Milvus Roadmap

🌌 Toward the Next-Gen Multimodal Database and Data Lake

Milvus Product Roadmap

Welcome to the Milvus Roadmap!

We are ushering Milvus into a new era — the next-generation multimodal database — spanning structured to unstructured data, real-time retrieval to offline analytics, and single-cluster performance to a global data lake architecture.

This roadmap outlines the core objectives for Milvus v2.6 (in progress), Milvus v3.0 (targeted for late 2026), and Milvus v3.1 (long-term development), along with the evolution plan for Vector Lake (data lake / Loon).

🧩 Milvus v2.6 (In Progress)

Timeline: Mid-2025 – End of 2025

Focus: Upgrading the data model, refactoring the streaming architecture, building hot/cold tiering capabilities, and launching the Vector Lake Prototype (v0.1).

🎯 Key Highlights

🔹 Data Model Upgrade

• Introduce a unified Tensor / StructList data type to support multi-vector embedding structures, enabling compatibility with ColBERT, CoLQwen, video, and multimodal vectors.

• Add Geo Data support, including points, regions, and spatial indexing (based on libspatial), to expand use cases in LBS and GIS.

• Support for Timestamp with Timezone data type.

🔹 StreamNode Architecture Refactor

• Rewrite the streaming ingestion pipeline to optimize incremental writes and real-time computation.

• Significantly improve concurrency performance and stability, laying the foundation for unified real-time and offline processing.

• Introduce a new message queue engine: Woodpecker.

🔹 Hot/Cold Tiering & Storage Architecture (StorageV2)

• Support dual storage formats: Parquet and Vortex, enhancing concurrency and memory efficiency.

• Implement tiered storage with automatic hot/cold data separation and intelligent scheduling.

🔹 Vector Lake Prototype (v0.1)

• Integrate with Spark / DuckDB / DataFusion via FFI, enabling offline schema evolution and KNN queries.

• Provide multimodal data visualization and a Spark ETL demo, establishing the foundational data lake architecture.

🌠 Milvus v3.0 (Targeted for Late 2026)

Timeline: Late 2025 – Early 2026

Focus: Comprehensive enhancements to search experience, schema flexibility, and unstructured data support, along with the release of Vector Lake (v0.2).

🎯 Key Highlights

🔹 Search Experience Overhaul

• Introduce More Like This (MLT) similarity search with support for searches with position or negative examples.

• Add semantic search capabilities such as highlighting and boosting.

• Support custom dictionaries and synonym tables, enabling lexical and semantic rule definitions at the Analyzer layer.

• Introduce aggregation capabilities for queries.

🔹 Multi-Tenancy & Resource Management

• Enable multi-tenant deletion, statistics, and hot/cold tiering.

• Improve resource isolation and scheduling strategies to support millions of tables in a single cluster.

🔹 Schema & Primary Key Enhancements

• Implement Global Primary Key Deduplication (Global PK Dedup) to guarantee data consistency and uniqueness.

• Support flexible schema management (adding/dropping columns, backup fill).

• Allow NULL values in vector fields.

🔹 Expanded Unstructured Data Types (BLOB / Text)

• Introduce the BLOB type, which provides native storage and referencing for binary data such as files, images, and videos.

• Introduce TEXT type, which provides enhanced full-text and content-based search capabilities.

🔹 Enterprise-Grade Capabilities

• Support Snapshot-based backup and recovery.

• Provide end-to-end tracing and audit logging.

• Implement Active-Standby High Availability (HA) across multi-cluster deployments.

🔹 Vector Lake (v0.2)

• Support TEXT / BLOB storage and multi-version snapshot management.

• Integrate Spark for offline indexing, clustering, deduplication, and dimensionality reduction tasks.

• Deliver ChatPDF cold-query and offline benchmark demos.

🪐 Milvus v3.1 (Long-Term Vision)

Timeline: Mid-2026

Focus: User-defined functions (UDF), distributed computing integration, scalar query optimization, dynamic sharding, and the official release of Vector Lake (v1.0).

🎯 Key Highlights

🔹 UDF & Distributed Computing Ecosystem

• Support User-Defined Functions (UDFs), allowing developers to inject custom logic into retrieval and computation workflows.

• Deep integration with Ray Dataset / Daft for distributed UDF execution and multimodal data processing.

🔹 Scalar Query & Local Format Evolution

• Optimize filtering and aggregation performance for scalar fields.

• Enhance expression evaluation and index-accelerated execution.

• Support in-place updates for local file formats.

🔹 Advanced Search Capabilities

• Add the following features: RankBy, OrderBy, Facet, and Fuzzy match queries.

• Enhance text retrieval with support for:

  • match_phrase_prefix

  • Completion Suggester

  • Term Suggester

  • Phrase Suggester

🔹 Dynamic Sharding & Scalability

• Enable automatic shard splitting and load balancing for seamless scaling.

• Improve global index building and ensure distributed search performance.

🔹 Vector Lake V1.0

• Deep integration with Ray / Daft / PyTorch to support distributed UDFs and Context Engineering use cases.

• Provide RAG (Retrieval-Augmented Generation) demos and import from Iceberg tables.

🤝 Co-Building the Future of Milvus

Milvus is an open-source project driven by a global community of developers.

We warmly invite all community members to help shape the next-generation multimodal database:

• 💬 Share feedback: Propose new features or optimization ideas

• 🐛 Report issues: File bugs via GitHub Issues

• 🔧 Contribute code: Submit PRs and help build core features

  • Pull requests: Contribute directly to our codebase. Whether it’s fixing bugs, adding features, or improving documentation, your contributions are welcome.

  • Development guide: Check our Contributor’s Guide for guidelines on code contributions.

• ⭐ Spread the word: Share best practices and success stories

👉 GitHub: milvus-io/milvus