The documentation currently recommends running a `mongos` on the primary shard of a sharded cluster. This is not a best practice in terms of field recommendations. A quick survey of #consulting reveals that this is not a common recommendation. It seems that recommending an intermediate layer for `mongos` routers (i.e. dedicated servers) makes more sense in most cases than co-locating instances.

I submitted a pull request here as an example:
https://github.com/mongodb/docs/pull/3146