This document summarizes the architecture of Allegro's recommendation system. It uses a lambda architecture with Apache Spark, Kafka, Cassandra and Elasticsearch. Over 78 million user events are processed daily to identify 18 million interesting items and group them into 210,000 meta items and 740,000 clusters. Collaborative filtering with Apache Mahout identifies similar items using implicit feedback to address cold start problems. Infrastructure is deployed on the cloud with monitoring, logging and automated testing.