Hard drives are an essential component of modern data storage. In order to reduce the risk of data loss, hard drive failure prediction methods using the Self-Monitoring, Analysis and Reporting Technology attributes have been proposed. However, these methods were developed from datasets not necessarily representative of operational systems. In this paper, we consider the Backblaze public dataset, a recent operational dataset from over 47,000 drives, exhibiting hard drive heterogeneity with 81 models from 5 manufacturers, an extremely unbalanced ratio of 5000:1 between healthy and failure samples and a real-world loosely controlled environment. We observe that existing predictive models no longer perform sufficiently well on this dataset. We therefore selected machine learning classification methods able to deal with a very unbalanced training set, namely SVM, RF and GBT, and adapted them to the specific constraints of hard drive failure prediction. Our results reach over 95% precision and 67% recall on a one year real-world public dataset of over 12 million records with only 2586 failures