Reimagining Linkedins Search Stack

LinkedIn's search architecture reduces the text its ranking model must read for each search. One model converts most of a job description into a compact numerical representation ahead of time. The ranking model combines that stored representation with the query and selected fields, such as job title or location, to judge relevance. LinkedIn reports substantially higher processing throughput than its raw-text setup, with a lower score on its top-ten ranking-quality measure. The system therefore makes a measured trade-off: it preserves much of the ranking quality while sharply reducing the amount of work performed during each search.