After increasing one of our clusters to 1600 nodes, we observed a large amount of invalid region moves(more than 30k moves) fired by the balance chore. Thus we simulated the problem and printed out the balance plan, only to find out many servers that had two regions for a certain table(we use by table strategy), sent out both regions to other two servers that have zero region.
In the SimpleLoadBalancer's balanceCluster function,
the code block that determines the underLoadedServers might have a problem:
if min is zero, some server that has load of zero, which equals to min would be marked as underloaded, which would cause the phenomenon mentioned above.
Since we increased the cluster's size to 1600+, many tables that only have 1000 regions, now would encounter such issue.
By fixing it up, the balance plan went back to normal.