Talk on "Cosine: A Cloud-Cost Optimized NoSQL Storage Engine" by Subarna Chatterjee at the MiDAS Seminar (April 24, 2020).
Abstract:
We present a key-value storage engine, Cosine, that guarantees an optimal cost- performance trade-off, given a workload and a budget. Cosine creates a massive search space comprising of the entire data structure design and hardware space of (LSM-tree/B-tree) key-value stores over diverse cloud pricing policies for the top three cloud providers – AWS, GCP, and Azure. In order to prune configurations from this massive search space, we present distribution-aware cost models that precisely estimate I/O costs of each possible data structure design, which in turn, helps in comparing pairs of designs. By using the cost models, Cosine reduces the massive dimensionality of the search space into a continuum of holistically optimal configurations. Cosine also enables decision makers in applications to quickly answer rich what- if questions about the changes in workload performance and cost as any of the design, hardware, or cloud provider change.
Speaker Bio:
Subarna Chatterjee is a post-doc at the Data Systems Lab (DASLab) at Harvard University since January 2019. Prior to this, she was also a post-doc with the Myriads team at Inria, Rennes in France. She completed her Ph.D. from Indian Institute of Technology Kharagpur, India from 2013-2017. Her current research interests are NoSQL storage engines and reasoning about their cost and performance on cloud.