@marcbrooker This is interesting, thanks. In real life, there is also the issue of temporal locality to consider. The usual model (i.e., the independent reference model) assumes no temporal locality other than the one that naturally arises from popularity (i.e., a popular key will have inter-request times that are smaller than an unpopular key). However, a real workload may exhibit strong temporal locality even for not-so popular keys (e.g., think of a medium popularity post that generates attention during an hour but then fades away to obliviousness). So the hotness of the shards is affected by the popularity distribution and by whether or not the workload exhibit strong temporal locality or not.