Techniques are presented for maintaining data distributed across a plurality of storage drives (drives) in a robust manner. A method includes (a) collecting physical state information from each drive of the plurality of drives, (b) generating a predicted failure probability of each drive based on the collected physical state information from that drive, the predicted failure probability indicating a likelihood that that drive will fail within a predetermined period of time, and (c) rearranging a distribution of data across the plurality of drives to minimize a probability of DU/DL. Systems, apparatuses, and computer program products for performing similar methods are also provided.