Very similar to #27, but unlike #27 and #28, this requires reconfiguring every compute node. New configuration can be tested and rolled out incrementally.