At the beginning of every balancing round, data size stats are collected from each shards in order to spot potential imbalance and schedule migrations. Currently, the balancer is requesting such stats in batches, collecting info for up to 20 collections per request.

Purpose of this ticket is to rethink the current batching policy because this hardcoded limit may become source of scalability issues in case of thousands of sharded collections.

[EDIT] REPURPOSED: this ticket also added a cache of imbalanced collections that lives between rounds in order to remember which collections had a migration scheduled during the previous round.