During the first week in May, I could share Mobiliar’s experience with the practical use of machine learning technology for DB2 systems tuning at IDUG’s annual North American DB2 Technical conference in Anaheim, CA.
Machine learning building blocks are already part of some major applications in our application landscape, such as the underwriting application for small and medium enterprises or in our fraud detection. More recently, we are also making use of cognitive technologies in the area of DB2 system tuning.
My presentation started with an introduction of machine learning, targeted to the audience of database administrators. Therefore, a simple use case was highlighted: Classify fruits according to some features into apples and oranges. Easy for humans, a bit more complex for an algorithm.
After this introduction, the real system tuning use case was discussed: The classification of DB2 tablespaces and indexes into the different buffer pools, which were separated by their individual page stealing algorithm. Until now, this classification has been made manually by a DBA, who not only needed a certain level of DB2 know-how, but also a quite deep level of application understanding. Now this manual approach has been replaced by a machine learning algorithm: All tablespaces and indexes already defined were used as training and test data for the algorithm during its learning phase (in this case, we selected Random Forest after comparing it to three other algorithms). For the most recent application release and its corresponding DB2 objects created, the algorithm was applied for the first time, saving manual efforts while producing results of at least similar quality compared to what I manually defined before.
The presentation produced a great interest, both among fellow DB2 users and the community of ISV product developers.
In addition to my proper presentation, I also attended many other sessions, discussions and hands-on workshops, basically for DB2 performance subjects, but also regarding the integration of DB2 in a big data architecture, experiences with the new DB2 V12 version, and last but not least database activity monitoring.
After this first step, the usage of machine learning won’t stop: Next steps are already in planning, e.g. for query performance tuning, and for database activity monitoring. Stay tuned!
The next stop on this journey will be Munich: Learn how to classify items into different groups such as “no risk”, “limited risk”, ‘”high risk” or “fraud highly probable” and detect how to audit an application composed of machine learning building blocks at ISACA’s 2017 EuroCACS Conference in Munich, Germany.
Attendees will discuss solutions and strategies in assurance, risk and security, including the question how assurance professionals can advance their careers and impact their enterprises. In my presentation, titled “Machine learning for auditors” I will discuss practical application of machine learning algorithms both from a user’s perspective and from an auditor’s point of view: “Individuals will accept the results of machine-learning-based applications only if the results are explainable and understandable. To gain trust in these models and algorithms will become an important part of the auditor’s profession.”
ISACA, a global association serving more than 130,000 members and certification holders in more than 180 countries, will offer 60 sessions in five tracks for the EuroCACS Conference. The conference will feature valuable career guidance from renowned keynote speakers. Pre- and post-conference workshops will offer hands-on training on privacy programs, database security and audit, risk strategies and data analysis.