Kafka Storage Load Balancer

Storage skew sets a ceiling for a whole cluster: the fullest broker decides what the fleet can accept. The standard remedy is a planned window and one large reassignment, which is unsafe in proportion to how much it moves.

This rebalancer treats redistribution as a continuous operation instead. Batches of 200 replicas move at a time, each batch admitted only when target brokers are below their disk thresholds, each one resumable so an interruption costs a batch rather than a run. Rollout and configuration ride on Salt.

Built single-handedly and run in production across a fleet holding more than 17,000 partition leaders.