Technical Perspective

From a historical perspective, relational database management systems (RDBMSs) have integrated many specialized systems and data models back into the RDBMS over time. New workloads motivated specialized systems for performance, but over time, general-purpose RDBMSs absorbed this functionality to avoid boundary crossing. We already witnessed this process for object-relational functionality, XML and JSON data types, OLAP/HTAP systems, and RDF/graph processing, while for natural language processing (NLP), time series, and machine learning (ML), the outcomes remain unclear. Interestingly, graph processing, NLP, and time series are largely ML workloads too. For this reason, integrating data management and ML is of high practical relevance and has been addressed by (1) integrating ML into RDBMSs, and (2) specialized ML systems. The paper "Declarative Recursive Computation on an RDBMS" [3] by Jankov et al. makes a very valuable contribution by reconciling these two areas and showing the potential of recursive computations on an RDBMS, as the backend-not necessarily frontend-for large-scale machine learning.