HubSpot's AI Infrastructure Behind 20 Billion Vectors
HubSpot manages 20+ billion vectors with its Qdrant-based VaaS platform, evolving from Helm to a Kubernetes Operator architecture for scale and reliability.
HubSpot built a centralized vector storage and search platform called Vector as a Service (VaaS) to power dozens of use cases—from RAG and agentic workflows to record deduplication—used by more than 38 teams. Built on the open-source Qdrant database, VaaS provides access control, embedding generation, data versioning, and feedback collection behind a single API. HubSpot runs Qdrant self-managed rather than in the cloud to benefit from deep integration with internal tooling, tighter security controls, and better cost efficiency at corporate AWS rates.
Starting from a small 2023 proof-of-concept, the system now spans 5 regions and 140+ clusters hosting 200+ indexes and over 20 billion vectors, with the largest single index holding 9.5 billion vectors. The platform sustains 5,000+ write requests per second (spiking to 100,000 RPS) and 1,000+ read RPS across all regions.
As scale grew, the original Helm-based manual cluster management became untenable, prompting a shift to the Kubernetes Operator pattern via HubSpot's internal Kube-operators framework. Three translator components—Cluster, Qdrant nodeset, and Indexer nodeset—continuously reconcile each cluster's namespaces, StatefulSets, Kafka topics, and shard management against a desired state, cutting operational overhead while improving reliability and elastic scaling.
The case offers engineers building large-scale semantic search systems concrete lessons on vector database selection and how Kubernetes operator patterns support the leap from proof-of-concept to enterprise-grade scale.