Datacenter schedulers (Borg, Kubernetes, Mesos) face the partition problem at every node-pair: given a set of pods with CPU/memory requests, assign them to two replicas to minimize the maximum load. That's the makespan version, equivalent to two-machine partition. Real schedulers solve it by approximation — longest-processing-time-first (also known as Graham's rule) is a 4/3 approximation and what Kubernetes' default scheduler effectively runs. The same problem appears in transaction sharding (split keys to balance across two shards), in VLSI bisection (split a chip's modules across two halves of a die), and in cryptographic backup secret sharing (Shamir-style schemes balance share sizes).
60
Load balancing across racks
Splitting jobs across two machines to finish at the same time is the scheduling-floor face of partition.