“On the Design and Scalability of Distributed Shared-Data Databases” (2015)
Summary: This paper is motivated by the fact, that distributed shared-nothing architectures are not flexible enough for cloud deployments. In particular the scalability as well as the elasticity w.r.t. storage and compute are limited especially with non-partition workloads. Therefore Loesing et al. propose a different architecture and provide a prototypical implementation, called Tell. Tell is based on two design choices contrary to classical distributed databases.
- Decouple compute and storage to allow independent scale-out of compute and storage to support different workload requirements.
- Opposed to distributed partitioned databases, processing nodes do not exclusively own a partition. Instead data as well as indexes are shared across query processing nodes.
This paper answers the technical challenges which arises with the given design decisions like efficient data accesses, concurrency control, and data buffering.