In WT-4438, we determined that any statistics that do not use atomic increment may be inaccurate. For many cases we don't care about a fractional percentage drift.  But in particular when a statistic holds a counter, and the drift can accumulate in one direction, we may care.

A check of the source tree looking for matches of STAT.*DECR should identify most or all cases where a statistic value represents a resource count.

For connection-wide statistics counters, we have: 

 lsm_work_queue_switch
 lsm_work_queue_manager
 lsm_work_queue_app
 session_open
 txn_prepare_active

and for statistics that are per-data source, we have:

 cursor_open_count
 cursor_update_bytes

We should fix these to use the STAT.*ATOMIC macros, or document that they may accumulate drift.