With the old logic, having a single stale/lagging replica would not
allow us to use caught up replicas, which would increase the load to
primary where we could avoid it. Also adding the comments on the
possible improvements of the similar method.

Changelog: performance