Scaleable Consistency Adjustable Data Storage (SCADS)
Michael Armbrust, Armando Fox, Michael Franklin, Nick Lanham, David A. Patterson, Beth Trushkowsky and Jesse Trutna
Sun Microsystems, Google, Microsoft, Hewlett-Packard, Cisco Systems, Oracle, Cisco Systems, IBM, Fujitsu, NetApp, Siemens, VMWare and Facebook
Modern user-facing web applications such as Facebook, Flickr, Yelp, the Amazon storefront, and the various Google properties present new challenges for storing and querying data at the multiple-terabyte scale.
Generally tolerant of stale reads, such systems have requirements for response times measured in milliseconds, face availability requirements nearing 100%, and must scale under bursty and exponentially increasing data storage loads. Under these requirements, traditional usage patterns change, with ad-hoc queries against the production system becoming dangerous and and intrusive migration schemes becoming infeasible.
We believe there exists an opportunity for a highly-scalable system which provides reasoned tradeoffs between consistency, availability, and performance in this space.
We propose a new system, SCADS, which provides for the declarative specification of the consistency and performance requirements of an application, takes advantage of utility computing to provide cost-effective rapid scale-up and scale-down, pre-computes queries to decrease response time, and uses machine learning models to anticipate performance problems and predict the runtime of new queries before execution.