Drives of a RAID group are classified as either healthy or failing using a trained learning process. The failure probability is then determined for each failing drive using a Venn-ABERS framework which provides a boundary range on the failure prediction probability. A variable sparing mechanism is used to enable one or more drives of the RAID group to be used as dual-purpose drives. In a first state, the dual-purpose drives are user-addressable drives and are available to be used to process IO workload on the RAID group. Spreading the IO workload on the RAID group across a larger number of drives results in increased performance in the form of reduced latency. In a second state, the dual-purpose drives are not user-addressable and are spare drives in the RAID group, which improves the level of protection provided to data stored in the RAID group.