Vald

Vald answers the question of how to run approximate nearest neighbor search over billions of vectors as a cloud-native service. Developed by LY Corporation, formerly Yahoo Japan, it is a distributed engine written in Go that runs on Kubernetes 1.19 or later and uses NGT, a graph and tree based ANN algorithm, at its core. The architecture splits into modular components: Agent pods that hold NGT indexes, load-balancing and filtering gateways, a Discoverer that tracks pods through the Kubernetes API, an Index Manager that coordinates automatic index rebuilds, and an optional Helm operator. This design provides horizontal scaling, automatic backups, and high availability so a node failure does not lose index data. Hosts need AVX2 support, and official client SDKs communicate over gRPC. Released under Apache 2.0 and listed in the CNCF landscape, it serves production workloads at LY Corporation and Japan Search, and suits platform teams who prefer operating vector search on their own Kubernetes clusters over managed databases.