Abstract: Methods, systems, and apparatus are described for throttling a distributed processing system. In one aspect, a method includes identifying records being processed by a distributed processing system that performs agent processes, each of the records including a corresponding timestamp; determining, based on timestamps of the records that have been processed by a first agent process, a first agent progress; identifying a dependent agent process performed by the distributed processing system, wherein the dependent agent process processes only records that have been processed by the first agent process; determining, based on timestamps of records that have been processed by the dependent agent process, a dependent agent progress; and throttling performance of the first process based on the first agent progress and the dependent agent progress.

Abstract: Methods, systems, and apparatus are described for throttling a distributed processing system In one aspect, a method includes identifying records being processed by a distributed processing system that performs agent processes, each of the records including a corresponding timestamp; determining, based on timestamps of the records that have been processed by a first agent process, a first agent progress; identifying a dependent agent process performed by the distributed processing system, wherein the dependent agent process processes only records that have been processed by the first agent process; determining, based on timestamps of records that have been processed by the dependent agent process, a dependent agent progress; and throttling performance of the first process based on the first agent progress and the dependent agent progress.

Abstract: Methods, systems, and apparatus are described for throttling a distributed processing system. In one aspect, a method includes identifying records being processed by a distributed processing system that performs agent processes, each of the records including a corresponding timestamp; determining, based on timestamps of the records that have been processed by a first agent process, a first agent progress; identifying a dependent agent process performed by the distributed processing system, wherein the dependent agent process processes only records that have been processed by the first agent process; determining, based on timestamps of records that have been processed by the dependent agent process, a dependent agent progress; and throttling performance of the first process based on the first agent progress and the dependent agent progress.

Abstract: Methods, systems, and apparatus are described for throttling a distributed processing system. In one aspect, a method includes identifying records being processed by a distributed processing system that performs agent processes, each of the records including a corresponding timestamp; determining, based on timestamps of the records that have been processed by a first agent process, a first agent progress; identifying a dependent agent process performed by the distributed processing system, wherein the dependent agent process processes only records that have been processed by the first agent process; determining, based on timestamps of records that have been processed by the dependent agent process, a dependent agent progress; and throttling performance of the first process based on the first agent progress and the dependent agent progress.

Abstract: Methods, systems, and apparatus are described for performing streaming graph computations in a distributed processing system. In one aspect, a method includes receiving, during a first superstep, streaming graph input that comprises continuous input of first graph updates for updating the state of an existing graph, the existing graph comprising a plurality of vertices and edges, and each first graph update specifying a vertex and a message; scheduling, during the first superstep, the first graph updates for processing during a second superstep; providing, during the first superstep, the first graph updates as input to a graph computation for the existing graph; determining that the graph computation has completed processing for the first superstep, and in response: initiating the second superstep; and processing, during the second superstep, each first graph update; and providing, during the second superstep, graph output based at least in part on the first graph updates.

Abstract: Methods, systems, and apparatus are described for throttling a distributed processing system. In one aspect, a method includes identifying records being processed by a distributed processing system that performs agent processes, each of the records including a corresponding timestamp; determining, based on timestamps of the records that have been processed by a first agent process, a first agent progress; identifying a dependent agent process performed by the distributed processing system, wherein the dependent agent process processes only records that have been processed by the first agent process; determining, based on timestamps of records that have been processed by the dependent agent process, a dependent agent progress; and throttling performance of the first process based on the first agent progress and the dependent agent progress.

Abstract: Methods, systems, and apparatus are described for load balancing in a distributed processing system. In one aspect, a method includes monitoring, for each data processor in a distributed processing system, a rate of cursor progress for the data processor based on timestamps of data units that have been processed, wherein the rate of cursor progress for each data processor specifies a rate of change of an oldest timestamp of an oldest data unit being processed by the data processor; determining a progress threshold for the distributed processing system based on the respective rates of cursor progress of the data processors; determining, based on a first rate of cursor progress for a first data processor, that the first rate of cursor progress does not meet the progress threshold; and in response to the determination, performing one or more load balancing operations on the distributed processing system.