MIT, “Relational Cloud: A Database-as-a-Service for the Cloud,” CIDR, 2011. [PDF]

Divyakant Agrawal, Amr El Abbadi, Sudipto Das, Aaron J. Elmore, “Database Scalability, Elasticity, and Autonomy in the Cloud,” DASFAA, 2011. [PDF]

Relational Cloud

The key idea of the Relational Cloud project is to define the concept of transactional Database-as-a-Service (DBaaS), identify the key challenges toward materializing it, and finally to address each one individually (in separate papers). The authors identify workload awareness as the key ingredient in addressing these challenges. Since this is only an overview paper, they do not go into the details, but they do identify three high-level goals:

1. Efficient Multi-tenancy: This deals with packing databases from different tenants in a single machine while maintaining individual SLAs. The paper suggests that just creating VMs for each database is not the ideal solution due to the lack of proper isolation. Instead they propose using accurate resource models for colocation.
2. Elastic Scalability: The goal here is to dynamically scale up/down databases based on current load from one machine to more and vice versa. The challenge is partitioning the data while avoiding cross-machine dependency as much as possible.
3. Database Privacy: The last key challenge, according to the authors, is adjustable security. This will allow querying encrypted data without decryption on the cloud. Btw, this IS cool!

Database Scalability, Elasticity, and Autonomy in the Cloud

Instead of rooting for one particular solution, this paper more or less surveys the design space of DBaaS in terms of scalability, elasticity, and autonomy. The authors believe that there can be two different approaches to simultaneously achieve scalability and atomicity in databases: one is to add some level of atomicity to existing key-value stores (Data Fusion) and the other is to scale traditional transactional databases by intelligent partitioning (Data Fission). For elasticity, they argue for live migration of instances during runtime. However, it is not sure how this technique will affect SLAs. The paper also self-cites some of the systems the authors have built at different points of the design space.

Comments

I like the proposed techniques presented in both papers, specially the Relational Cloud ones, from a technical perspective. However, I’m not sure whether DBaaS will be a successful business model. Yes, Amazon and Microsoft have DBaaS or similar products, but who are using them? If they are big enough to care about performance and security isolation, they might not be comfortable in sharing DBs with random entities. If they are not, may be they can do without all the complicated solutions anyway.