Milvus与Qdrant选型实测：千万级向量下资源占用与查询延迟对比

# docker-compose-milvus.yml
version: '3.5'
services:
  etcd:
    image: quay.io/coreos/etcd:v3.5.5
    environment:
      - ETCD_AUTO_COMPACTION_MODE=revision
      - ETCD_AUTO_COMPACTION_RETENTION=1000
      - ETCD_QUOTA_BACKEND_BYTES=4294967296
    volumes:
      - ${DOCKER_VOLUME_DIRECTORY:-.}/volumes/etcd:/etcd
  minio:
    image: minio/minio:RELEASE.2023-03-20T20-16-18Z
    environment:
      MINIO_ACCESS_KEY: minioadmin
      MINIO_SECRET_KEY: minioadmin
    volumes:
      - ${DOCKER_VOLUME_DIRECTORY:-.}/volumes/minio:/minio_data
  standalone:
    image: milvusdb/milvus:v2.3.4
    command: ["milvus", "run", "standalone"]
    environment:
      ETCD_ENDPOINTS: etcd:2379
      MINIO_ADDRESS: minio:9000
      KNOWHERE_GPU_ENABLE: "true"  # 开启GPU加速
    ports:
      - "19530:19530"
    volumes:
      - ${DOCKER_VOLUME_DIRECTORY:-.}/volumes/milvus:/var/lib/milvus
    depends_on:
      - etcd
      - minio

# docker-compose-qdrant.yml
version: '3.5'
services:
  qdrant:
    image: qdrant/qdrant:v1.9.2
    ports:
      - "6333:6333"
      - "6334:6334"  # gRPC端口
    volumes:
      - ./qdrant_storage:/qdrant/storage
    environment:
      QDRANT__STORAGE__OPTIMIZER__CPU_LIMIT: "8"  # 限制优化器CPU
      QDRANT__SERVICE__MAX_REQUEST_SIZE: "10485760"
    command: ["/qdrant/qdrant", "--config", "/qdrant/config/production.yaml"]

from pymilvus import (
    connections, CollectionSchema, FieldSchema, DataType, Collection, utility
)

# 连接
connections.connect(alias="default", host="localhost", port="19530")

# 建集合（64维，余弦距离）
fields = [
    FieldSchema(name="id", dtype=DataType.INT64, is_primary=True),
    FieldSchema(name="vector", dtype=DataType.FLOAT_VECTOR, dim=64)
]
schema = CollectionSchema(fields, description="user_behavior")
col = Collection("behavior", schema, consistency_level="Bounded")

# 批量写入（每次5000条）
import numpy as np
data = [
    [i for i in range(5000)],  # id
    np.random.rand(5000, 64).tolist()  # 向量
]
col.insert(data)
col.flush()  # 强制刷盘

# 创建索引（IVF_FLAT，nlist=1024，HNSW也可）
index_params = {
    "index_type": "IVF_FLAT",
    "metric_type": "COSINE",
    "params": {"nlist": 1024}
}
col.create_index("vector", index_params)
col.load()

# 查询（Top-10）
query_vector = np.random.rand(64).tolist()
results = col.search(
    data=[query_vector], anns_field="vector", param={"nprobe": 16},
    limit=10, output_fields=["id"]
)
print(f"Milvus查询耗时: {results[0].distance[0]:.4f} ms")

from qdrant_client import QdrantClient
from qdrant_client.models import Distance, VectorParams, PointStruct
import numpy as np

client = QdrantClient(host="localhost", port=6333, prefer_grpc=True)

# 建collection（使用HNSW索引，m=16，ef_construct=200）
client.recreate_collection(
    collection_name="behavior",
    vectors_config=VectorParams(size=64, distance=Distance.COSINE),
    hnsw_config={"m": 16, "ef_construct": 200}
)

# 批量写入（使用upload_collection，比逐条insert快5倍）
vectors = np.random.rand(5000, 64).tolist()
points = [
    PointStruct(id=i, vector=vec) for i, vec in enumerate(vectors)
]
client.upsert(collection_name="behavior", points=points)

# 查询（Top-10，ef=64）
hit = client.search(
    collection_name="behavior",
    query_vector=np.random.rand(64).tolist(),
    limit=10,
    search_params={"hnsw_ef": 64}
)
print(f"Qdrant查询耗时: {hit[0].score:.4f}")