Brief Announcement: Consistency and Complexity Tradeoffs for Highly-Available Multi-cloud Store

Cloud storage services are becoming increasingly popular due to their flexible deployment, convenient pay-per-use model, and little (if any) administrative overhead. Today they are being offered by ever growing number of Internet companies, such as Amazon, Google, Microsoft as well as numerous smaller providers, such as Rackspace, Nirvanix and many others. Although cloud storage providers make tremendous investments into ensuring reliability and security of the service they offer, most of them have suffered from wellpublicized outages where the integrity and/or availability of data have been compromised for prolonged periods of time. In addition, even in the absence of outages, the customers can still lose access to their data due to connectivity problems, or unexpected alterations in the service contract (data lock-in). To address these concerns, multi-cloud storage systems whereupon data is replicated across multiple cloud storage services have become a hot topic in the systems community. Despite the significant progress in building practical multi-cloud storage systems (see e.g., [1]), as of today, little is known about their fundamental capabilities and limitations. The primary challenge lies in a wide variety of the storage interfaces and consistency semantics offered by different cloud providers to their external users. For example, whereas Amazon S3 supports a simple read/write interface, other storage services also expose a selection of more advanced transactional primitives, such as conditional writes. In this paper, we outline the results of our recent study [2] that explored the space and time complexity of building reliable multi-cloud storage services.