Described is the differentiation of replicas in a large distributed object store as either being smoothing replicas based on an amount of load on storage nodes (bricks), or as spreading replicas based on a substantially random distribution among the system's bricks. The smoothing replicas are placed among the lowest usage bricks, while the spreading replicas are placed randomly throughout other bricks in the system independent of load. As a result, fast, primarily parallel data repair is facilitated by selecting a spreading replica when repair is needed, while load balancing is facilitated by placing a smoothing replica on a low-usage brick when a new replica is checked in, and selecting a smoothing replica when load balancing is triggered by overloading of a brick and/or addition of a new brick. Check-in, data repair and load balancing policies specify how to use smoothing replicas and spreading replicas.