Ray for large-scale time-series forecasting of disk utilization in a data center

Forecasting of disk utilization is crucial for one of KocDigital's clients, a leading private corporate cloud company that operates three data centers in Turkey. Full disk utilization leads to severe availability issues that undermine the company's reputation, so the company needs to be able to forecast disk utilization for several weeks, observe the risk of exceeding a critical utilization threshold, and take precautions to prevent availability issues. Forecasting a single server's disk utilization is a relatively simple task that requires time-series forecasting capabilities, but producing forecasts for tens of thousands of servers each and every day would be extremely challenging. In this talk, we'll share how Ray helped KocDigital ML engineers accelerate the training speed (finding out the best fitting model for each server) and generate inference almost 10 times to create a robust forecasting solution for the company. Ray Core (task and actor) and Ray Clusters were implemented successfully to overcome the computational complexity that stems from the scale of the problem.