I think that given the difficulty of doing this programmatically, even if we knew the load per token (which we don't), the best thing to do here is expose relocate to the user and let them manage it themselves, if need be.

Since we already expose everything we need via jmx, simple patch to expose 'taketoken' to nodetool, which will relocate the tokens from wherever they exist to the node it is called against.

What happens if I run this against a non-vnode cluster?

You'd almost get a very weird way to decommission a node, except the system table would throw an assertion. Updated patch won't let you relocate a token if the source node doesn't have more than one token.

+1

Committed.

I have a question about the partitioning schema of Cassandra. Is all the ranges of each vnodes in the Cassandra system are equal?

No, currently they are randomly allocated resulting in a variability of balance to the ring. See CASSANDRA-7032

Cassandra uses moving node to balance, moving data manually, just by using the command of nodetool move? Is there an automatic service to complete data moving?

For others who come here and happily find there's a tool for moving tokens, please note that this tool was removed in CASSANDRA-7601.