Currently the average query duration is tracked for servers only when a readwritesplit service uses it and is configured with slave_selection_criteria=ADAPTIVE_ROUTING. In addition, only queries classified as reads are tracked.

As query latency is a good metric for measuring the overall performance of a cluster, it should be tracked both at the service level as well as on the server level. Initially, the average should take into account all queries regardless of whether they are writes or reads.

The new statistic should be separate from the existing adaptive_avg_select_time statistic exposed in the servers REST API resource. This will keep the semantics of the existing statistic the same as it was in older releases.