Our file cluster stores hundreds of terabytes of media files for international cable TV distribution. Effective management of this online resource is necessary to support distribution to our international clients. Thus, we sought to develop a machine learning system that could learn from a combination of factors (eg. file age, future schedule, days since last airing, etc) to predict whether a file is likely to be unused in the future and therefore can be taken off line. In the development of this system, several methods were investigated before settling on Spark MLlib’s Support Vector Machines as the best method due to it’s accuracy and robustness. The system has been tested in production for a couple of months and the results are positive, and therefore plans are to move it into full production usage later this year.
Christopher Burdorf is a software engineer with extensive research and development experience in distributed processing, computer graphics, and machine learning in both the entertainment and defense Industries. He holds a Phd in Mathematics from the University of Bath (UK).