Abnormal AI builds OpenSearch vector search at scale

An engineer at Abnormal AI details how the company built a shared vector search system using OpenSearch to handle billions of email embeddings across US and EU clusters. The system utilizes daily indices, batch writes through Kafka, and a five-minute refresh interval to balance freshness with I/O efficiency. Production scaling required partitioned tenant routing, memory-optimized search with page-caching, and multi-tier write retries with dead-letter queues, ultimately reducing replicated store costs by roughly 44 percent compared to managed alternatives. Open article →