Big Data Normalization for Massively Parallel Processing Databases

Computer Standards and Interfaces. 2017. Vol. 54. No. P2. P. 86–93.

Golov N., Rönnbäck L.

High performance querying and ad-hoc querying are commonly viewed as mutually exclusive goals in massively parallel processing databases. Furthermore, there is a contradiction between ease of extending the data model and ease of analysis. The modern 'Data Lake' approach, promises extreme ease of adding new data to a data model, however it is prone to eventually becoming a Data Swamp - unstructured, ungoverned, and out of control Data Lake where due to a lack of process, standards and governance, data is hard to find, hard to use and is consumed out of context. This paper introduces a novel technique, highly normalized Big Data using Anchor modeling, that provides a very efficient way to store information and utilize resources, thereby providing ad-hoc querying with high performance for the first time in massively parallel processing databases. This technique is almost as convenient for expanding data model as a Data Lake, while it is internally protected from transforming to Data Swamp. A case study of how this approach is used for a Data Warehouse at Avito over a three-year period, with estimates for and results of real data experiments carried out in HP Vertica, an MPP RDBMS, is also presented. This paper is an extension of theses from The 34th International Conference on Conceptual Modeling (ER 2015) (Golov and Rönnbäck 2015) [1], it is complemented with numerical results about key operating areas of highly normalized big data warehouse, collected over several (1-3) years of commercial operation. Also, the limitations, imposed by using a single MPP database cluster, are described, and cluster fragmentation approach is proposed. 

Keywords: analytics, big data, MPP, Database, Normalization, Ad-hoc, Querying, Performance, Modeling, Data Lake