TechPrufer
Get started
Milvus
Milvus logoMilvus · Feature trail

Milvus Milestones and Evolution

3 posts · Apr 2023 – Dec 2025

What this trail covers

Milvus continues to evolve its architecture to address the complexities of manual sharding in vector databases. The focus remains on automated scaling, distributed architecture, and efficient query processing to enable seamless growth from millions to billions of vectors without operational burden. Key innovations include a segment-based architecture for data management and a two-layer routing system that dynamically distributes data across multiple nodes, eliminating the need for manual shard management. Milvus 2.5 introduced native hybrid search combining full-text and vector search. Milvus 2.6 delivered architectural improvements for cost and performance, including tiered storage, memory savings via RaBitQ 1-bit quantization, a redesigned full-text engine, Path Index for JSON-structured metadata, AiSAQ for billion-scale compression, Semantic + Geospatial Search with R-Tree, hybrid CAGRA+ Vamana indexing, a simplified data ingestion/retrieval workflow, and support for up to 100K collections in a single cluster.

All Milvus trails
  1. 7 Years, 2 Major Rebuilds, 40K+ GitHub Stars: The Rise of Milvus as the Leading Open-Source Vector DatabaseThis post chronicles Milvus's journey from its inception in 2017 to surpassing 40,000 GitHub stars, highlighting key architectural shifts and feature releases. It details the open-sourcing in 2019, the graduation from LF AI & Data and BigANN win in 2020-2021, the ground-up reinvention for Milvus 2.0 with a decoupled cloud-native architecture, and Zilliz's recognition by Forrester. It emphasizes the introduction of native hybrid search in Milvus 2.5 and significant cost and performance improvements in Milvus 2.6, including tiered storage, RaBitQ 1-bit quantization, a faster BM25 full-text engine, JSON Path Index, AiSAQ, hybrid GPU-CPU CAGRA indexing, and increased collection support. The post also introduces two new open-source tools: DeepSearcher for local deep research and Claude Context for AI coding assistants. It reiterates Milvus's role in powering production workloads for over 10,000 teams, delivering sub-50ms retrieval, billions of documents managed, and multi-tenant architectures.
  2. Why Manual Sharding is a Bad Idea for Vector Database And How to Fix ItThis post details the engineering challenges and operational burdens associated with manual sharding in vector databases, using pgvector as an example. It contrasts this with Milvus's automated scaling capabilities, highlighting its segment-based architecture and two-layer routing system. The post explains how Milvus distributes data dynamically across multiple nodes, enabling seamless scaling from millions to billions of vectors without manual intervention. It also touches upon future plans for dynamic shard splitting in Milvus 3.0.
  3. Milvus at Its Best: Exploring v2.2 to v2.2.6This post details the evolution of Milvus through releases v2.2 to v2.2.6, highlighting new features such as bulk inserts from files, query result pagination, RBAC, quotas and limits, collection-level TTL, DiskANN for disk-based ANNS, and data backup tools in v2.2. It also covers minor releases v2.2.1 and v2.2.2 with bug fixes and improvements like Pulsar tenant/TLS support and scheduler optimizations. v2.2.3 introduced rolling upgrades and coordinator HA. v2.2.4 added resource grouping, collection renaming, GCS support, and search API options. Finally, v2.2.6 addressed critical issues from v2.2.5, focusing on binlog recycling and DataCoord GC failures.