Follow up improvements for user configuration:
Tests from Tristan seem to indicate there is no drawback in a model of 1 cache per "table".