It is unclear from the documentation how to use the query profiler in a sharded cluster, and the results are confusing.

I can't use setProfilingLevel on a mongos server, so I have to set it individually on each shard. This leads to a 'system.profile' collection appearing when connecting to the mongos, but it claims to contain only 92 documents whereas each shard contains much more than that.

I expected the system.profile collection on the mongos to act like any other sharded collection, and represent the combined contents of the system.profile collection on each shard. This is not happening, so either it's a bug, or the usage/docs are misleading.

In short, I'm missing a reliable and clearly-documented way to get system.profile data from the entire cluster.