We currently increment the topology id only when we "rebalance" the consistent hash (i.e. we add new owners). This allows us to to do some optimizations after a leave, like not forwarding commands (because there are no new owners).

Unfortunately, it is not very intuitive, because it doesn't match how JGroups works, so it can cause bugs like ISPN-2417.

Additionally, it turns out there are many places where we care that a node left, so the code is more complex to handle this (e.g. TransactionTable.useStrictTopologyIdComparison()), or it is more slow for the common case when there is no leaver (e.g. LocalTransaction.getCommitNodes()).