Productionizing Semantic Search: How We Built and Scaled Vector Infrastructure at Airtable

What this post added

Airtable details their experience building and scaling a vector search platform using Milvus for production semantic search. Key contributions include: evaluating and selecting Milvus for its self-hosting, scalability, and multi-tenancy features; designing a partitioning strategy with "one base per partition" and mitigating performance degradation by capping partitions per collection and using multiple collections per cluster; selecting HNSW index for its recall and performance; detailing ingestion and query flows; and addressing operational challenges like deployment via Kubernetes CRD and Milvus operator, observability with infrastructure and service-level metrics, node rotation strategies, cold partition offloading for cost efficiency, and a data recovery plan involving re-embedding commonly used bases.