Minimum Drain Rate for a Distributed Backlog
A distributed job system runs `n` partitions, where partition `i` holds `piles[i]` pending jobs. Before a maintenance window opens, the control plane must publish one integer drain rate, `k`, measured in jobs per round. During each round the dispatcher selects exactly one non-empty partition and completes up to `k` jobs from it. Two partitions never drain in the same round.