When the caching was implemented, the query classifier cache was not present. That meant that if a rule required the presence or absence of a database/table, then the statement had to be parsed and the parsing had a significant impact on the performance. Now with the query classifier cache the classification information is in general available without having to parse the statement. Hence, the impact of rules on the performance should be much smaller than what it used to be. Measure what the impact is and document it.