Salesforce's engineering team shares how they evolved the Conversation Storage Service (CSS) from handling 10,000 to 100,000 concurrent AI-driven conversations. The journey involved migrating from a PostgreSQL-based transactional system to a horizontally scaled NoSQL database, introducing Kafka with conversation-level partitioning to smooth traffic spikes, and adding VegaCache to restore read-after-write consistency masked by asynchronous streaming delays. Additional optimizations include payload compression, pagination, back-pressure controls, and a metadata-driven integration layer for downstream systems like Data 360 and AI pipelines. The architecture now serves as the backbone for Salesforce's Unified Agentic Communication Platform (UACP).

6m read timeFrom engineering.salesforce.com
Post cover image

Sort: