Milvus
Zilliz
Star45.2KBook a DemoTry Managed Milvus
Home

Milvus Roadmap

🌌 Toward the Next-Gen Multimodal Database and Vector Lakebase

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 Vector Lakebase architecture.

This roadmap outlines the core objectives for Milvus v3.0 (public beta), and Milvus v3.1 (long-term development), along with the evolution plan for Zilliz Vector Lakebase.

🌠 Milvus v3.0 (Public Beta)

Public Beta: May 2026

Focus: Building a semantic-native query engine with in-engine sorting, aggregation, and multi-vector retrieval, and the lake-native foundation of Zilliz Vector Lakebase so compute reaches data without migration.

🎯 Key Highlights

🔹 Schema & Data Type Evolution

  • Support ALTER COLLECTION ADD COLUMN and DROP COLUMN at runtime without rebuilding indexes or interrupting serving.
  • Provide two backfill paths for new columns: external via Spark Connector, and internal with BM25 sparse vectors auto-generated at write time.
  • Introduce TEXT as a first-class data type that stores original text alongside vectors with BM25 and text-match support.

🔹 Query Execution Overhaul

  • Push Order By into the engine with per-segment sort and merge-sort across query nodes.
  • Add SQL-style query aggregation (GROUP BY with COUNT, SUM, AVG, MIN, MAX) computed in the kernel.
  • Introduce search facets over ANN results with per-bucket statistics and nested sub-facets server-side.
  • Support custom dictionaries and synonym tables registered cluster-side for improved CJK and domain-specific recall.

🔹 Multi-Vector & Late-Interaction Support

  • Introduce StructList to represent one entity as a single row with many vectors, with native late-interaction support (ColBERT, ColPali) via MAX_SIM.
  • Support element-level and entity-level search on StructList fields, with configurable match policies for entity-level results.
  • Add three multi-vector retrieval strategies: TokenANN (exhaustive), Muvera (projection-based, no training), and Lemur (learned compression).

🔹 Retrieval & Index Overhaul

  • Overhaul the sparse inverted index with block compression, weight quantization, and a persisted format; introduce SINDI as the default sparse IP algorithm.
  • Expand index coverage with the full Faiss family (SVS, Panorama, PQ, IVFPQ, ScaNN) and MinHash DIDO for near-duplicate detection.
  • Support nullable vector fields for async embeddings and missing modalities, with auto-filtering at search time.

🔹 Vector Lakebase Storage & Compute Architecture

  • Introduce External Collection to index and query data in S3 / GCS / Azure in place, with support for Lance, Parquet, Iceberg, and Vortex table formats.
  • Add Vortex, an open columnar format, and Loon (Storage V3), a mixed-format storage layer for efficient point reads from object storage.
  • Support point-in-time snapshots with MVCC-style isolation for batch processing while serving continues to write.
  • Integrate as a Spark DataSource v2 for reading from and writing to Milvus directly in Spark / Databricks / EMR pipelines.

🪐 Milvus v3.1 (Long-Term Vision)

Timeline: Late 2026 and beyond

Focus: Storage intelligence, write-path integrity, compute extensibility, and expanded Vector Lakebase interoperability.

🎯 Key Highlights

🔹 Storage & Write Path

  • Add predicate pushdown with page-index and bloom-filter pruning at the storage layer.
  • Implement primary-key dedup on ingest to prevent duplicates at write time.

🔹 Compute & Elasticity

  • Support User-Defined Functions (UDFs) for running custom logic in the engine, on the data plane.
  • Enable shard splitting to resplit shards as data grows, with custom sharding key support.

🔹 Spark & Vector Lakebase Expansion

  • Expand the Spark connector with a richer library of native batch operators.
  • Add table format capabilities including time-travel, schema evolution, and snapshot rollback.
  • Expand Vector Lakebase interoperability with CDC-fresh external indexes, Apache Paimon support, and additional data formats.

🤝 Co-Building the Future of Milvus

Milvus is an open-source project driven by a global community of developers. We invite all community members to help shape the next-generation multimodal database:

  • 💬 Share feedback: Propose new features or optimization ideas on GitHub Discussions.

  • 🐛 Report issues: File bugs through GitHub Issues.

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

    • Pull requests: Contribute directly to our codebase. Whether you’re fixing bugs, adding features, or improving documentation, your contributions are welcome.
    • Development guide: Check our Contributor’s Guide for guidelines on code contributions.

  • 🗣️ Join the conversation: Ask questions and meet maintainers on Discord, at Milvus Office Hours, or across all community channels.

  • ⭐ Spread the word: Share best practices and success stories, and follow Milvus on X, LinkedIn, and YouTube.

👉 GitHub: milvus-io/milvus

Table of contents

Try Managed Milvus for Free

Zilliz Cloud is hassle-free, powered by Milvus and 10x faster.

Get Started
Feedback

Was this page helpful?