It's difficult to see the impact of particular mongos/mongod algorithms for refreshing config data because there's very few (if any) statistics gathered on how often we need to do this. We should at least record the number of times config data is accessed, could also record buckets like :

• local query/update/insert stale version (doesn't actually hit config server, usually)
• remote query stale version
• remote update/query stale writeback
• migrate / split
• repeated stale version resulting in force

Could also be a good test case for labeled log levels of whatever implementation - logging every interaction with the config server would be extremely useful, but only when debugging particular issues and would otherwise be very verbose.