PreSight: A Vision for an Instantaneous Web

We present the position that the time for general, scalable, privacy-preserving web cache speculation is at hand. At its core, web speculation is about predicting what resources users will want. Machine learning (ML) seems like a natural fit for this predictive task. We show, however, that ML is not fundamentally necessary to achieve a ~10% performance improvement worldwide across a diverse set of sites. We also present a set of constraints that large-scale CDNs are subject to, how these constraints limit the possibilities of using ML, what we did instead to achieve our performance results, and a discussion on the place of ML in large, Internet-scale, systems.