DevConf.CZ 2020

More than a million terabytes of data gets generated every day, and every bit of that can be valuable. Therefore, modern data storage solutions need to be reliable, scalable, and efficient. Storage systems like RAID and Ceph use replicas or erasure-coded redundancy to provide fault-tolerance. So, while scaling up to exabyte-level is possible, it can be resource-intensive and expensive.
However, these issues can be mitigated by some clever use of machine learning. We can use ML models to predict the remaining-useful-life or time-to-failure of hard drives, and then create or destroy replicas according to those predictions. In this way, storage can be made more resource-efficient. This talk will discuss the techniques we used and the models we built for this task. Our open-source-built model outperforms the ProphetStor model that is currently on upstream Ceph. Additionally, we frame the problem in a Kaggle competition format to provide a platform to the community to contribute their ideas