Scale-out technique in which data is divided across partitions using a deterministic function (e.g. a hash of the entity ID) so the same resource always routes to the same partition. Enables per-partition atomic guarantees (needed for idempotency-key lookups and uniqueness) while letting the overall system scale. Presented as the CAP-theorem response to the worry that the more a receiver scales out to handle load, the harder central atomic guarantees become — sharding makes partitions disjoint and independently consistent.