Write tests of HA3 where we force new allocators to be allocated, do an abort and then make sure that the allocators are consistent (rolledback) across all nodes. Do this both with and without group commit.
When group commit is enabled there is an additional question since isolated writes will be true in RWStore.reset(). Note that there is no call to something similar to postHACommit() in the abort code path. Are the allocators appropriately synchronized?