HubSpot's Vector Search Platform Hits 20 Billion—and Shows Why Scale is an Ops Problem

HubSpot's semantic search infrastructure has grown from experimental prototype to a production backbone handling over 20 billion vectors across dozens of internal teams. What began as a proof of concept has evolved into VaaS (Vector as a Service)—an abstraction layer that now powers AI agents, retrieval-augmented generation, and contact deduplication across the company.

From POC to Mature Platform#

The journey reveals a familiar scaling pattern: the database itself was only the beginning. HubSpot chose Qdrant as its underlying vector store, but wrapped it with layers of custom infrastructure—access controls, embeddings generation, data versioning, and feedback collection. Running Qdrant in-house rather than through a managed service let the team integrate with internal tooling for tracing, cost tracking, rate limiting, and security, all while maintaining control over customer data.

The platform now spans more than 200 indexes, 140-plus clusters across five regions, and two environments, with write traffic peaking at 100,000 requests per second.

Operations as the Real Bottleneck#

Early infrastructure relied on Helm for cluster management, a setup that worked when the fleet was small. As usage grew, manual operations became untenable. The team migrated to a custom Kubernetes Operator framework that could do what Helm could not: make API calls, auto-scale based on external metrics, and handle stateful lifecycle decisions.

That shift introduced Translators—components that continuously reconcile desired and actual system state every 60 seconds. This automation now handles:

• Cluster creation and decommissioning
• Shard movement and replication recovery
• Horizontal scaling and rebalancing

The payoff was dramatic: cluster spin-up time dropped from hours to minutes, and the need for standby clusters vanished. More importantly, operational burden on the team plateaued even as demand kept rising.

A Lesson Across the Ecosystem#

HubSpot's experience mirrors broader patterns in vector search. Qdrant's own scaling guidance focuses on keeping latency and accuracy stable across hundreds of millions of vectors. Pinecone, which scaled to 600 million vectors, published similar priorities: maintaining recall, decoupling storage from compute, and treating reliability as a first-class concern rather than an afterthought.