Case Study: Groupon achieves dramatic search relevance and performance improvements (sub-1s latency, <0.1% fallback) with Elastic

Groupon - Customer Case Study

Groupon moved its search and ranking stack from an unsharded Solr frontend to Elasticsearch to support large-scale, personalized deal ranking across web, mobile and email. The team struggled with Elasticsearch native scripting: scripts were verbose, error-prone, limited to returning a single score, and couldn’t easily expose debug or per-request feature data. At the same time Groupon’s ML models needed rich per-user and per-deal feature vectors computed on every request (up to ~16k deals ranked/request), creating high CPU load and operational complexity.

To solve this they built a Scoring API that wraps ES scripting with clean abstractions (Scorable, ScoringFunction, Score) and request-scoped storage so models can return numbers plus debugging and meta-data, reuse caching, and be developed independently of ES internals. The change made ranking simpler and portable, cut fallback rates from 5% to <0.1%, reduced latency from ~2s to <1s, enabled ranking more deals per user and reuse of the same ranking code for email sends, and set the stage for open sourcing, dynamic model updates, and multi-stage scoring.

Groupon

Brian Humphrey

Relevance Infrastructure Engineer