Nowadays, the balancer interleaves consecutive batches of migrations issued by its main thread with a fixed delay of one second to avoid starving CRUD operations with continuous cache invalidations on shards/routers.

While this solution works, it may introduce excessive or unnecessary delays (as the time spent computing the batch of migrations within each round may be already enough for avoiding the problem).

A more efficient alternative would consist in measuring the time between consecutive balancer rounds and applying a variable throttling only when required (similarly to what has been already implemented in the secondary thread of the balancer).