TechPrufer
Get started
Milvus
Milvus logoMilvus · Feature trail

Milvus Deployment Modes

9 posts · Mar 2020 – Apr 2025

What this trail covers

Milvus continues to evolve its deployment and configuration guidance, focusing on optimizing performance and operational control across various components. This post details critical configuration parameters within `milvus.yaml`, categorizing them into dependency, internal, and functional configurations. It provides specific recommendations for tuning `etcd`, `minio`, `mq`, `rootCoord`, `proxy`, `queryNode`, `dataCoord`, `log`, `security`, and `quotaAndLimits` to improve performance, stability, and security in production environments. Key areas of focus include managing partitions, access logs, memory-mapped I/O, segment sizes, GC tolerances, logging verbosity, authentication, and rate limiting.

All Milvus trails
  1. The Developer’s Guide to Milvus ConfigurationThis post provides a deep dive into the `milvus.yaml` configuration file, detailing critical parameters for optimizing Milvus performance and operational control. It categorizes configurations into Dependency Component Configurations (`etcd`, `minio`, `mq`), Internal Component Configurations (`rootCoord`, `proxy`, `queryNode`, `dataCoord`), and Functional Configurations (`log`, `security`, `quotaAndLimits`). Specific recommendations are given for tuning parameters like `rootCoord.maxPartitionNum`, `proxy.maxShardNum`, `dataCoord.segment.maxSize`, `log.level`, `common.security.authorizationEnabled`, and `quotaAndLimits.flushRate.collection`, offering practical values and use-case specific guidance. It also highlights the importance of proper configuration for multi-tenancy and production deployments.
  2. What Milvus Users Taught Us in 2024This post analyzes user discussions from 2024 to identify common challenges and provide guidance. It elaborates on Milvus deployment strategies, detailing the use cases and characteristics of Milvus Lite, Standalone, and Distributed deployments, and mentions Zilliz Cloud. It also highlights the importance of resource sizing tools and provides detailed explanations on choosing between in-memory, on-disk (DiskANN, MMap), and GPU indexes, as well as selecting appropriate distance metrics. The post addresses common deployment issues, including troubleshooting distributed setups and deploying on Windows. For performance, it discusses measuring metrics via the monitoring system and Milvus WebUI, observing internal states, and understanding Milvus internals. Finally, it touches upon the selection of embedding models, recommending specific models and resources.
  3. How to Deploy the Open-Source Milvus Vector Database on Amazon EKSThis post details the deployment of Milvus on AWS EKS, focusing on the distributed deployment mode. It outlines the prerequisites, considerations for Amazon MSK configuration (autoCreateTopics, message.max.bytes, replica.fetch.max.bytes, SASL/SCRAM authentication), and the process of creating an EKS cluster using eksctl. It also covers the necessary IAM roles for S3 access and load balancing, and the configuration of networking components like VPC CNI, CoreDNS, and kube-proxy. The post emphasizes the use of Amazon S3 for object storage and Amazon MSK for message storage, highlighting the importance of proper security group configurations for inter-service communication.
  4. Introducing Milvus Lite: Start Building a GenAI Application in SecondsIntroduces Milvus Lite, a lightweight, in-application vector database that runs locally within Python applications. It reuses core Milvus components for indexing and query parsing but removes elements for high scalability in distributed systems, making it suitable for laptops, Jupyter Notebooks, and edge devices. Milvus Lite integrates with AI development stacks like LangChain and LlamaIndex, enabling its use as a vector store in RAG pipelines without server setup. It is included in the Python SDK (version 2.4.3+) and uses a local SQLite file for persistence. The API is consistent with the full Milvus API, allowing for easy transition to distributed deployments. It supports basic operations like creating collections, inserting, searching, and deleting vectors, with future support for hybrid search planned. The post highlights its role in the Milvus ecosystem for all stages of development, from prototyping to production and edge computing.
  5. What Milvus version to start withThis post details the three primary open-source Milvus deployment modes: Milvus Lite, Milvus Standalone, and Milvus Cluster. Milvus Lite is presented as a lightweight, single-binary option for local development and prototyping, integrating with Colab/Jupyter and offering core vector search features. Milvus Standalone is described as a single-instance mode suitable for smaller datasets and traffic, detailing its essential components (Milvus, etcd, MinIO). Milvus Cluster is explained as a distributed mode for large-scale, enterprise-grade workloads, outlining its eight microservice components and three third-party dependencies (etcd, MinIO, Pulsar). The post provides guidance on choosing the appropriate version based on dataset size, traffic volume, scalability, and production constraints, also mentioning Zilliz Cloud as a managed option.
  7. Introducing Milvus Lite: the Lightweight Version of MilvusIntroduces Milvus Lite, a lightweight, self-contained, in-application Python library version of Milvus. It requires no external dependencies and can be integrated directly into Python applications, Colab, and Jupyter Notebooks for local development and experimentation. It is not recommended for production environments requiring high performance, availability, or scalability. Details installation via pip, usage as a Python module with `default_server.start()` and `connections.connect()`, CLI usage with `milvus-server`, debug mode, data/log persistence via `set_base_dir()` or `--data` flag, and configuration via Python APIs or CLI flags.
  8. Using Embedded Milvus to Instantly Install and Run Milvus with PythonIntroduces 'embedded Milvus' as a new Python-user-friendly installation method for Milvus 2.1. This method allows for quick installation and running of Milvus directly within Python applications without external dependencies like etcd, MinIO, or Pulsar. It is suitable for development and testing but not recommended for production or high-performance scenarios. The post details installation, starting, stopping, and provides a comparison with other Milvus deployment modes.
  9. Announcing General Availability of Milvus 2.0Milvus 2.0 GA release introduces several key engineering capabilities: entity deletion by primary key, automatic load balancing for query and data nodes, handoff tasks for improving search performance by indexing sealed growing segments, data compaction for merging small segments and cleaning logical deleted data, and support for embedded etcd and local data storage in standalone mode to reduce external dependencies. It also announces the availability of multi-language SDKs (Python, Node.js, Java, Go) and the Milvus K8s Operator for simplified deployment and management on Kubernetes. Future development will focus on string type support, TTL, ACL management, multi-memory replicas for segments, performance enhancements, broader platform support (MacOS, ARM), and embedded PyMilvus.
  10. Run Milvus 2.0 on WindowsThis post provides detailed instructions for running Milvus 2.0 on Windows, covering two primary methods: compiling from source code and installing from a binary package. For compilation, it outlines prerequisites like the MinGW64/MSYS toolchain, cloning the specific `windows-dev` branch, installing dependencies via `scripts/install_deps_msys.sh`, and compiling using `make` and `scripts/package_windows.sh`. For binary installation, it guides users to download pre-compiled releases, unzip the package, and start the necessary services (MinIO, etcd, Milvus) using provided batch scripts, followed by running a `hello_milvus.py` example.
  11. How to Modify Milvus Advanced ConfigurationsThis post details the process of modifying advanced configurations for Milvus deployed on Kubernetes. It explains how to update ConfigMap objects for parameters not directly exposed by the Helm chart, and subsequently modify Deployment YAMLs to mount these updated configurations into the pods. It provides specific examples and verification steps using `kubectl exec`. The post also mentions that future releases will consolidate configurations and improve Helm chart support.
  12. Technical Sharing:Apply Configuration Changes on Milvus 2.0 using Docker ComposeThis post provides a practical guide for applying configuration changes to Milvus 2.0 when using Docker Compose. It details the process for both standalone and cluster deployments, including downloading and modifying the `milvus.yaml` and `docker-compose.yml` files. Specific instructions are given for mapping local configuration files into the Docker containers and for changing the log file path. The post emphasizes that configuration changes must be made before restarting Milvus.
  13. What's Inside Milvus 1.0?This post details the initial release of Milvus v1.0, highlighting its CPU-only and GPU-enabled distributions, hybrid search capabilities, and support for integer vector IDs. It introduces partitions for accelerating search on large datasets, optimizations to underlying index libraries (Faiss, NMSLIB, Annoy) including IVF and FLAT, and specific optimizations for search and index building performance. The post also covers supported distance metrics, logging, the Mishards distributed solution, Prometheus compatibility for monitoring, and the default use of SQLite for metadata management (with MySQL recommendation for production).
  14. Milvus 1.0 The World's Most Popular Open-Source Vector Database Just Got BetterMilvus v1.0 is released, marking its first major release with long-term support (3+ years). Key features include support for mainstream similarity metrics, integration with advanced ANNs algorithms, scale-out capability via Mishards, and support for various AI processors. The engine development will use Golang, with C++ for ANNS components, and future distributed versions will utilize cloud components.
  15. Set Up Milvus in Google Colaboratory for Easy ML Application BuildingThis post details how to set up and run Milvus in Google Colaboratory by compiling from source code, as Docker is not supported in the Colab environment. It provides step-by-step instructions for downloading the source code, installing dependencies, building Milvus (both CPU and GPU versions), launching the server, and performing basic operations using the pymilvus Python SDK, including creating collections, inserting data, searching, and managing collections and indexes.
  16. Mishards — Distributed Vector Search in MilvusThis post introduces Mishards, a distributed solution for Milvus designed to handle datasets larger than a standalone instance. It details the distributed architecture, including service discovery, load balancing, Mishards nodes (stateless proxy for request breaking and merging), write-only Milvus nodes (single point of failure, not scalable), and read-only Milvus nodes (stateful, scalable for search). It also covers shared storage, metadata service (MySQL), service discovery frameworks (ZooKeeper, etcd, Consul), load balancing strategies, and tracing (OpenTracing, Jaeger, Zipkin) for debugging distributed systems. The post highlights the current limitations of Mishards, such as latency costs, single-point-of-failure for writable nodes, and deployment complexity.