Page History

...

Cruse Control is a LinkedIn project to automate some aspects of Kafka broker cluster administration. One thing it does is dynamically rebalance the partition assignment over the broker instances based on a large number of metrics it collects including CPU load, disk usage, etc. It structures the assignment optimization task as a Goal-Based Optimization problem. See https://en.wikipedia.org/wiki/Goal_programming for more information about this paradigm. Note that "goal programming" often implies that you represent the goals as a system of linear equations, and then solve the system to maximize some variables (aka Linear Programming), but that's not how Cruise Control is implemented. It just expresses the optimization problem as a system of goals and seeks a satisfactory (not optimal) solution by sequentially satisfying each goal.

The structure of the implementation is that you get a list of Goal implementations, each one corresponding to an optimization goal, like disk usage. The overall optimizer sorts the goals by priority, and then passes in the cluster assignment to the highest priority goal. That goal proposes modifications to the assignment (typically using an iterative algorithm similar to the StickyAssignor's). Once it's happy with the proposal, it returns. Then the optimizer passes in the new proposed assignment to the next goal, and so on. Each goal is responsible for ensuring that its proposals do not violate any of the prior, higher priority, goals. The API provides a hook that the goal can call during its optimization passes to ensure it doesn't violate the higher priority goals.

...