The current sharding balancer policy attempts to move chunks from shards which have more than the cluster-wide per-shard chunk count average to shards which have less than it. However there is a bug due to which chunks will move from highly loaded to less-loaded shard if the difference between the two is sufficiently large.
In particular consider the following case:
Shard1 - 1000 Shard2 - 900 Shard3 - 900 Shard4 - 800 Shard5 - 0 Shard6 - 0
Intuitively, this should produce only two migrations: Shard1 -> Shard5 and Shard2 -> Shard 6. However it will also cause a third migration from Shard3 -> Shard4, because 900 is above the ideal per-shard chunk number of 600.