Overcoming the “D” in CAP: Using Isis

The CAP theorem establishes that a cloud service can only guarantee two of {Consistency, Availability and Partition Tolerance}, motivating developers to reject transactional ACID properties. Instead, they use BASE: a methodology whereby one transforms an application into a faster and more scalable version by running it as a series of asynchronous steps that each use local data replicas (even if potentially stale), eschew locking, and are designed to tolerate unplanned failures and service launches. Along the way, consistency is substantially weakened. But CAP and BASE may not be the final word. The new Isis platform supports consistent, locally responsive cloud services. The system is scalable, highly available, and fast. Responses to client requests can be computed using purely local data, hence delays are limited only by local computational costs. Updates propagate asynchronously and map to a single IP multicast; locking is usually avoided by employing primary-copy replication, and otherwise is performed with an inexpensive token-passing scheme. The approach relaxes durability for softstate updates: this yields an “ACI and mostly D” model. Durability violations are concealed using a form of firewall.