Database Scalability, Elasticity, and Autonomy in the Cloud

Agrawal, D., Abbadi, A. E., Das, S., and Elmore, A. J. (2011). Database scalability, elasticity, and autonomy in the cloud - (extended abstract). In Database Systems for Advanced Applications - 16th International Conference, DASFAA 2011.

This paper, like the previous one, also discusses the challenges of designing a 'Database-as-a-service' model in the cloud. However, unlike Relational Cloud, it discusses the pros and cons of each design choice and analyzes their inherent trade-offs for a non-relational DBMS. The paper highlights and attempts to tackle 2 key challenges:

1. Traditional DBMSs cannot be scaled very easily. This is because in traditional DBMSs, the databased is always treated as a "whole" and the DBMS tries to guarantee the consistency of the entire data. This is in contrast with the popular key-value stores in which each key-value is treated as an independent unit and consistency is guaranteed for each individual key. However, the authors claim that either of these two approaches in the strict sense are not sufficient and DBMS generally requires consistency between a group of key-value pairs on a per-application and even on a time-to-time basis. Such a system can either be implemented by having multi-key atomicity in key-value stores (data-fusion) or by partitioning traditional DBMS into 'entity groups' (data-fission)
2. Traditional DBMSs are statically provisioned. Since traditional DBMSs were always intended for enterprise infrastructure, they are statically provisioned. A query plan is in some sense binds-early to the resource. The lack of automatic workload dependent management not only prevents the queries to scale-out or scale-in depending on the load, it also prevents individual DBMSs to share optimally physical resources. To this end, similar to Relational Cloud, the authors also presented a load balancer with both static and dynamic components along with Zephyr, a technique for live-migration of instances

Comments/Critiques:

It was very interesting to see that both- this paper and the relational cloud paper identify similar challenges to implement DBaaS solutions. I liked reading this paper mainly because it explained and analyzed various trade-offs in a neutral fashion and even helped me understand the Relational Cloud paper better (eg. Data Fission/Partitioning arguments). In general, the paper presents an interesting case for implementing non-relational DBMS in the cloud because they scale well and most applications don't require strong consistency semantics anyways as opposed to the Relational Cloud paper. However, a key question to answer would be that who are the customers of this pay-per-use DBaaS model? Are these small enterprises/websites or are these fairly large enterprises? If they are the former, do we really need to worry about scaling problems of individual DBMSs given that it is highly likely that their databases would easily fit in the same datacenter?