Milvus与Qdrant实测对比：亿级向量下的RAG检索延迟与资源消耗

CPU: Intel(R) Xeon(R) Gold 6248R @ 3.00GHz (16核)
内存: 32GB DDR4
磁盘: NVMe SSD 1TB (顺序读 3.5GB/s)
OS: Ubuntu 22.04 LTS, Kernel 5.15.0
Docker: 26.1.1, Docker Compose v2.27.0

# 1. 下载standalone部署文件
wget https://github.com/milvus-io/milvus/releases/download/v2.4.1/milvus-standalone-docker-compose.yml -O docker-compose.yml

# 2. 修改关键配置：限制内存和并发
sed -i 's/- "19530:19530"/- "19530:19530"\n    mem_limit: 16g/' docker-compose.yml
# 注意：Milvus的etcd和minio会吃大量内存，必须单独限制

# 3. 启动并检查健康状态
docker compose up -d
sleep 30
curl http://localhost:9091/healthz  # 返回OK表示就绪

# 1. 创建数据目录
mkdir -p /opt/qdrant/storage

# 2. 运行容器，重点：用host网络避免NAT性能损耗
docker run -d \
  --name qdrant \
  --network host \
  -v /opt/qdrant/storage:/qdrant/storage \
  -e QDRANT__SERVICE__GRPC_PORT=6334 \
  qdrant/qdrant:v1.9.7

# 3. 验证gRPC接口
python3 -c "from qdrant_client import QdrantClient; c=QdrantClient(host='localhost', port=6333); print(c.get_collections())"

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

connections.connect(host='localhost', port='19530')

# 创建集合：HNSW参数必须初始化
fields = [
    FieldSchema(name="id", dtype=DataType.INT64, is_primary=True),
    FieldSchema(name="embedding", dtype=DataType.FLOAT_VECTOR, dim=768)
]
schema = CollectionSchema(fields, "product_embedding")
col = Collection("product_recall", schema, consistency_level="Strong")

# 关键：HNSW参数设置，M=16, efConstruction=200
index_params = {
    "index_type": "HNSW",
    "metric_type": "IP",
    "params": {"M": 16, "efConstruction": 200}
}
col.create_index("embedding", index_params)

# 批量写入 200万条，batch_size=10000
import random
for i in range(200):
    batch = [[i*10000+j for j in range(10000)], 
             [[random.random() for _ in range(768)] for _ in range(10000)]]
    col.insert(batch)
col.flush()
print(f"实体数量: {col.num_entities}")

from qdrant_client import QdrantClient
from qdrant_client.models import VectorParams, Distance, HnswConfigDiff

client = QdrantClient(host='localhost', port=6333)

# 创建集合：同样用HNSW，但Qdrant的配置更细
client.recreate_collection(
    collection_name="product_recall",
    vectors_config=VectorParams(size=768, distance=Distance.DOT),
    hnsw_config=HnswConfigDiff(
        m=16,                # 与Milvus对齐
        ef_construct=200,    # 与Milvus对齐
        full_scan_threshold=10000  # 低于1万条时全扫描
    )
)

# 批量上传，Qdrant的批量接口是异步的
from qdrant_client.models import PointStruct
import random

points = []
for i in range(2000000):
    points.append(PointStruct(id=i, vector=[random.random() for _ in range(768)]))
    if len(points) == 10000:
        client.upsert(collection_name="product_recall", points=points, wait=True)
        points = []
if points:
    client.upsert(collection_name="product_recall", points=points, wait=True)
print("Qdrant 写入完成")