Google patents method to balance data workloads across servers

1. A computer-implemented method executed by data processing hardware that causes the data processing hardware to perform operations comprising: partitioning a data set for an application job into a plurality of partitions based on a key; assigning, to each worker computer in a set of worker computers, one or more partitions of the plurality of partitions; receiving, from each respective worker computer in the set of worker computers, a respective load measure indicating a computational load of the respective worker computer; selecting, based on the respective load measure of each respective worker computer, a first worker computer in the set of worker computers for a replication operation for replicating at least one partition from the first worker computer to a second worker computer in the set of worker computers; replicating, the at least one partition from the first worker computer to the second worker computer; after replicating the at least one partition from the first worker computer to the second worker computer, distributing requests for a workload associated with the at least one partition between the first worker computer and the second worker computer such that the first worker computer and the second worker computer share the workload for the at least one partition; and in response to distributing the requests for the workload associated with the at least one partition between the first worker computer and the second worker computer, de-replicating the at least one partition from the second worker computer by deleting the at least one partition from the second worker computer.