An approach is provided enabling efficient access to and storage of data within a database system comprising a plurality of data repository sites distributed across different physical locations. A processor receives a request for data, the request specifying a key value. The processor determines a bucket associated with one of a plurality of data repository sites of a virtual cache based on the key value, the data repository sites being in different physical locations. The association between the bucket and the one data repository site is based on probability of serving the request.