Scaling Tweet Reply Ranking to Handle Tens of Millions of QPS
The Tweet Detail page on Twitter shows ranked replies to a particular tweet. With various product and ranking improvements, this surface has been seeing organic growth in usage. Furthermore, when external websites embed viral tweets, the reply serving service experiences a sharp increase in traffic. With a fixed time budget for serving a response, we developed a Light Ranking module that incorporates various Machine Learning signals and system performance signals to adaptively cut down the low-quality candidates under higher system load, allowing the service to reliably handle tens of millions of QPS.
Rishabh Misra is an ML Engineer at Twitter, Inc, and co-author of the book "Sculpting Data for ML". He combines his past engineering experiences in designing large-scale systems, working at Amazon and Arcesium (a D.E. Shaw company), and research experiences in Applied Machine Learning, with publications at top venues, to develop distributed Machine Learning relevance systems as part of the Content Quality team. In his downtime, he enjoys watching sci-fi shows, gaming, and spending time with his family.