No abstract

Users of websites such as Facebook, Ebay and Yahoo! demand fast response times, and these sites replicate data across globally distributed datacenters to achieve this. However, it is not necessary to replicate all data to all locations: if a European user's record is never accessed in Asia, it does not make sense to pay the bandwidth and disk costs to maintain an Asian replica. In this paper, we describe mechanisms for selectively replicating large-scale web databases on a record-by-record basis. We introduce a flexible constraint language to specify replication policy constraints. We then present an adaptive scheme for replicating data to where it is most frequently accessed, while respecting policy constraints and using minimal bookkeeping. Experiments using a modified version of our PNUTS system demonstrate our techniques work well.