Milvus

Managed Milvus, an open-source vector database for embeddings and similarity search, with standalone or clustered high-availability deployments.

Creating a Milvus instance

From the dashboard, click New service, then Database, then Milvus. Milvus is a vector database: it stores high-dimensional embeddings and finds the most similar vectors to a query, which makes it a common building block for AI features such as semantic search, recommendations, and retrieval-augmented generation (RAG).

Attach Milvus to your service from the Environment tab. When you connect it, StackBlaze injects MILVUS_URI into your service automatically, so your application can read the connection target from the environment.

Connecting

Milvus speaks gRPC and listens on port 19530. Connect over the private network using the internal hostname of your service. The injected MILVUS_URI points at this address.

MILVUS_URI (internal)

[service-name]:19530

Standard vs high availability

Standard runs Milvus in standalone mode: a single process that holds the query, data, and index roles together. It is simple and well suited to development and smaller workloads.

High availability runs Milvus in cluster mode. The query, data, and index components run as separate, independently scalable services, each with replicas. This lets the database scale out across nodes and tolerate the loss of a single node without going down. You choose Standard or HA when creating the instance.

Aspect	Standalone (Standard)	Cluster (HA)
Mode	Standalone	Cluster
Components	Single combined process	Query, data, and index run as separate services
Replicas	Single instance	Multiple replicas per component
Scaling	Vertical (resize the instance)	Horizontal (scale components out)
Node loss	Causes downtime	Tolerates loss of a node
Best for	Development and smaller workloads	Production and larger workloads

Connecting from Python

Install pymilvus and read the connection target from MILVUS_URI. The snippet below connects, creates a collection with a vector field, inserts nothing yet, and runs a similarity search.

milvus_example.py

import os
from pymilvus import MilvusClient, DataType

client = MilvusClient(uri="http://" + os.environ["MILVUS_URI"])

# Create a collection with a 768-dimensional vector field
schema = client.create_schema(auto_id=True)
schema.add_field("id", DataType.INT64, is_primary=True)
schema.add_field("embedding", DataType.FLOAT_VECTOR, dim=768)

client.create_collection("documents", schema=schema)
client.create_index(
    "documents",
    index_params=client.prepare_index_params(
        field_name="embedding",
        index_type="IVF_FLAT",
        metric_type="L2",
    ),
)
client.load_collection("documents")

# Search for the 5 nearest vectors to a query embedding
query = [[0.1] * 768]
results = client.search(
    "documents",
    data=query,
    anns_field="embedding",
    limit=5,
)
print(results)

Backups

Milvus instances are backed up automatically. Backups are encrypted and follow the standard StackBlaze backup policy, including schedule and retention. For details on how backups run, how they are encrypted, and how to restore, see Backups & recovery.

Under the hood

Milvus is an open-source vector database. StackBlaze runs it through a Kubernetes operator with persistent volumes for durable storage. In cluster mode (HA), the operator spreads Milvus components and their replicas across nodes so the database can scale out and survive the loss of a single node.