Case Study: Outbrain achieves scalable, low-latency personalized content recommendations with Elastic

Looking at Content Recommendation Through a Search Lens

Outbrain, a leading content-discovery platform that serves personalized recommendations to hundreds of millions of users, faced a hard engineering challenge: enforce hundreds of market rules and deliver highly personalized recommendations across massive inventory in real time with very high throughput and sub-100ms latency.

They reframed recommendation as a search problem by indexing each article as an Elasticsearch document with rich semantic features, translating user profiles into queries and market rules into filters, and adding a custom plugin for machine-learned scoring. Combined with read-only, optimized indexes (forceMerge to single segment), tmpfs storage, a small indexing cluster with a large search cluster, tuned thread/queue settings and dual node instances per server, the system now serves 800,000 requests per minute under 100 ms on 48 data nodes and supports more complex personalization models at scale.

Outbrain