Machine Learning Models for SSD and HDD Reliability Prediction

This paper compares hard disk drives (HDDs) and solid-state drives (SSDs), the two most used storage devices in datacenters, which frequently fail and are among the main causes ofdata center downtime. Using a six-year field data of 100,000HDDs from the Backblaze dataset and a six-year field data of30,000 SSDs from a Google data center, we characterizeworkload conditions that prompt drive failures. %and show thatthey differ from common expectation. We develop machinelearning models that accurately predict the drive failure stateseveral days in advance and provide highly interpretable resultsthat are useful to identify the causes and symptoms of drivefailures.