEY, embedding vector({dim}));"
+                ).format(table_name=sql.Identifier(self.table_name), dim=dim)
+            )
+            self.cursor.execute(
+                sql.SQL(
+                    "ALTER TABLE public.{table_name} ALTER COLUMN embedding SET STORAGE PLAIN;"
+                ).format(table_name=sql.Identifier(self.table_name))
+            )
+            self.conn.commit()
+        except Exception as e:
+            log.warning(
+                f"Failed to create pgvector table: {self.table_name} error: {e}"
+            )
+            raise e from None
+
+    def insert_embeddings(
+        self,
+        embeddings: list[list[float]],
+        metadata: list[int],
+        **kwargs: Any,
+    ) -> Tuple[int, Optional[Exception]]:
+        assert self.conn is not None, "Connection is not initialized"
+        assert self.cursor is not None, "Cursor is not initialized"
+
+        try:
+            metadata_arr = np.array(metadata)
+            embeddings_arr = np.array(embeddings)
+
+            with self.cursor.copy(
+                sql.SQL("COPY public.{table_name} FROM STDIN (FORMAT BINARY)").format(
+                    table_name=sql.Identifier(self.table_name)
+                )
+            ) as copy:
+                copy.set_types(["bigint", "vector"])
+                for i, row in enumerate(metadata_arr):
+                    copy.write_row((row, embeddings_arr[i]))
+            self.conn.commit()
+
+            if kwargs.get("last_batch"):
+                self._post_insert()
+
+            return len(metadata), None
+        except Exception as e:
+            log.warning(
+                f"Failed to insert data into pgvector table ({self.table_name}), error: {e}"
+            )
+            return 0, e
+
+    def search_embedding(
+        self,
+        query: list[float],
+        k: int = 100,
+        filters: dict | None = None,
+        timeout: int | None = None,
+    ) -> list[int]:
+        assert self.conn is not None, "Connection is not initialized"
+        assert self.cursor is not None, "Cursor is not initialized"
+
+        # TODO add filters support
+        result = self.cursor.execute(
+            self._unfiltered_search, (query, k), prepare=True, binary=True
+        )
+
+        return [int(i[0]) for i in result.fetchall()]

{vectordb_bench-0.0.8 → vectordb_bench-0.0.9}/vectordb_bench/backend/runner/serial_runner.py

@@ -46,11 +46,9 @@ class SerialInsertRunner:
                 del(emb_np)
                 log.debug(f"batch dataset size: {len(all_embeddings)}, {len(all_metadata)}")
 
-                last_batch = self.dataset.data.size - count == len(all_metadata)
                 insert_count, error = self.db.insert_embeddings(
                     embeddings=all_embeddings,
                     metadata=all_metadata,
-                    last_batch=last_batch,
                 )
                 if error is not None:
                     raise error