Milvus: Billion-Scale Vector Search

Why Milvus?

Built for scale:

• Billions of vectors
• GPU acceleration
• Horizontal scaling
• S3/MinIO storage
• Kubernetes-native

Used by: Shopify, NVIDIA, Salesforce

Docker Setup

DEVELOPERbash
# Standalone (development)
docker run -d --name milvus -p 19530:19530 -p 9091:9091 \
  milvusdb/milvus:v2.3.4 milvus run standalone

Distributed Deployment

DEVELOPERyaml
# docker-compose.yml (production)
version: '3.8'

services:
  etcd:
    image: quay.io/coreos/etcd:v3.5.5

  minio:
    image: minio/minio:RELEASE.2023-03-20T20-16-18Z

  milvus-proxy:
    image: milvusdb/milvus:v2.3.4
    command: ["milvus", "run", "proxy"]
    depends_on:
      - etcd
      - minio

  milvus-querynode:
    image: milvusdb/milvus:v2.3.4
    command: ["milvus", "run", "querynode"]
    deploy:
      replicas: 3  # Scale horizontally

Python Client

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

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

# Define schema
fields = [
    FieldSchema(name="id", dtype=DataType.INT64, is_primary=True, auto_id=True),
    FieldSchema(name="embedding", dtype=DataType.FLOAT_VECTOR, dim=1536),
    FieldSchema(name="text", dtype=DataType.VARCHAR, max_length=65535),
    FieldSchema(name="metadata", dtype=DataType.JSON)
]

schema = CollectionSchema(fields=fields, description="RAG documents")

# Create collection
collection = Collection(name="documents", schema=schema)

Indexing Strategies

DEVELOPERpython
# IVF_FLAT (balanced)
index_params = {
    "index_type": "IVF_FLAT",
    "metric_type": "COSINE",
    "params": {"nlist": 1024}
}

# HNSW (faster search, more memory)
index_params = {
    "index_type": "HNSW",
    "metric_type": "COSINE",
    "params": {
        "M": 16,
        "efConstruction": 256
    }
}

# GPU index (10x faster)
index_params = {
    "index_type": "GPU_IVF_FLAT",
    "metric_type": "COSINE",
    "params": {"nlist": 2048}
}

collection.create_index(field_name="embedding", index_params=index_params)

Inserting Data

DEVELOPERpython
# Batch insert
data = [
    [embedding1, embedding2, ...],  # embeddings
    ["text1", "text2", ...],        # text
    [{"category": "A"}, {"category": "B"}, ...]  # metadata
]

collection.insert(data)
collection.flush()  # Persist to disk

Searching

DEVELOPERpython
# Load collection into memory
collection.load()

# Search
search_params = {
    "metric_type": "COSINE",
    "params": {"nprobe": 16}  # Higher = more accurate but slower
}

results = collection.search(
    data=[query_embedding],
    anns_field="embedding",
    param=search_params,
    limit=10,
    output_fields=["text", "metadata"]
)

for hits in results:
    for hit in hits:
        print(f"Score: {hit.score}, Text: {hit.entity.get('text')}")

Filtering

DEVELOPERpython
# Metadata filtering with boolean expressions
results = collection.search(
    data=[query_embedding],
    anns_field="embedding",
    param=search_params,
    limit=10,
    expr='metadata["category"] == "tech"',
    output_fields=["text", "metadata"]
)

Partitioning

Split collection for faster queries:

DEVELOPERpython
# Create partitions
collection.create_partition("partition_2024")
collection.create_partition("partition_2025")

# Insert into specific partition
collection.insert(data, partition_name="partition_2025")

# Search in specific partition only
results = collection.search(
    data=[query_embedding],
    anns_field="embedding",
    param=search_params,
    partition_names=["partition_2025"],
    limit=10
)

Time Travel

Query historical data:

DEVELOPERpython
import time

# Get timestamp before deletion
ts_before = int(time.time() * 1000)

# Delete some data
collection.delete(expr="id in [1, 2, 3]")

# Query data as it was before deletion
results = collection.search(
    data=[query_embedding],
    anns_field="embedding",
    param=search_params,
    limit=10,
    travel_timestamp=ts_before
)

Resource Groups

Isolate workloads on different nodes:

DEVELOPERpython
from pymilvus import utility

# Create resource groups
utility.create_resource_group("rg1", config={"node_num": 2})
utility.create_resource_group("rg2", config={"node_num": 1})

# Assign collection to resource group
collection.set_properties({"resource_groups": ["rg1"]})

Monitoring

DEVELOPERpython
# Collection stats
stats = collection.get_stats()
print(f"Row count: {stats['row_count']}")

# Index progress
index = collection.index()
print(f"Index state: {index.state}")

# Query metrics (Prometheus endpoint)
# http://localhost:9091/metrics

Production RAG Pipeline

DEVELOPERpython
from pymilvus import Collection, connections
import openai

connections.connect("default", host="milvus-proxy", port="19530")
collection = Collection("documents")
collection.load()

def milvus_rag(query):
    # Embed query
    query_emb = openai.Embedding.create(
        input=query,
        model="text-embedding-3-small"
    )['data'][0]['embedding']

    # Search Milvus
    results = collection.search(
        data=[query_emb],
        anns_field="embedding",
        param={"metric_type": "COSINE", "params": {"nprobe": 32}},
        limit=5,
        output_fields=["text"]
    )

    # Build context
    context = "\n\n".join([hit.entity.get('text') for hit in results[0]])

    # Generate answer
    response = openai.ChatCompletion.create(
        model="gpt-4-turbo",
        messages=[{
            "role": "user",
            "content": f"Context: {context}\n\nQuestion: {query}"
        }]
    )

    return response.choices[0].message.content

# Usage
answer = milvus_rag("What is Milvus?")

Milvus handles billion-vector scale with ease. Perfect for enterprise RAG deployments.