Consider how to optimize system queries against system collections via CBR.
According to this discussion system queries may fail with CBR if it uses histogram CE.
This ticket should propagate into CBR the information that the query that is being optimized is a system one. Most likely this can be done by inspecting the collection.
Then choose a CBR mode and settings that is most appropriate for system queries irrespective of the CBR query knob (most likely sampling CE). The goal is to never crash, and to also provide better performance than multi-planning.

Consider whether it is possible to analyze all system collections/queries and determine if heuristic CE would do the job, and when. Also consider if such queries may benefit from custom settings, such as smaller sample size.

Reason how to ensure we have testing coverage for these cases?