Currently the plan is to enable persistent samples by default in 9.1, but without any automation of the `analyze` command. This means that most customers will not have any samples persisted, but will still incur the cost of the samples collection find query when CBR runs.

This overhead is expected to be constant across queries/collections, so a google benchmark should be used to determine how big of a planning time regression we can expect in these cases and to assess whether that poses a problem for release without automatic `analyze`.