Milvus says Loon is a new storage engine for Milvus and Zilliz Vector Lakebase, designed to manage vector datasets that change continuously across search, indexing, backfills, and external compute workflows.
Milvus has introduced Loon, a storage engine for Milvus and Zilliz Vector Lakebase that the company says is designed for AI datasets that are constantly changing.
In a Milvus blog post titled “Why We Built Loon: a Storage Engine for AI Data That Never Stops Changing,” the project describes Loon as a new storage engine for Milvus and Zilliz Vector Lakebase. The stated goal is to handle vector datasets that do not stay static after ingestion, including data that is updated, reindexed, backfilled, searched online, and accessed by external compute systems.
Milvus says this workload creates requirements that are different from traditional database storage. Vector data used in retrieval systems may need to support low-latency online search while also serving background tasks such as index building, data correction, and batch processing. According to the Milvus post, Loon is intended to provide a shared storage foundation for these changing states rather than treating them as separate data copies.
Milvus says Loon uses hybrid file formats to support multiple access patterns. The company’s description points to a design that must serve vector search, scalar filtering, indexing, and external analytics without forcing every operation into a single layout.
The Milvus post also highlights row ID alignment as part of the design. In practical terms, this means that vectors, metadata, and other stored representations can be kept in sync through a shared identity model. Milvus presents this as important for systems that need to combine online search results with filters, updates, and background maintenance operations.
Another element described by Milvus is versioned manifests. The company says these manifests help coordinate different views of data as changes occur. That matters for systems where online queries, index builds, and external compute jobs may be reading or producing data at different times.
Zilliz, the company behind Milvus, has separately described Vector Lakebase as the next evolution of Zilliz Cloud. In a Zilliz blog post titled “From Vector Database to Vector Lakebase,” the company says the product direction is based on unified lake-native storage built on Vortex, and is intended to support real-time retrieval, iterative discovery, and batch analytics.
That framing helps explain why Milvus is positioning Loon as more than an internal storage optimization. The storage engine is being described as a foundation for workloads that span database-style retrieval and lake-style data processing. Zilliz’s Vector Lakebase messaging emphasizes a broader environment for AI data, where vector search is only one of several operations applied to the same underlying data.
Milvus documentation also links this direction to the Milvus 3.0-beta release. The Milvus release notes state that Milvus 3.0 is the core kernel of Zilliz Lakebase and adds open lake ecosystem integration, including External Collection and Snapshot features.
Those release notes suggest that Milvus is moving toward tighter integration between vector database workloads and open data lake environments. The Loon announcement fits that direction by focusing on storage consistency, versioning, and access from systems beyond the serving path.
The Loon announcement reflects a wider infrastructure problem in AI applications: vector datasets are often live assets rather than static indexes. Documents are added, embeddings are refreshed, metadata changes, and indexes may need to be rebuilt while applications continue serving queries.
Milvus’s argument is that storage engines for this environment need to coordinate online retrieval, background processing, and external access without making each task maintain its own disconnected copy of the data. The company says Loon addresses that with hybrid formats, aligned row IDs, and versioned manifests across Milvus and Zilliz Vector Lakebase.
The claims are still vendor claims from Milvus and Zilliz, and the cited materials do not provide independent performance benchmarks. But the announcement is a notable sign of how vector database vendors are adapting their storage designs as retrieval systems become part of larger AI data platforms.
Milvus has introduced Loon, a storage engine for Milvus and Zilliz Vector Lakebase that the company says is designed for AI datasets that are constantly changing.
The stated goal is to handle vector datasets that do not stay static after ingestion, including data that is updated, reindexed, backfilled, searched online, and accessed by external compute systems.
Milvus says this workload creates requirements that are different from traditional database storage.
Continue reading
