A Database System with Amnesia

Big Data comes with huge challenges. Its volume and velocity makes handling, curating, and analytical processing a costly affair. Even to simply “look at” the data within an a priori defined budget and with a guaranteed interactive response time might be impossible to achieve. Commonly applied scale-out approaches will hit the technology and monetary wall soon, if not done so already. Likewise, blindly rejecting data when the channels are full, or reducing the data resolution at the source, might lead to loss of valuable observations. An army of well-educated database administrators or full software stack architects might deal with these challenges albeit at substantial cost. This calls for a mostly knobless DBMS with a fundamental change in database management. Data rotting has been proposed as a direction to find a solution [10, 11]. For the sake of storage management and responsiveness, it lets the DBMS semi-autonomously rot away data. Rotting is based on the systems own unwillingness to keep old data as easily accessible as fresh data. This paper sheds more light on the opportunities and potential impacts of this radical departure in data management. Specifically, we study the case where a DBMS selectively forgets tuples (by marking them inactive) under various amnesia scenarios and with different implementation strategies. Our ultimate goal is to use the findings of this study to morph an existing data management engine to serve demanding big data scientific applications with well-chosen built-in data amnesia algorithms.